A Cognitive Approach to the Verb: Morphological and Constructional Perspectivs [Reprint 2011 ed.] 9783110817461, 9783110170313


176 46 79MB

English Pages 317 [320] Year 2000

Report DMCA / Copyright

DOWNLOAD PDF FILE

Table of contents :
Introduction
Section I: Basic issues
Topic, subject, and possessor
Language as a cognitive network
Section II: Morphological aspects of the verb
The Norwegian verb
Building blocks or network relations: Problems of morphological segmentation
Past tense acquisition in Norwegian: Experimental evidence
Individual variation in past tense inflection: Experimental data from Norwegian SLI children
A connectionist model of past tense acquisition in Norwegian
Regular inflection in the mental lexicon: Evidence from Italian
Section III: Constructional aspects of the verb
Malefactive by means of GIVE
The Norwegian and Russian reflexive-middle-passive systems and Cognitive Grammar
Norwegian reflexive-middle-passive marker -s: Homonymy/polysemy and morphology
Control and transitivity: A study of the Norwegian verb love 'promise'
References
Index
Recommend Papers

A Cognitive Approach to the Verb: Morphological and Constructional Perspectivs [Reprint 2011 ed.]
 9783110817461, 9783110170313

  • 0 0 0
  • Like this paper and download? You can publish your own PDF file online for free in a few minutes! Sign Up
File loading please wait...
Citation preview

A Cognitive Approach to the Verb

W DE

G

Cognitive Linguistics Research 16

Editors Rene Dirven Ronald W. Langacker John R. Taylor

Mouton de Gruyter Berlin · New York

A Cognitive Approach to the Verb Morphological and Constructional Perspectives Edited by Hanne Gram Simonsen Rolf Theil Endresen

Mouton de Gruyter Berlin · New York 2001

Mouton de Gruyter (formerly Mouton, The Hague) is a Division of Walter de Gruyter GmbH & Co. KG, Berlin

© Printed on acid-free paper which falls within the guidelines of the ANSI to ensure permanence and durability.

Die Deutsche Bibliothek — Cataloging-in-Publication Data

A cognitive approach to the verb : morphological and constructional perspectives / ed. by Hanne Gram Simonsen ; Rolf Theil Endresen. - Berlin ; New York : Mouton de Gruyter, 2001 (Cognitive linguistics research ; 16) ISBN 3-11-017031-0

© Copyright 2000 by Walter de Gruyter GmbH & Co. KG, D-10785 Berlin All rights reserved, including those of translation into foreign languages. No part of this book may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher. Printing: Werner Hildebrand, Berlin Binding: Lüderitz & Bauer, Berlin Printed in Germany

Contents

Introduction Hanne Gram Simonsen and Rolf Theil Endresen

Section I: Basic issues Topic, subject, and possessor Ronald W. Langacker

11

Language as a cognitive network Richard A. Hudson

49

Section II: Morphological aspects of the verb The Norwegian verb Rolf Theil Endresen and Hanne Gram Simonsen Building blocks or network relations: Problems of morphological segmentation Helge Gundersen

73

95

Past tense acquisition in Norwegian: Experimental evidence Hanne Gram Simonsen

129

Individual variation in past tense inflection: Experimental data from Norwegian SLI children Kirsten Meyer Bjerkan

153

vi

Contents

A connectionist model of past tense acquisition in Norwegian Anders N0klestad

165

Regular inflection in the mental lexicon: Evidence from Italian Paolo. F. Matcovich

189

Section 111: Constructional aspects of the verb Malefactive by means of GIVE Ole Torfinn Fagerli

203

The Norwegian and Russian reflexive-middle-passive systems and Cognitive Grammar Hans-Olav Enger and Tore Nesset

223

Norwegian reflexive-middle-passive marker -s: Homonymy/polysemy and morphology Hans-Olav Enger

243

Control and transitivity: A study of the Norwegian verb love 'promise' Kristian Emil Kristoffersen

259

References

281

Index

305

Introduction Hanne Gram Simonsen and Rolf Theil Endresen

1. Background In June 1998, a small conference took place at Sanner Hotel, Gran, Norway, with the title "The Verb in Cognitive Grammar: Morphological, Semantic, and Syntactic Aspects". The bulk of the participants at the conference had all for a longer or shorter time been affiliated to a Cognitive Linguistics research group based at the Department of Linguistics at the University of Oslo. This group, named The Oslo Project in Cognitive Grammar (OPCoG), is headed by Rolf Theil Endresen and Hanne Gram Simonsen, and has been active since 1992. The conference was financed through a prize of NOK 75.000 awarded to the research group by the Faculty of Arts, University of Oslo. There were two invited keynote speakers: Ronald W. Langacker from the University of California at San Diego, and Richard A. Hudson from University College, London. In addition, 18 other papers were presented. The present book consists of a selection of the papers, which have been through a thorough reviewing and rewriting process.1 Given the topic of the conference, "The Verb in Cognitive Grammar"—or, rather—"The Verb in Cognitive Linguistics", the papers concentrated on a central theme and addressed this theme from different angles, illustrated through different languages, but all within a cognitive linguistics perspective. A majority of the papers described different aspects of Norwegian—a language not earlier investigated at such a scale within cognitive linguistics. Norwegian offers valuable contributions through being closely related to English, yet differing in certain important and interesting ways. In addition, several other languages were included, providing a wider typological range.

2

Hanne Gram Simonsen and Rolf Theil Endresen

In their papers, the two keynote speakers provided a broad perspective and a general theoretical background for the main theme. Richard A. Hudson drew mainly upon his own theory of Word Grammar in explaining his view of language as a network, while Ronald W. Langacker discussed and developed new ways of characterising semantic and grammatical organisation within his theory of Cognitive Grammar.2 We have placed these two papers together in Section One of the book. The rest of the papers fall naturally into two groups: those mainly addressing morphological aspects of the verb—placed in Section Two, and those discussing constructional aspects of the verb—placed in Section Three. However, the division into three main sections does not mean that there is a strict border between the sections, as several of the papers touch upon themes crossing the section borders. In this way all the papers can be seen as being in a network relation to each other ...

2. Section I: Basic issues The first section starts with Ronald W. Langacker: Topic, subject, and possessor. This paper explores the special, intimate relationship between the grammatical notions topic, subject and possessor, illustrating this relationship with examples from a wide range of languages, and analysing it within the framework of Cognitive Grammar. The notions are discussed in terms of reference point relationships, where they are presented as prominent reference points with respect to propositions, processes, or things, creating asymmetric relationships with an inherently dynamic, temporal dimension. The differences between them are further discussed as related to different levels of organisation: the lexical, the grammatical, and the discourse levels. This way, the paper presents a new and integrated, conceptually based view of these phenomena. Then follows Richard Hudson: Language as a cognitive network. As indicated by the title, it develops the central idea of language being a network, and discusses several consequences of this view, relating them to cognitive theories like Langacker's Cognitive

Introduction

3

Grammar (Langacker 1987a, 1990, 1991), Goldberg's Construction Grammar (Goldberg 1995a), Bybee's morphology (Bybee 1985, 1995a), and most thoroughly to his own theory, Word Grammar (Hudson 1984, 1990, 1995, 1998). The basic similarities between these theories, in spite of different terminologies and and starting points, are underlined. He illustrates his arguments with examples from two areas of language analysis: the polysemy/homonymy distinction, and regular/irregular morphology. The examples from morphology include a Word Grammar analysis of Norwegian verbs, pointing forward to subsequent papers in the book.

3. Section II: Morphological aspects of the verb This section contains six papers, most of them concerning Norwegian. To give an overview of some relevant aspects of this language which may function as necessary background knowledge, we begin with Rolf Theil Endresen & Hanne Gram Simonsen: The Norwegian verb. This paper gives a short survey of the Norwegian verbal system, starting with some phonological notes, followed by an outline of the morphological and periphrastic constructions and their use, and summing up by an overview of the verb classes, their relative sizes, and their phonological and morphological make-up. To be able to evaluate different theoretical approaches to morphology detailed knowledge both of phonological patterns and of frequency patterns in the relevant language is needed. Then follows Helge Gundersen: Building blocks or network relations: Problems of morphological segmentation. Here two morphological models are contrasted: the morpheme or building block model, and the network model as presented in cognitive linguistics. It is demonstrated how the requirement of exhaustive segmentation of the building block model creates problems which do not appear within a network model, illustrated through a special type of Norwegian verbs, derived by the suffix -er. Gundersen argues that while segmentation may be an important part of morphology, it is neither necessary nor sufficient to account for morphological

4

Hanne Gram Simonsen and RolfTheil Endresen

relations, and he posits Langacker's and Bybee's theories as alternative, superior accounts. The remaining papers in this section all concern the acquisition of past tense inflection of verbs, addressing the role of input factors (phonological properties and type and token frequencies) during the acquisition process. In this connection they also address the ongoing debate as to which kind of model is better suited to account for the acquisition and processing of morphology: a single mechanism model or a dual mechanism model. According to the dual mechanism model, there is a qualitative difference in processing between regular and irregular verbs, while the single mechanism model assumes essentially the same kind of processing for both regular and irregular verbs, input factors playing an important role for both types of inflection. These questions are investigated through different kinds of experimental data in the following papers. Past tense acquisition in Norwegian: Experimental evidence by Hanne Gram Simonsen investigates the performance of Norwegian 4-, 6-, and 8-year-olds in an experimental task eliciting past tense forms of verbs from different verb classes, and also comparing their performance with that of adults. The results clearly indicate that input factors play an important role, both for correct performance and for the errors made, and also that they play a role for all verb classes. Furthermore, the relative importance of each factor changes in the course of development, in such a way as to lend support to a single mechanism, exemplar-based model for morphological acquisition. Kirsten Meyer Bjerkan: Individual variation in past tense inflection: Experimental data from Norwegian SLI children reports on results from the same elicitation task applied to 6- and 8-year-old Norwegian children with specific language impairment. The different response patterns in the children are discussed as results of different kinds of schema formation and entrenchment. Anders N0klestad: A connectionist model of past tense acquisition in Norwegian presents a connectionist model of past tense acquisition of Norwegian, comparing the model's behaviour to results of the past tense elicitation experiments presented in the preceding papers. The

Introduction

5

modelling results show that a single mechanism is able to learn to produce past tense forms of Norwegian verbs, on the basis of the corresponding present tense forms. The developmental profile of the model shows striking similarities with the other experimental data, but also some very interesting differences, which are given a thorough discussion, both in the light of the simulation architecture, and in the light of the different models proposed. The final paper in this section is Paola F. Matcovich: Regular inflection in the mental lexicon: Evidence from Italian. Closely related to the preceding papers, it investigates the acquisition and processing of Italian verb inflection in 4-, 6-, and 8-year-old children and adults through two sets of experiments—eliciting past participle forms of existing verbs as well as nonce verbs. Again, the results offer support for a single mechanism model, since no qualitative difference is found between regular and irregular inflection with respect to the influence of phonological or token frequency factors.

4. Section III: Constructional aspects of the verb The second section contains four papers, starting with Öle T. Fagerli: Malefactive by means of GIVE. Fagerli discusses semantic aspects of certain African languages—more specifically the benefactive and the malefactive, both in serializing languages using a word identical to the verb GIVE, and in languages using verbal extensions to express these semantic relations. In both language types malefactive events may be marked by expressions used also for benefactive events. Seeing both types of expressions as special instances of a cognitively and semantically basic GIVE relation, the author argues that the malefactive meaning may be seen as an extension of the originally benefactive marker, in combination with the semantics of the host verb and its complements. In the next paper, Hans-Olav Enger and Tore Nesset: The Norwegian and Russian reflexive-middle-passive systems and Cognitive Grammar, the authors give an extensive analysis and comparison of the reflexive-middle-passive systems in Norwegian

6

Hanne Gram Simonsen and RolfTheil Endresen

and Russian, and show how they can be presented in Cognitive Grammar networks with Schemas and extension relations. Their analysis illustrates the value of networks of this type both as heuristic devices and as valuable tools for gaining new insight, particularly in typological research. Hans-Olav Enger: The reflexive-middle-passive marker -s in Norwegian: Homonymy/polysemy and morphology singles out one Norwegian suffix marking a set of the meanings examined in the previous paper, discussing whether it should be counted as one polysemous suffix or several homonymous ones. Arguing in favour of the polysemy position, the author examines the status of the suffix as inflectional or derivational in different functions, relating his discussion to Bybee's (1985) concept of relevance. Both in its content and its discussions, this paper also ties back to Hudson's paper. Finally follows Kristian Emil Kristoffersen: Control and transitivity: A study of the Norwegian verb love 'promise'. Like the preceding papers, this one also exploits the notion of polysemy, but this time not of lexical items, but of constructions. The author investigates the relation between transitivity and argument control in one Norwegian verb in a Construction Grammar perspective. It is argued that argument control should be treated as a property not of the individual verbs, but of the constructions in which the verb may occur. The specific, different patterns of transitivity and argument control concerning the verb love are explained through the notion of multiple inheritance from different construction types. Furthermore, the frequency of use of the different love constructions is seen as a result of differences in type frequency of the constructions.

5. Final remarks and acknowledgements All the authors have profited from the fruitful and active discussion of all the participants during the conference. Of particular value was the active participation of the two keynote speakers, Richard A. Hudson and Ronald W. Langacker, who deserve special thanks.

Introduction

7

The editors of the Cognitive Linguistics Research series gave very useful advice and comments during the publishing process. We are also grateful to the Faculty of Arts, University of Oslo, for providing financial support. Finally, as convenors of the conference, we would like to thank our conference secretary, Ame Martinus Lindstad. He has been an invaluable coworker at all stages, both at the conference and through the finalising of the book manuscript.

Notes 1. Reviewing is a time-consuming job. We wish to thank the following persons for taking of their time to review the papers in this volume, giving valuable suggestions, comments and evaluations: Hans Basb011, Elizabeth Bates, Andrew Carstairs-McCarthy, Bruce Derwing, Wolfgang Dressier, Lena Ekberg, Hans-Olav Enger, Richard Hudson, Laura Janda, Suzanne Kemmer, Ronald W. Langacker, Virginia Marchman, Tore Nesset, Eric Pederson, Trond Trosterud, Äke Viberg, Wolfgang Wurzel. 2. This paper by Ronald Langacker was first presented at the Bielefeld Symposium in April 1998, and will be published in the proceedings from that symposium. We are grateful that the arrangement committee for the Bielefeld Symposium has agreed to our publication of Langacker's contribution in the present book.

Section I Basic issues

Topic, subject, and possessor Ronald W. Langacker

Language offers ample evidence that the grammatical notions topic, subject, and possessor have special affinities for one another. The basis for their close relationship is explored from the standpoint of cognitive grammar. They are all described as instantiating a basic aspect of dynamic conceptualization involving successive foci of attention: a salient reference point affords mental access to a context (its dominion,) in which a target is located. A possessive construction is characterized schematically as specifying a reference point relation between two things. A topic construction establishes a reference point relation between a thing and a proposition. Subject and object are analyzed as reference points with respect to the relationship designated by a verb or a clause (called a process): they represent the initial and secondary focal participants accessed in building up to the full relational conception. Evidence is provided for their focal prominence, for their asymmetry in this regard, for this asymmetry having a temporal dimension, and for its assimilation to reference point phenomena. The relation between topic and subject is discussed in terms of function and levels of organization.'

1. Evidence of special affinities I start from the observation, not at all a novel one, that the grammatical notions topic, subject, and possessor display certain affinities for one another. Numerous phenomena, in particular languages and cross-linguistically, suggest that these constructs have a special and intimate relationship. My goal is to find a precise description of the abstract similarity that is presumably responsible for their affinity. This supports the semantic characterization of subject and object

12

Ronald W. Langacker

proposed in cognitive grammar (Langacker 1991:7.3). It also bears on the internal conceptual organization of verbs and clauses.2 Let me start by offering some empirical evidence for topic, subject, and possessor being specially related in some fashion. I will merely list certain phenomena that have come to my attention. The following inventory is not intended to be exhaustive, nor have I undertaken a systematic survey. Other linguists could no doubt easily add to the list. The close connection between subject and topic is of course well known, if not notorious. Linguists often have difficulty determining whether a given nominal element is best analyzed as a subject, a topic, or both. Moreover, topic status is commonly ascribed to subjects as part of their characterization, at least prototypically. Givon, for instance, describes subject and object as primary and secondary clausal topic (1984:138). I basically agree with this description and will try later to render it less vague. Diachronically, topics often evolve into subjects. Li and Thompson (1976:484) describe this recurrent path of grammaticization as follows: "...it seems clear that subject and topic are not unrelated notions. Subjects are essentially grammaticalized topics; in the process of being integrated into the case frame of the verb (at which point we call them subjects), topics become somewhat impure, and certain of their topic properties are weakened, but their topicness is still recognizable...That is why many of the topic properties are shared by subjects in a number of languages. For example, some [subject-prominent] languages do not allow indefinite subjects.". A second observation concerns the notion pivot. I define a pivot as the entity within a proposition that corresponds to a topic and thus enables the proposition to function as a comment with respect to that topic. In (1), for example, the topic is your friend, and the pivot is the coreferential pronoun in the following clause. I am reasonably confident that systematic investigation would reveal a strong tendency for the subject of a comment clause to function as pivot. Secondarily, the object of the comment clause often serves as pivot, or else the possessor of the subject or object. The expressions in (1) are, I believe, quite typical.

Topic, subject, and possessor

(1)

a. b. c. d.

Your friend, Your friend, Your friend, Your friend,

13

he will never amount to anything. I really like him. his mother just died. I really liked his mother.

Data showing an intimate relationship between subject and topic in Russian is presented in Nichols, Rappaport, and Timberlake (1980). The authors discuss a variety of control phenomena—involving reflexivization, adverbial participles, predicate agreement, and case agreement of predicate nominatives—in which the controller is typically the grammatical subject. For each, they claim that topic status (signaled by preverbal position) is a factor that affects the possibility of control in less typical instances. The authors conclude that subject and topic status (also animacy) are special cases of nominal prominence, which favors control. This is quite compatible with the account I will offer below. Let us next consider phenomena indicating a special relationship between subject and possessor. Some of these are well known. For instance, when a verb or a clause is nominalized, it is extremely common cross-linguistically for its subject or object to be specified periphrastically by means of a possessive construction: (2)

Booth's assassination [of Lincoln]; Lincoln 's assassination [by Booth]

Another such phenomenon, found in French, Norwegian, and numerous other languages, is for a subject nominal to simultaneously function as possessor (in this case of a bodypart object) even in the absence of any specific possessive marking, consider (3): (3)

a. b.

II leve la main. / Han rekker opp handa. 'He raises his hand.' J'ouvre la bouche. /Eg opnar munnen. Ί open my mouth.'

14

Ronald W. Langacker

Phenomena specifically relating topic and possessor may be less common, but they do exist. One example is the prevalence of nonsubject possessor pivots, as in (Id). Another interesting case is found in spoken French, as described by Lambrecht (1996). It is possible for multiple topics to occur preposed to the comment clause, as seen in (4a). Sentence (4b) is slightly different, however, since only the sister has any direct role in the comment clause. Strictly speaking, then, only sa soeur 'his sister' functions as topic with respect to that clause. The first preposed nominal, Pierre, instead functions as possessor for the second, sa soeur. While this looks just like a double-topic construction, in actuality it represents the combination of a possessive and a topic relationship. (4)

a. b.

Nicole, Pierre, eile ne I'aimepas. 'Nicole, Peter, she doesn't like him.' Pierre, sa soeur, je la deteste. 'Peter, his sister, I hate her.'

There are also phenomena where topic, subject, and possessor all figure crucially, suggesting that all three have some special affinity. Li and Thompson (1976) report that in Indonesian the pivot in a topic construction can only be the subject of the comment clause or a possessor of the subject: (5)

a. Ibu anak itu membeli sepatu. mother child that buy shoe 'That child's mother bought shoes.' b. Ibu anak itu, dia membeli sepatu. mother child that she buy shoe That child's mother, she bought shoes.' c. Anak itu, ibu-nja membeli sepatu. child that mother-his buy shoe That child, his mother bought shoes.' d. *Sepatu itu, ibu anak itu membeli. shoe that mother child that buy Those shoes, that child's mother bought them.'

Topic, subject, and possessor

15

For Korean, Cho (1991) points out that certain phenomena which normally involve subjects can instead be controlled by the possessor of an inanimate subject, provided that the possessor is part of a topicalized constituent. One such phenomenon is subject honorification: (6)

a. Apeci-ka Kim sensayngnim-ul mana-si-ess-ta. father-SUBJ Kim teacher-OBJ meet-HON-PAST-ASSR 'Honorable father met teacher Kim.' b. Halmeny-uy sayngsi-eyse-nun samsiptay-ka kacang gr.m.-POSS life-LOC-TOP thirties-SUBJ most hayngpokha-si-ess-ta. happy-HON-PAST-ASSR 'In honorable grandmother's life, the thirties were the happiest.'

Control of the reflexive marker casin works analogously. Finally, I note the prevalence in many languages of what are often called "double-subject constructions" (Langacker 1999a). Illustration is provided by Luiseno, a Uto-Aztecan language of southern California (Steele 1977): (7)

a. Noo=n no-puush konoknish. I=1SG:PRES my-eye green Ί have green eyes.' b. Xwaan=up po-toonav qala. Juan=3SG:PRES his-basket sits 'Juan has a basket.' c. Noo-up no-te' tiiwu-q. I=3SG:PRES my-stomach hurt-PRES Ί have a stomach ache.'

16 Ronald W. Langacker

These expressions have the following abstract form: (8)

[ noo-up outer subject

[ no-te' tiiwu-q ] ] inner subject inner clause

full clause Depending on the language, there is evidence for both the "outer subject" and the "inner subject" having subject status at some level, and also for the entire expression constituting a single clause. In the case of Luiseno, for instance, the tense/agreement clitic which regularly follows the first element of a clause appears in this construction after the outer subject (indicating that the full clause is in fact a single clause). Moreover, it can agree with either the outer subject (as does -n in (7a)) or else the inner subject (as does -up in (7c)). I analyze the outer subject in this construction as being both a subject and a topic with respect to the full clause. Furthermore, the construction is strongly associated with possessive relationships, prototypically involving bodypart relations. More specifically, the outer subject is usually understood as the possessor of the inner subject. Luiseno requires that this possessive relationship be explicitly marked by means of a possessor prefix. In other languages (e.g. Japanese and Korean) it remains implicit (cf. (3)). I am sure that this inventory can be greatly expanded. But even from this short, random list, it should be evident that something of interest is going on. To what can the intimate relationship among topic, subject, and possessor reasonably be attributed?

Topic, subject, and possessor

17

2. Some basic notions of cognitive grammar Naturally, I will approach this question from the standpoint of cognitive grammar (Langacker 1987a, 1990, 1991). Its central claim is that grammar is symbolic in nature. More specifically, lexicon, morphology, and syntax form a continuum fully describable as assemblies of symbolic structures. A symbolic structure resides in the pairing between a semantic structure (some kind of conceptualization) and a phonological structure (its semantic and phonological poles). It follows that all elements posited in grammatical description must have some kind of semantic value (often highly schematic). The possibility of a symbolic account of grammar presupposes a conceptualist semantics that properly accommodates construal, i.e. our multifaceted ability to conceive and portray the same content in alternate ways. Lexical and grammatical elements incorporate particular ways of construing conceptual "content", as an inherent part of their conventional meanings. Some facets of construal selected for brief illustration are level of specificity (conversely schematicity), perspective, ana prominence. Degrees of specificity are exemplified by lexical hierarchies such as (9). Each expression is schematic with respect to one that follows, which elaborates its coarse-grained specifications (i.e. spells them out in finer-grained detail). (9)

thing>creature>bird>bird of prey>hawk>red-tailed hawk

The meanings of grammatical elements tend to be quite schematic. To put it another way, the more schematic an element is semantically, the more inclined linguists are to consider it "grammatical" rather than "lexical". Naturally I view this as a false dichotomy. Perspective includes numerous factors, only one of which will be exemplified:

18

Ronald W. Langacker

(10) a. A big crack extends from the corner of the window up to the ceiling. b. A big crack extends from the ceiling down to the corner of the window. These sentences describe precisely the same, static situation— nothing actually moves. Yet intuitively they exhibit some kind of motion or dynamicity, and contrast semantically owing to its opposite directionality in the two cases. I attribute this to subjective motion on the part of the conceptualizer (i.e. the speaker, and secondarily the addressee), who mentally scans through the scene in one direction or the other in "building up" to its full conceptualization (Langacker 1986). Subjective or "fictive" motion is actually quite prevalent and varied (Matsumoto 1996a, 1996b; Talmy 1996; Langacker to appear). Two kinds of prominence are especially relevant to grammar. The first is profiling. Within its conceptual base (i.e. the array of content it evokes as the basis for its meaning), an expression designates (refers to, directs attention to) some substructure, called its profile. This is best seen from expressions that evoke essentially the same content but nonetheless differ in meaning. (a)

Base

iris

(")

Base

admire

pupil

admirer

O-O O-«O O--O tr

Im

Figure 1. Base and profile

As shown in Figure l(a), for instance, the conception of the human eye provides the common conceptual base for terms like im and pupil, which are certainly not synonymous (even if one's knowledge of their referents is limited to the configurational properties depicted). They contrast semantically by virtue of profiling—

Topic, subject, and possessor

19

referring to—different sub-parts of the base. Observe that profiling is indicated by heavy lines. Expressions can profile not only things (abstractly defined) but also relationships. An example is given in Figure l(b). The verb admire profiles a relationship where one individual manifests a certain mental attitude in regard to another.3 The noun admirer evokes the same conceptual content—one cannot apprehend its meaning without conceptualizing the process of admiring—yet it surely contrasts semantically with the verb. They have different meanings because the noun profiles a thing, specifically the participant displaying the mental attitude. Expressions that evoke the same content and profile the same relationship can also differ in meaning due to another kind of prominence, namely the degree of salience accorded to the participants in the profiled relation. For instance, as illustrated in Figure 2, above and below are semantically distinct despite invoking the same elements and profiling the same spatial relation (referentially, an above relationship is also a below relationship). (a)

above

(b)

below

?

Figure 2. Trajector/landmark alignment

Their semantic contrast resides in the fact that above takes the lower participant as a landmark for purposes of specifying the location of the higher one, whereas below does just the opposite. To capture such differences, I say that a relational expression accords focal prominence to one or more participants. It is usual for one

20

Ronald W. Langacker

participant—termed the trajector (tr)—to stand out as the primary focal participant in the sense of being the one the expression is concerned with locating, characterizing, or assessing in relation to others. A secondary focal participant, with respect to which the trajector is located or evaluated, is called a landmark (Im). An expression's grammatical class is not determined by its conceptual content, but by the nature of its profile (Langacker 1987b). A noun profiles a thing. A verb profiles a process, defined as a relationship whose evolution through time is rendered salient. As shown above in Figure l(b), for example, admirer is a noun because it profiles a thing, whereas admire is a verb because it profiles a temporally extended relation. Classes like adjectives and prepositions involve the profiling of non-processual relations. For instance, although the relationships profiled by above and below typically endure through time, this temporal dimension is not essential to their semantic characterization. Alice admires Bill

admires Bill

Alice

admires Figure 3. A symbolic assembly

Topic, subject, and possessor

21

A grammatical construction comprises a set of component symbolic structures that are integrated to form a composite symbolic structure. Integration depends on correspondences between elements of the component structures. Consider the lower right-hand portion of Figure 3, showing how the component structures admires and Bill are integrated to form the composite structure admires Bill. Their integration hinges on a correspondence between the profile of Bill and the schematic landmark of admires, which it serves to elaborate.4 Corresponding elements are superimposed and merged (or "unified") to form the composite structure. The composite structure admires Bill is therefore specific in regard to its landmark, which inherits the semantic specifications of the component Bill (abbreviated B). It is usual for the composite structure to inherit its profile from one of the components. The component structure which bequeathes its profile upward is called the profile determinant, identified by being enclosed in a heavy-line box. Admires is thus the profile determinant in admires Bill, for the composite expression profiles the same relationship as the verb. The profile determinant constitutes the grammatical head, at a given level of organization. Grammar resides in conventional patterns for combining symbolic structures into complex symbolic assemblies. I claim that such patterns are represented by symbolic assemblies that are partially or fully schematic, i.e. templates that embody the abstract commonality exhibited by particular complex expressions. These constructional schemas govern the formation of novel expressions. For instance, admires Bill instantiates a constructional schema that specifies the integration of a verb with its object by virtue of a correspondence between the former's landmark and the latter's profile. Constructional schemas will not be a prime concern here, however. A composite structure at one level of organization (in one construction) can in turn function as component structure at a higher level of organization (in a higher-order construction). In Figure 3, for example, we see that the composite structure admires Bill is integrated at a higher level of organization with the component Alice, which elaborates its schematic trajector to form the higher-order composite structure Alice admires BilL Symbolic assemblies with

22

Ronald W. Langacker

this character thus exhibit a kind of constituency. However, I have argued elsewhere (1995a, 1997a) that constituency of the classic sort is far less consistent or important than is assumed in most theories. Symbolic assemblies with the properties just described support what I believe to be a viable characterization of the grammatical notions subject and object: at a given level of organization, a subject is a nominal whose profile corresponds to the trajector of a profiled relationship, and an object, one whose profile corresponds to the landmark. Thus, considering both the "horizontal" and the "vertical" correspondences in Figure 3, Bill is identified as the object of admires, admires Bill, and Alice admires Bill, for its profile corresponds to their landmark. Likewise, Alice is identified as their subject, since its profile corresponds to their trajector. Observe that this characterization is not based on constituency. It is not configurational but rather conceptual, and alternate constituencies (e.g. a flat structure) yield the same grammatical relations. This illustrates the notion that constituency is neither fundamental nor essential to grammatical structure. It is a feature of symbolic assemblies that sometimes emerges, but not always, and is often flexible and variable when it does. The key work of grammar and semantics is done instead by correspondences, profiling, and other aspects of conceptual structuring.

3. Reference point phenomena As I view it, conceptualization is inherently dynamic (Langacker 1997b). By this I mean that it develops and manifests itself through processing time, the specific nature of its temporal development being essential to its value. This was shown quite clearly in (10), where the semantic contrast resides in the directionality of the conceptualized mental scanning through the scene. One aspect of dynamic conceptualization is our very general cognitive ability to direct attention to some entity for purposes of then establishing "mental contact" with another. This attention may be perceptual, as in (1 la). A more abstract example is given in (1 Ib).

Topic, subject, and possessor

23

(11) a. Do you see that boat out there in the middle of the lake ? There's an otter in the water just alongside it. b. Do you remember the surgeon we met at your sister's party? His wife just had quintuplets. In such cases I speak of a conceptual reference point (R) providing mental access to a target (T) found in its dominion (D). By serving as an initial focus of attention, the reference point tends to activate an array of associated entities which collectively constitute its dominion. This enables attention to shift to a secondary focus within it, the target (which activates its own dominion and may in turn function as a reference point). The components of a reference point relationship are depicted in Figure 4. This reference point ability has many linguistic manifestations. For instance, it plays a role in the sanction of definite articles, whose presupposition of contextual uniqueness can be explicated in terms of the speaker and addressee establishing mental contact with the same conceived entity (Langacker 1991:89-103). Consider examples like (12), where the definite article occurs even in the absence of previous mention: (12) a. My car is no longer dependable. The motor is just about shot. b. / would never buy this house. The roof leaks badly. The noun car evokes an established frame, including the standard knowledge that cars have motors, and that a given car has just one. Likewise, house evokes a frame that includes the notion of a (single) roof. Thus in (12a), my car directs attention to a specific car, which can thereby function as a reference point. The frame consisting of our general knowledge of cars specifies a set of entities accessible via this reference point, among which are its basic parts (motor, steering wheel, tires, etc.). Because the dominion defined in this manner contains just a single motor, the intended referent of motor in the second sentence is contextually unique, hence the definite article is used.

24

Ronald W. Langacker

C = conceptualize! R = reference point T = target D = dominion = mental path

Figure 4. A reference point relationship

Another reference point phenomenon is metonymy, where an expression's usual referent (R) is able to evoke its intended referent (T). The term Chicago, for example, refers primarily to a city, but in the proper context it might instead be construed as referring to any number of associated entities comprising its dominion. Possibilities include its populace, a sports team located there, the city government, the physical embodiment of a grant application its government has submitted, etc. (13) a. b. c. d.

Chicago is all excited about the Bulls. Chicago has just won another championship. Chicago was late in sending in its application. Chicago is right here in the stack between Dallas and Memphis.

I will mention only in passing a reference point analysis of pronominal anaphora worked out in great detail by Karen van Hoek in the context of cognitive grammar (van Hoek 1995, 1996, 1997a, 1997b). The basic idea is that a pronoun needs to occur in the structural dominion associated with a coreferential reference point salient in the current discourse space, whereas a full NP may not occur in such a context. I cannot go into any detail here. Instead we will concentrate on two other grammatical phenomena that I analyze in terms of reference points: possessive and topic constructions.

Topic, subject, and possessor

25

A central problem posed by possessive constructions is the extraordinary variety of relationships they are able to encode. The expressions in (14a) are but a small sample of the kinds of situations coded by prenominal possessives in English. While a case can be made for ownership, kinship, and whole/part relations being prototypical, it is by no means obvious that all the others can reasonably be considered metaphorical extensions from these. It seems as though virtually any kind of association is susceptible to being expressed in this manner. Yet that is not quite adequate. In and of itself, an associative relationship is symmetrical. As we see in (14b), however, possessive expressions are usually not reversible. (14) a. the boy's shoe; Jeffs uncle; the cat's paw; their lice; the baby's diaper; my train; Sally's job; our problem; her enthusiasm; its location; your candidate; the city's destruction b. *the shoe's boy; *the paw's cat; *the diaper's baby; *the destruction's city Apparently the semantic value of a possessive element or construction needs to be both abstract (hence broadly applicable) and asymmetrical (resulting in irreversibility). Elsewhere (Langacker 1993, 1995b) I have argued in some detail for a reference point characterization (see also Taylor 1996). Specifically, in a possessive construction the possessor can be described schematically as a reference point, and the possessed, as a target found in its dominion. In all the reference point phenomena described so far, the reference point and target have each been things. In pronominal anaphora, these two things are the profiles—taken as being referentially identical—of a pronoun and its antecedent. For metonymy, the things in question are alternate referents of the same nominal expression: its usual referent, and the metonymic referent intended in the context. In a possessive construction, the things in question are distinct and are designated by separate nominals, though one may be a constituent of the other. For instance, [[the boy]'s shoe] invokes a reference point relation between a particular boy and a

26

Ronald W. Langacker

particular shoe. The reference point is coded by the boy, and shoe specifies the intended target in its dominion. Hence the full nominal the boy's shoe singles out and profiles the shoe being referred to. Frame-based uses of the definite article, as in (12), also involve the referents of separate nominals (often a possessive construction could in fact be employed to describe their relationship, e.g. my car's motor). The difference is that the two nominals are separate and the reference point relation remains implicit. Whereas a possessive construction specifically encodes the reference point relationship, in (12) it is merely presupposed as part of the background knowledge that assures the contextual uniqueness required by the definite article. The reference point notion is not inherently limited to things, however. In particular, I propose that a topic construction expresses a reference point relationship between a thing and a proposition. This is diagrammed in Figure 5. The target proposition (given as a rectangle) is coded by a clausal expression minimally consisting of a thing (shown as a circle) engaging in some process (solid arrow).

Figure 5. A topic relationship

I believe this characterization properly captures the dynamic nature of topic constructions, wherein the topic is presupposed as the basis for interpreting the comment clause and thus has a certain conceptual priority with respect to it. A topic is sometimes described as a "mental address" to which the comment proposition is intended to be delivered. In terms of the reference point model, we can say that the topic evokes a domain of knowledge (its dominion) in which the target proposition is to be integrated. Equivalently, we can describe

Topic, subject, and possessor

27

the topic as providing mental access to a proposition that can be found in its dominion. Whether we speak of the proposition as (already) being located in the topic's dominion, or as needing to be incorporated into it, is simply a matter of whether we take the speaker's perspective or the listener's. Consider (15), which involves two reference point relations: (15) My car, the motor is just about shot. First, my car functions as reference point with respect to the motor, its dominion providing the contextual uniqueness required for felicitous use of the definite article. There are many motors in the world, but the speaker has a particular motor in mind, uniquely identifiable in relation to the speaker's vehicle in accordance with the frame semantics of car. For the speaker, the motor is already identified and already in the car's dominion. On the other hand, the listener may not have been thinking about motors at all, let alone any particular one. For the listener, the motor constitutes an instruction to conceive of an instance of the motor category, and signals its contextual uniqueness, but that alone does not identify it. Identification comes from interpreting the motor in relation to my car—it is a matter of integrating the conceived instance of motor in the dominion of that reference point. As a topic, my car is also a reference point with respect to the comment clause the motor is just about shot. For the speaker, this proposition is already known and ascribed to the proper domain—it belongs to the dominion of the speaker's car. It is "identified" in the sense that the speaker knows it pertains to a particular motor (associated with a particular car), as opposed to all the other motors an analogous proposition might describe. On the other hand, the listener may have no prior knowledge of the situation. The motor is just about shot might represent any number of distinct propositions, depending on which particular motor it pertains to, so to identify the specific proposition actually intended the listener needs instructions for delivering it to the correct mental address. The topic my car

28

Ronald W. Langacker

provides these instructions, enabling the listener to integrate the proposition in the proper domain of knowledge. Two points worth noting emerge from this example. First, a nominal put in focus by being invoked as a reference point for one purpose is quite susceptible to being simultaneously exploited as a reference point for some other purpose. Second, one reference point relationship may create the potential for another. In (15), for example, it is by virtue of the motor being in the car's dominion that the comment clause the motor is just about shot is applicable to the topic my car. If the motor were not construed as the car's motor, no basis for integrating the comment into the topic's dominion would be apparent.

4. Reference point analysis of subject and object I hope to have made a reasonable case for analyzing both possessive and topic constructions in terms of reference point relations. Not only does this make good sense for each type of construction individually, but it also accounts for the possessor-topic affinity: possessors and topics are each characterized as nominal reference points (with respect to things and propositions, respectively), and an entity invoked as a reference point for one purpose is likely to be invoked for another as well. Recall now that subjects show a strong affinity to both possessors and topics. This suggests that subjects should also be characterized as nominal reference points. I will therefore sketch an analysis along these lines and try to support it. I have described subjects and objects as nominal expressions which respectively elaborate the trajector and (if there is one) the landmark of a profiled relationship. A reference point analysis of subjects and objects thus implies that a relational participant achieves the status of trajector or landmark by virtue of serving some kind of reference point function. This proposal has initial plausibility given the independently developed descriptions of trajector/landmark status on the one hand, and a reference point/target relation on the other. Trajector and landmark are characterized as the primary and secondary focal participants of a profiled relationship. Similarly,

Topic, subject, and possessor

29

reference point and target are reasonably described as initial and secondary foci of attention, the first affording access to the second. As theoretical constructs, the notions trajector and landmark are first posited and strongly motivated for purposes of achieving minimally adequate semantic descriptions. An asymmetry in participant salience is the only evident basis for distinguishing the meanings of pairs of expressions like above and below, which evoke the same content and profile the same relationship (Figure 2). Many such pairs exist: in front of/in back of, precede/follow, like/please, etc., as well as actives and their corresponding passives. Metaphorically, I conceive of trajector and landmark status in terms of spotlights of focal prominence, a stronger and a weaker spotlight, that can be directed at various elements of a complex scene. By way of making this idea more precise, we can isolate the factors listed in (16). (16) a. The trajector and landmark of a profiled relationship are distinguished from other relational elements by the focal prominence conferred on them. b. Their prominence is asymmetrical, trajector and landmark being describable as primary and secondary focal elements. c. This prominence asymmetry has a temporal dimension; the trajector's primary focal prominence resides at least partially in its role as initial point of access. d. As successive foci, trajector and landmark are properly described as reference point and target, or equivalently, as first and second elements in a reference point chain. Let me briefly consider some of the evidence supporting each of these points, taking them in the order listed. That trajector and landmark—respectively coded by subject and object nominals—are distinguished by some kind of special prominence is, I think, fairly evident. Linguistic scholars are in virtually unanimous agreement that the subject and object are in some sense the "core" arguments of a verb, and typically they

30

Ronald W. Langacker

represent its only obligatory complements. Subject and object are especially active grammatically, which I take as being symptomatic of their special salience or accessibility. In particular, subject and object are the only elements with which verbs commonly show "agreement". I view such agreement as a matter of the greater specificity that is naturally associated with entities placed in the focus of attention, and the greater concern with clearly indicating their role in the event. Granted the focal prominence of trajector and landmark, their asymmetry in this regard is quite apparent. For one thing, subjects are generally more active and grammatically accessible than objects (specific cases are mentioned below). A more fundamental observation is that a relational expression may have a trajector without a focused landmark, but not conversely. A full justification of this statement requires numerous points of clarification and analysis that I cannot go into here (e.g. with respect to ergativity, as well as various types of impersonal constructions). For present purposes, let us simply note the basic pattern across languages for a single-participant clause to have a grammatical subject rather than an object, even when the subject's semantic role is the same as that of a transitive object: (17) a. Alexander broke the vase, b. The vase broke. c. *Broke the vase.

[subject and object] [subject only] [object only]

I believe one can ultimately sustain the generalization that a landmark presupposes a trajector, whereas a trajector does not presuppose a landmark. This follows from their respective characterizations as primary and secondary focal participants: a focal participant can be secondary only if there is also a primary one— otherwise it is the primary focal participant by default. Is it plausible in psychological terms to assume that our apprehension of a situation may involve two foci of attention, one primary and the other secondary? We can readily observe this ability in a non-linguistic example. Suppose you are watching a boxing

Topic, subject, and possessor

31

match. You can easily focus your primary attention on one of the boxers, observing his footwork, which hand he prefers, how well he defends himself, how many blows he absorbs, etc. For a given action, this focusing is possible whether the boxer in question has an active or a passive role; in either case, the action is viewed and interpreted in relation to him. But the other boxer has a prominent role as well. If you focus on boxer A and see him deliver a blow to boxer B, it is evident that B is also central to your conception. As a main participant in the observed event, he clearly stands out as a secondary focus (in contrast to the referee, the announcer, the audience, and so on). Experimental work by Tomlin (1995, 1997; Forrest 1996) yields striking evidence for characterizing subject as primary focus of attention. In multiple experiments with slightly different designs, the experimental subjects were made to observe a scene and immediately describe it with a simple sentence. The scenes consisted of either dynamic events or static situations, each with two salient participants. Attention was directed to one of the two participants (e.g. by an arrow flashed on the screen) just prior to presentation of the critical event or configuration, the interval being too brief for attention to wander in between. Tomlin's hypothesis (specifically for English) was that the speaker would choose the focused participant as the syntactic subject of the utterance produced. This proved to be the case, virtually 100% of the time. Tomlin's experiments also bear on (16c), the claim that the trajector/landmark asymmetry has a temporal dimension. Since attention was directed to the focused participant prior to presentation of the critical event or configuration, his results fit well with the notion that the trajector's primary focal prominence resides at least partially in its role as initial point of access. Pointing in the same direction is some recent experimental work indicating that agents are accessed more quickly than patients: (18)

"...When a prototypical causal relation can be perceived directly and when the participants in the event maintain their role for an extended time interval, agent information is accessed more rapidly than patient

32

Ronald W. Langacker information. The prominence of the agent over the patient in grammatical phenomena is parallelled by a privileged status of the agent at the level of perceptual processing. It is tempting to speculate that the observation that the agent is a prototypical subject and that the subject in canonical sentences surfaces before the object, can be traced back to the fact that, during visual event perception, agent information is available earlier for further processing than patient information" (Verfaillie and Daems 1996:144).

The observation that subjects canonically surface before objects needs a bit of elaboration. There is indeed a cross-linguistic tendency for subject to precede object in sentences reasonably considered unmarked (Greenberg 1963:77), and I take this as lending support to their supposed temporal asymmetry. This universal tendency is not however absolute. Moreover, subjects and objects have to be posited even in non-canonical sentences where the object precedes. The reason this is unproblematic is that temporal ordering in linguistic processing is not limited to a single, left-to-right pass strictly correlated with the phonological dimension of speech time. We can draw an analogy to eye movements, which do not trace a simple, continuous beginning-to-end path in the reading of a text. The overall progression through the text incorporates numerous saccades, both forward and back, resulting in momentary fixations at different points and piecemeal scanning through local domains. By the same token, speaking or understanding undoubtedly encompasses cognitive processing representing myriad levels and dimensions of structure, each with its own dynamics and to some degree independent of the linear progression of speech. Certainly it is common for all the elements within a local domain to be simultaneously available for conceptual integration whose inherent temporal directionality runs counter to their sequencing in speech time. Thus we can easily process either (19a) or (19b) despite the fact that the temporal sequencing of events correlates with speech time only in the former. (19) a. They finished the job before I got there. b. Before I got there, they finished the job.

Topic, subject, and possessor

33

Observe the greater naturalness of uttering (19a) without a pause, and (19b) with one. I interpret this as indicating that the correlation of event sequencing with speech time makes it easier to process the entire expression in a single "window of attention". With the reverse order, apprehension is more difficult because the subordinate-clause event has to be kept in mind until the main clause is processed; only then can the entire scenario be conceptualized in accordance with its inherent temporal dynamic. This more complex processing task is facilitated by placing each clause in its own attentional frame. Claiming that the trajector/landmark asymmetry has a temporal dimension therefore does not imply that a subject must invariably precede an object in terms of word order. It does however predict a general tendency for the subject to precede, since apprehension of the subject/object asymmetry can then be made to coincide with the temporal flow of speech and thus be effected without "backtracking". Analogous comments hold for reference point relationships in general. If a reference point affords mental access to a target and a basis for interpreting it, we should expect a strong tendency for this inherent directionality to correlate with linear ordering. We do indeed find that a topic normally precedes a comment, the antecedent precedes a pronoun, and a possessor precedes the possessed. Yet the opposite sequence can often occur and may have its own functional motivation (for "backwards" anaphora, see van Hoek 1997b). Presented with a pronoun, for example, we are able—within a local domain —to defer the resolution of its reference until we encounter a potential antecedent, at which point we apprehend their reference point/ target relationship in accordance with its intrinsic directionality. Further supporting (16c) are various phenomena naturally describable in terms of the subject providing a point of access for interpreting other clausal elements. One of these is the asymmetry inherent in equative sentences, as in (20): (20) a. My wife is the mayor. b. The mayor is my wife.

34

Ronald W. Langacker

These expressions are not equivalent, despite the fact that they profile the identity of the same two individuals. Thus (20a) would be more appropriate if we were talking about my wife, and (20b) if we were talking about the mayor. The difference emerges more clearly in sentences describing class membership: (21) a. A lynx is a feline. b. * A feline is a lynx. I analyze these expressions also as profiling the identity of two individuals. Here, though, we are no longer dealing with actual instances of the lynx and feline classes, but rather with arbitrary instances, i.e. instances "conjured up" for a specific local purpose and having no existence outside the mental space thus created (Langacker 1991:2.2).5 Given that the lynx class is a proper subset of the feline class, the well-formedness contrast between (2la) and (21b) can be seen as a consequence of the subject functioning as point of access for interpreting the profiled identity relation. In (2la), we initially focus on a lynx and describe it as coinciding with a feline; in view of the subset relationship, this will always be valid. On the other hand, (21b) initially focuses on a feline and describes it as coinciding with a lynx. Since the subset relationship is a proper one, the validity of such a description cannot be assumed. We come now to (16d), the suggestion that trajector and landmark are properly described as reference point and target, or—since a target, once focused, can itself function as a reference point—as the first and second elements in a reference point chain. In a way this follows as a matter of definition. If trajector and landmark are successive foci, and the trajector an initial point of access, the landmark must in some sense be accessible via the trajector, hence a target in its dominion. Certain other considerations reinforce this conclusion. It can first be noted that reference point relations do form chains, the target of one functioning in turn as reference point for the next. The clearest example is with possessives, as in (22a). The construction in (22b) invokes a chain of spatial reference points leading to the target of search.6

Topic, subject, and possessor

35

(22) a. my mother's cousin's friend's sister's lawyer b. In the kitchen, on the counter, next to the toaster sat the missing kitten. To a limited extent, topics also participate in chains, as we saw in (4). Beyond this, we can observe that subject and object have a privileged status with respect to grammatical constructions for which a reference point analysis is strongly motivated. One of these is pronominal anaphora, analyzed by van Hoek (1995, 1997a) as requiring that a pronoun occur in the dominion of a salient reference point with which it can be construed as being coreferential. It has long been recognized that subject and object status are especially important in this regard. In van Hoek's terms, the subject functions as reference point for all other clausal elements (including their modifiers), and an object, for all other clausal elements apart from the subject: (23) a. b. c. d.

My sister lost her keys. *She lost my sister's keys. I observed the baboons in their natural habitat. */ observed them in the baboons' natural habitat.

In short, subject and object act as initial and secondary reference points for anaphoric purposes. Van Hoek (to appear) has also shown that they do so for quantifier scope as well. Another reference point construction where trajector and landmark enjoy a privileged status is the use of possessives to specify the focal participants of a nominalized verb, as (24): (24) the army's destruction [of the city]; the city's destruction [by the army] The trajector and landmark of a profiled relationship are claimed to function as reference points for purposes of conceptualizing it. When a verb is nominalized, the process it profiles undergoes a conceptual reification, creating an abstract thing to which the profile is shifted; because it profiles a thing, the derived expression is classed as a noun

36

Ronald W. Langacker

(Langacker 1987b, 1991:1.2). If the trajector or the landmark of a process is a reference point with respect to it, and the process is reified to form an abstract thing, the reference point relation between them then holds between two things. That is precisely how a possessive construction is described: as coding a reference point relation between two things. Figure 6(a) shows the trajector of a process as a reference point, and that process as a whole (represented by a box) as the target accessed through it.7 The effect of nominalization is shown in Figure 6(b): by virtue of conceptual reification, the target process becomes an abstract thing, indicated diagrammatically by changing the box to an ellipse. Hence R and T are both things, just as for possessives (cf. Figure 4), except that R is necessarily intrinsic to the conception of T (which is true of many but not all possessives). Thus the proposed characterization of subject and object neatly accounts for the widespread use of possessive periphrasis with nominalizations. It further predicts (correctly I believe) that such periphrasis should be used primarily for the specification of trajector and landmark, and only by extension (if at all) for non-focal participants.

Figure 6. Trajector as reference point and the effect of nominalization

Topic, subject, and possessor

37

5. Levels of organization Even if accepted as basically correct, the characterization does still need to be sharpened, clarified, and situated in a broader context. Let us start by addressing the most fundamental issue. If trajector and landmark are first and second reference points, with respect to what experiential domain or dimension of structure do they so function? If they serve as initial and subsequent points of access, just what is it that they give access to? The clearest examples of reference point organization pertain to particular dimensions of linguistic structure and/or basic aspects of our everyday conceptual experience. For instance, a chain of locative reference points, as in (22b), pertains to the basic experience of finding things in space. Specifically, it reflects the common strategy of locating something in relation to salient landmarks, and also that of progressively narrowing the search. Possession pertains to identification, the problem of directing attention to a particular thing out of all those that populate our mental world. It exploits the typical circumstance that different instances of the same thing type are associated with distinct individuals whose prior identification enables them to function as reference points. Pronominal anaphora pertains to reference tracking, necessitated by the practice of referring to multiple individuals in a single discourse, and to the same individual on multiple occasions. Metonymy pertains to symbolization. In a conventional form-meaning pairing, the form is an initial point of access, hence a reference point. It prompts the listener to evoke a certain domain of knowledge—the conceptual base, which we can identify as the reference point's dominion—and focus attention on a particular element within it, the profile, which is thus a target accessible via the reference point. In the case of a metonymic interpretation, the expression's profile (its usual referent) functions in turn as a reference point affording mental access to its intended referent. The overall symbolizing relationship thus constitutes a reference point chain. With respect to what facet of linguistic structure or conceptual experience can we plausibly say that trajector and landmark function as reference points? I suggest that they are reference points with

38

Ronald W. Langacker

respect to the very process of conceptualizing a relationship. A relationship is conceptually dependent on its participants, i.e. it cannot be conceptualized without conceptualizing its participants, if only schematically. Since a relational conception presupposes and incorporates the conception of relational participants, the latter provide mental access to it, which makes them reference points. For reasons of space, I will mainly limit discussion to trajectors. By definition, the trajector is the initial focus, or starting point (cf. MacWhinney 1977; Chafe 1994). If we first conceive of a person, for example, we can then go on to conceive of that person engaging in some activity, or participating in some other kind of relationship. The set of relationships that a trajector potentially "anchors" in this fashion constitutes its dominion, and the profiled relationship—the one accessed via the trajector—is the target.8 This was sketched in Figure 6(a), where the dashed arrows lead to alternate relations potentially anchored by the reference point (trajector), and a box encloses the one that is actually conceived and profiled. The conceptualizer initially focuses on the trajector and uses that as a basis for conceptualizing the full relationship incorporating it. A resemblance can be noted between Figure 6(a) and Figure 5, the diagram given earlier for topic constructions. The resemblance would be even stronger if Figure 5 gave some indication of the pivot, i.e. the reference point's role in the target proposition. The similarity is not unexpected in view of the close connection between subjects and topics. Each can be characterized in terms of a reference point relation between a thing and a process. The problem then is to clarify the nature of their connection, especially since not every subject is a topic in the traditional sense. The key to the matter is to recognize that comparable relationships can be manifested at different levels of organization. It is commonly observed that topics can function as such for discourse units of varying size, e.g. for chapter, section, paragraph, and sentence, or for analogous units in speech. Likewise, a trajector has to be posited as part of the inherent semantic structure of expressions that profile relationships, ranging in size from single morphemes to full sentences. In Figure 3, for example, trajector/landmark alignment is

Topic, subject, and possessor

39

shown as part of the internal organization of admires, admires Bill, and Alice admires Bill. Roughly speaking, the higher the level of organization, the looser the connection may be between a nominal reference point and a process in its dominion. Consider first a "paragraph"-level discourse topic: (25) It's really hard to write a dissertation. You have to find a subject. Then you have to come up with some ideas and do lots of preliminary analyses. When you do the background reading, you find that most of those ideas have already been proposed and rejected. So you have to work for a number of years before anything viable starts to take shape. You have to worry about continued financial support. Then you have to satisfy five committee members with mutually incompatible notions about what you should be doing. You have to go through about seven drafts. Then ... The first sentence establishes the dissertation as the discourse topic. The following sentences are all integrated into its dominion, i.e. interpreted in relation to the dissertation, even though none of them mentions it explicitly. Structurally, the topic nominal is thus external to each of the comment sentences. Conceptually, its referent—the pivot—need not be a focal participant in any of these sentences, and in some it may be quite peripheral indeed (e.g. You have to worry about continued financial support). Consider next a sentence-level topic: (26) a. Jack, when I go to see him, he's never home, and he's always complaining that his friends ignore him. b. Jack, he's always complaining. The comment sentence may be complex, as in (26a), in which case a number of propositions are integrated in its dominion. Or it may be monoclausal, as in (26b), in which case the target consists of just a single proposition. Either way, the nominal establishing the topic is

40

Ronald W. Langacker

still external to the comment from a structural standpoint. Yet their connection is tighter than it is with a paragraph-level topic, as in (25), if only because just one sentence functions as comment.9 More importantly, the pivot tends strongly to be a central participant in the comment clause; note that in every clause in (26) it is coded by a subject, object, or possessor pronoun. There are of course occasional exceptions, e.g. (27), cited by Lambrecht (1996). (27) La mer, tu vois de l'eau. The ocean, you see water.' At the next lower level of organization is the situation where the nominal that establishes a topic is structurally internal to the clause expressing the comment proposition. Though different syntactic arrangements are possible, the one with the closest topic-comment integration finds the topic nominal functioning simultaneously as the grammatical subject or object. A well-known example is object preposing in English: (28) Bill I admire. Here clause-initial position marks Bill as the local topic, and at the same time this nominal directly instantiates the grammatical role of clausal object. Not only is the pivot a focal participant in the profiled relationship, but also topic and pivot coincide in the sense that the same nominal expresses them both. Is this option also available for a subject nominal? It is certainly possible in principle, but since English subjects normally occur clause-initially, that alone does not unambiguously mark a subject as clause-internal topic. Note however that expressions like (28)—where topic and pivot coincide—often occur in contrastive pairs, which makes the instantiating nominal accentually prominent. An example involving objects is (29a), where accentual prominence is indicated by capital letters: (29) a. BILL I admire, JACK I don't. b. ALICE admires Bill, JANET doesn't.

Topic, subject, and possessor 41

The same kind of contrast is possible with clausal subjects, as in (29b). Here at least I take the subject as establishing a clause-internal topic. Each clause in (29b) establishes the subject's profile as a salient reference point in whose dominion the clausal process is to be integrated. Although I have not investigated the matter thoroughly, I suspect that the contrast between "thetic" and "categorical" judgments (Kuroda 1972; Lambrecht 1987) resides in whether a focal participant is additionally established as a clause-level discourse topic. The clauses in (29b) represent categorical judgments concerning their respective subjects. Personally, I see no reason not to similarly regard the clauses in (29a) as representing categorical judgments concerning their respective objects. The thetic/categorical contrast is usually discussed in relation to non-contrastive examples, e.g. the responses in (30), recently described by Lambrecht (1994) in terms of "sentence focus" vs. "predicate focus": (30) a. What's wrong? My neck hurts. [thetic; sentence focus] b. What's wrong with your neck ? My neck hurts. [categorical; predicate focus] Whereas the response in (30a) accords no special discourse status to the subject, in that the entire sentence represents new information, the response in (30b) takes the subject as a local topic and makes a predication concerning it. I would say, then, that my neck in (30a) is merely the clausal subject, while in (30b) it is both the subject and the topic for that clause. The difference between the two ways of expressing categorical judgments is that in (30b) my neck has already been established as a topic by the preceding question, whereas in (29) the topic is established as such through means internal to the comment clause itself (initial position and/or accentual prominence). In other words, in (29) the topic and pivot roles are instantiated by the same nominal, within a single clause, while in (30b) this is done by separate nominals, in successive clauses.

42

Ronald W. Langacker

It should be evident that I am distinguishing between a clauseinternal discourse topic, on the one hand, and a subject or object, on the other. Although they may be conflated in a single nominal expression, as in (28) and (29), these represent distinct functions which have to be kept separate for analytical purposes. A discourse topic can serve as reference point for any number of comment clauses, and the nominal establishing it is usually external to each of them. The constructions in (28)-(29) represent the special, limiting case where the topic holds sway for just one clause, and moreover is established by a nominal that simultaneously has a central grammatical role in it. For the most part, however, the nominal expressing the topic is external to the comment clause. We can see this as being iconic. Whether a clause is to be integrated in the dominion of a particular reference point is basically an extrinsic matter that depends on contingencies of the current discourse. By contrast, a subject or object specifically functions as such only for a single clause, as an inherent part of its internal structure. It does serve as a reference point with respect to the clausal process, but the reference point role of a subject or object per se—qua subject or object—is independent of any particular discourse status. Instead, a subject (or object) specifies the trajector (or landmark) of a profiled relationship, which functions as reference point with respect to the very act of conceptualizing that relationship. This is the closest connection possible between a nominal reference point and a processual target. Access to a profiled relation via its focal partici-pants is intrinsic rather than contingent (note the contrast between Figure 6(a) and Figure 5). Trajector/landmark alignment is inherent in the meanings of relational expressions at any level of organization. At the lexical level, it is a facet of how we categorize the world in terms of conventionally recognized relation types. While there is of course an interaction, this categorization function is in principle distinct from the problem of constructing a coherent discourse, and certainly from the factors involved in any particular discourse. Agentivity is a particularly strong factor determining the trajector/landmark alignment of English verbs, for example. At the lexemic level, verbs that

Topic, subject, and possessor

43

profile events with a clearcut agent/patient asymmetry virtually always choose the agent as trajector (initial focal participant). In devising an inventory of lexically-coded categories well suited for describing what happens around us, one natural strategy is to consistently "take the perspective" of the agent in the sense of invoking it as the initial focal participant to anchor the conception of its interaction with a patient. In most any theory, it is recognized that the trajector/landmark alignment ("argument structure") inherent in a lexical item is projected "upward" to higher levels of grammatical organization. The default alignment imposed by lexical categorization can however be overridden if there is good reason to do so. That is where discourse comes in. There is no guarantee that a participant we wish to invoke as a starting point for discourse purposes will invariably coincide with the one selected as starting point by inherent lexical coding. When it does not we often resort to a special grammatical construction that imposes its own choice of subject, thereby ensuring that the current discourse topic also functions as point of access for conceptualizing a clause in its dominion. In (31), for example, the first sentence establishes my car as a local topic, and the speaker then finds it pertinent to mention its participation in certain events. (31) You really should think about buying my car. It was just repainted, and it drives very smoothly. The lexical items repaint and drive are readily available for describing the types of events in question. Since the repainting and the driving are to be interpreted with reference to the car, discourse will flow most smoothly if the pivot (coded by if) has the gram-matical role of subject, the natural starting point for conceptualizing the clausal process. The car however functions as patient with respect to these events, while the verbs select the agent as trajector and initial point of access. For this reason, the speaker resorts to the passive construction, in the case of repaint, and the "patient subject construction", in the case of drive (cf. van Oosten 1977, 1986). Both

44

Ronald W. Langacker

constructions confer trajector status on the semantic patient. The pronoun it which specifies the patient is therefore both the subject, affording access to the clausal process, and the pivot serving to connect each clause with the topic. Thus a clause, especially in its choice of subject, is a primary locus for the interaction between discourse considerations and the trajector/landmark alignment imposed by lexical categorization. As the expression of a single proposition, a clause is the smallest unit of discourse over which a topic holds dominion as the active reference point. At the same time, a clause is the largest unit for which we can posit a single overall trajector, either inherited by default from the lexical verb or imposed by a special construction like the passive. Whether the lexical default is sustained at the clause level depends on discourse factors like topic/comment organization, which arise at that level and above. In particular, discourse tends to flow most smoothly when the participant chosen as subject is also the pivot corresponding to a salient topic. It does so because a pivot is the point of connection between a topic and a comment clause in its dominion, while a clause's subject, by virtue of coding the trajector, is the point of access for conceptualizing the process it profiles. I have traced the progression from a paragraph-level discourse topic, at a high level of organization, down to the lexically determined trajector of a verb, at the lowest level. Viewed abstractly, all these levels are susceptible to uniform characterization: they all involve a reference point relation between a thing (the topic or trajector) and a process in its dominion. This abstract commonality is held responsible for the affinities noted earlier between topics and subjects. As we move from the highest to the lowest level, the connection between the nominal reference point and the processual target becomes progressively closer. The span of discourse over which the reference point holds dominion becomes progressively smaller: from a series of sentences, to a single sentence, to a clause, to a phrase, to just a verb. The reference point becomes progressively more intrinsic to the target process, both structurally and conceptually. Naturally, as we move from larger, loosely integrated structures to those which are smaller and more compact (at the

Topic, subject, and possessor

45

extreme, a single verb stem), their processing occurs on a progressively smaller time scale, and we are less and less likely to be consciously aware of either the reference point organization or its dynamic character. I would not expect the temporal dimension of trajector/landmark alignment to be easily accessible to introspection at the lowest level. Despite this commonality and continuity, an important discontinuity between topic and trajector has also been observed. It stems from the fact that the reference point relations at the two extremes have different functional motivations. A high-level topic helps to organize a discourse and render it coherent by evoking a single dominion for interpreting an open-ended series of target propositions. As full-fledged discourse entities, these propositions (generally expressed by finite clauses) make numerous specifications concerning the process they designate: they specify its participants (with subject and object nominals), its location in time and reality (via tense and modals), its precise nature (through modifiers), etc. Furthermore, the topic is often quite extrinsic to the target clause, which could just as well be assigned to some other topic's dominion. At the opposite extreme, the trajector/landmark alignment inherent in lexical verbs contributes to their function of supplying an inventory of process types with which to categorize events and situations. At the level of lexical categorization, the profiled relationship is not a full-fledged proposition, nor even a specific process instance, but merely a process type. Moreover, the trajector is wholly intrinsic to the target process, serving as a reference point with respect to the very act of conceptualizing it. Starting from the extremes, these two functions converge at the level of the clause—the smallest unit expressing a full proposition, and the largest one with a single overall trajector. A nominal that specifies a clause's trajector is its subject, as a matter of definition. It may or may not be a discourse topic as well. In a "thetic" judgment, like the response in (30a), it has no such status. If it is a topic, resulting in a "categorical" judgment, there are two possibilities. On the one hand, the topic may already be established in the discourse. In this case the subject nominal merely expresses the pivot, e.g. my

46

Ronald W. Langacker

neck in (30b). Alternatively, the clause itself may serve to establish the topic. In this case the subject nominal simultaneously expresses both topic and pivot, as in (29b).

6. Conclusion Though some will regard these proposals as speculative, I see them as both plausible and motivated by a variety of considerations. The notion that subject and object should be ranked as the first and second elements on a scale of prominence or accessibility is hardly unprecedented. Shibatani (1985) has made cogent observations pointing in this direction. Keenan and Comrie (1977) list subject and object at the top of their "accessibility hierarchy" describing the possibilities observed cross-linguistically for relative clause formation. Relational grammarians found good reason to label them "1" and "2" (Perlmutter 1983), and the same ranking is posited in other formal theories. Here we have concentrated on motivations for describing the parameter in question as some kind of prominence hierarchy, and specifically for characterizing subject and object as the first and second foci on a reference point chain inherent in the conception of a profiled relationship. Corroborating this account is the fact that subject and object clearly function as first and second reference points for purposes of pronominal anaphora. Furthermore, the characterization of subject and object fits nicely into a much broader view of conceptual and linguistic structure that has been emerging of late in cognitive linguistics from the efforts of numerous scholars working on diverse problems. There is a growing realization that inherently dynamic aspects of conceptual structure— involving notions like "access", "mental scanning", "fictive motion", "viewpoint shift", etc.—are both pervasive and fundamental to semantic and grammatical organization.10 More concretely, I have presented a substantial body of evidence suggesting that trajector and landmark are properly described as primary and secondary focal participants, and that the asymmetry in their degree of prominence has a temporal dimension. It is possible,

Topic, subject, and possessor

47

moreover, to see their focal prominence and reference point status as verb- and clause-level analogs of reference point phenomena observable at higher levels of organization, where temporality and sequenced access are more apparent. If we merely evoke the meaning of a transitive verb, it may not be introspectively obvious that we focus on its two participants sequentially in building up to the full conception of the profiled relationship.11 However, focusing and sequential access very similar to what is posited for verbs are quite evident at the level of clause-external topics, especially with multiple topics (cf. Lambrecht 1996): (32) Henri, ma cousine, il ne I'aime pas. 'Henry, my cousin, he doesn't like her.' Here there can hardly be any doubt that nominal referents are invoked as successive foci subsequently used to anchor the conception of a relationship in which they both participate. The same can be observed at the discourse level when topical participants are separately introduced and then described as interacting: (33) Jack came home late in the afternoon. There was Jill, working at her computer. He walked over and hugged her. In short, the conceptual operations I am positing are quite unremarkable even if, at lower levels of organization, they are too compressed to be readily discernible by introspection. A great deal of work is clearly needed before this proposal is widely accepted as a viable and well-supported characterization of subject and object. Required in particular are detailed analyses of individual languages of diverse types (e.g. Cook 1988; Kumashiro in preparation), as well as further experimental work approaching matters from the psychological and processing standpoint. Perhaps the proposed description will be enough to give investigators some idea of what to look for. At the very least, I suggest it offers a coherent, integrated view of topic-like elements at the lexical, grammatical, and discourse levels, in both synchronic and diachronic

48

Ronald W. Langacker

perspective. I further believe that it explains and clarifies the special relation among topic, subject, and possessor, accounting for both their affinities and their non-equivalence. Notes 1. A preliminary version of this paper appeared in Linguistic Notes from La Jolla 19, 1-28, 1998.1 would like to thank the anonymous reviewers of this paper for helpful comments. 2. In the discussion I will often concentrate on subjects: I believe the characterization can however be succesfully extended to objects. 3. Note that I generally use circles to represent things, and various kinds of lines and arrows to indicate relationships. 4. Observe that correspondences are represented with dotted lines. 5. Similarly, in the non-specific reading of Phil wants to pet a lynx, the referent of a lynx is an arbitrary instance conjured up to characterize Phil's desire and has no status outside the mental space representing it. 6. I analyze this locational reference point organization as being superimposed on the one defining trajector/landmark alignment. Each reflects its own natural path of access. 7. The box is included just for diagrammatic convenience - it is not distinct from the trajector and the process it anchors (given as a solid arrow), but is rather a notational device meant to represent their status as a single entity. 8. Recall that the topic my car in (15) is a reference point with respect to both the motor and the entire comment clause the motor is just about shot. In similar fashion, a trajector functions as reference point with respect to both a landmark (if there is one) and a profiled relationship. A landmark can be characterized as an initial target and a secondary reference point in a reference point chain required for a relationship's full conceptualization. 9. This does not preclude a sentence-level topic also being adopted as a higherlevel topic that remains active through a longer stretch of discourse. 10. See, for example, Elman (1990); Cutrer (1994); Fauconnier and Sweetser (1996); Matsumoto (1996a, 1996b, 1996c); Sweetser and Fauconnier (1996); Talmy (1996); Sweetser (1997); Langacker (1997b, 1999b); van Hoek (1997a, to appear). 11.1 attribute this to the small time scale inolved, the automatization of lexical meanings, and the general non-salience of conceptual components that are sublexical hence not individually symbolized (cf. Postal 1969).

Language as a cognitive network Richard Hudson

The paper develops the leading idea of Word Grammar and other "cognitive" theories of language, which is that language is a network. It reviews some of the consequences of this view: spreading activation, effects of conceptual distance, default inheritance, the unity of grammar and lexicon and, more generally, non-modularity; the unity of permanent and temporary representations, degrees of accessibility and binary relations. It then shows briefly how these ideas apply to two specific areas of language analysis: the contrast between polysemy and homonymy, and the treatment of regular and irregular morphology. The last section discusses Pinker's contrast between "mentalese" and "connectoplasm", and argues that the networks defined in this paper have all the symbolic qualities of mentalese, so maybe the mind uses "networks all the way down".1

1. Networks in Cognitive Linguistics and Word Grammar The main focus of this paper is one aspect of the approach to language which has come to be called "Cognitive Linguistics" (Lakoff 1987; Langacker 1987a, 1990; Taylor 1989), namely the widespread assumption that the whole of language is a network, a collection of units linked pair-wise by relations. These units are concepts—concepts such as "word", "the word word' or "the phoneme /p/"—and the leading claim of Cognitive Linguistics is that the concepts of language are organised in the same way as all other concepts. In short, "knowledge of language is knowledge" (Goldberg 1995: 5). General knowledge is widely believed to be structured as a network, so this view allows language to be integrated with other kinds of knowledge.

50

Richard Hudson

The network view contrasts most clearly with the view that knowledge is divided into distinct modules, each with its own formal properties and internal structure and with very limited links to other modules (Fodor 1983; Chomsky 1986; Pinker 1994, 1997). As far as language is concerned, the lexicon and the grammar form distinct modules, as do phonology, semantics and so on; and in an extreme version, each lexical item is held in a distinct "lexical entry" which is a kind of mini-module. Cognitive Linguistics exists as a rather general ideology about language, but it also includes three well-developed theories of language structure: Cognitive Grammar (Langacker 1987a, 1991, 1994), Construction Grammar (Goldberg 1995) and my own theory, Word Grammar2, as well as work by a wide range of other linguists among whom the most directly relevant is Bybee (1985, 1995a). Networks have played a central role in Word Grammar since its earliest days, as shown by the following quotation: "At the most general level, [Word Grammar theory] consists of [the following] generalisation: A language is a network of entities related by propositions." (Hudson 1984: 1) However, its historical roots are rather different from those of the other theories. It was influenced heavily by Chomsky's work, and especially by his stress on generativity, but it does not, in any way, grow out of Transformational Grammar. (In personal terms, I have never been a Chomskyan—witness Hudson 1976.) The main theories of grammar that contributed to it are also outside the Chomskyan mainstream: Systemic Functional Grammar (Halliday 1985; Martin 1992; Hudson 1971), Stratificational Grammar (Lamb 1966; Bennett 1992) and Dependency Grammar (Tesniere 1959/66). As will become apparent below, some of the leading ideas come directly from Cognitive Science. The undoubted similarities between Word Grammar and the other "cognitive" theories (Cognitive Grammar and Construction Grammar) are due to independent evolution. This in itself is encouraging for those of us who would like to believe that our theories are driven by the facts rather than by our personal histories. We shall see some differences between Word Grammar and the other

Language as a cognitive network 51

theories, but the most obvious differences are matters of notation and the substantive differences are relatively unimportant compared with the shared stress on networks. We shall now consider some of the consequences of building a theory of language in terms of networks.

2. Networks 2.1. Networks and activation For a cognitive linguist, one of the attractions of applying networks to language structure is that cognitive psychologists often model other parts of cognition as a network. One of the most widely accepted ideas in this field is "that memory connections provide the basic building blocks through which our knowledge is represented in memory [...] all of knowledge is represented via a sprawling network of these connections, a vast set of associations." (Reisberg 1997: 257-258). If language is like other parts of cognition, then we might expect at least some parts of language to constitute a network; and if some parts are a network, the most parsimonious theory is that they all are. This is the basis for the Word Grammar claim quoted above: "a language is a network of entities related by propositions." Cognitive Grammar and Construction Grammar also present language as a network of units (Langacker 1990: 2, 266; Goldberg 1995: 5). What does this claim mean, and how is it distinctive? The network view has several interesting and important consequences, but I shall highlight just a selection. First, knowledge must be declarative, rather than procedures which have an "input" and an "output". We know relationships, not "rules" for "doing" things. For example, we know that a past-tense verb has a suffix after its stem, and that the stem and suffix together comprise the whole of the verb; these are relationships among the concepts "past-tense verb", "its suffix", "its stem" and "its whole" which are diagrammed (using Word Grammar notation) in Figure 1 (see the next page). We do not know "how to form a past-tense verb"—such knowledge cannot be accommodated,

52

Richard Hudson

as such, in a network. A network is like a map, which is simply static; it is not like a set of instructions for getting from A to B. On the other hand, because of its declarative nature, the network may be used like a map as a guide to activity. When our goal is to find a pronunciation we can move from "past-tense verb" to the pattern "stem+suffix" by following these links in the network, and when our goal is to find a meaning we can move in the opposite direction. This is possible in a network because links may be activated like the links on a circuit board. A widely accepted model in cognitive psychology is spreading activation, whereby activation of one node spreads (as weaker activation) to the nodes to which it is directly attached. Spreading activation can only be assumed in a network; it makes no sense if knowledge is envisaged as a collection of rules or separate lexical items. whole

verbpast suffix Figure 1

It is perhaps in this sense that Langacker (1990: 15) states that "linguistic knowledge is procedural rather than declarative". Activation is central to Cognitive Grammar, and the "structured inventory of conventional linguistic units" (1990: 15) which are activated is closely analogous to the network of concepts in Word Grammar.

2.2. Conceptual distance and the networked lexicon The conceptual distance between two concepts can be measured as the number of intervening links—a notion which would be meaningless if the same information was presented as a list of rules

Language as a cognitive network

53

or propositions. We shall explore this consequence below, but meanwhile it should be noticed that distance is relevant because of spreading activation—the further apart two concepts are, the less likely they are to activate one another. The classic experiments on semantic priming (Reisberg 1997: 265-267; Swinney 1979) show that activation can spread across several links, so it does not require strict adjacency. For example, when presented with the word ink, subjects take less time to decide that it is an existing English word when it has been primed by an immediately preceding word than when it has not; and it is primed by pen, but not by king. Now the point to notice is that the link between the words ink and pen lies in their meaning: (the word) ink - (the thing) "ink"—"pen" - pen. This involves two intervening concepts, and perhaps more (depending on how direct the link between "ink" and "pen" is). Of course the conclusion is not that ink is related to pen but not at all related to king; after all, in a network everything is ultimately related to everything else. The difference lies in the relative distance between the concepts, which is shown in Figure 2.

pen

pen

MEANING

ink

king

WORDS

Figure 2

The same kind of spreading activation which leads to semantic priming also produces well-known effects in the recognition of words. These effects are the basis of the cohort model of speech recognition (Marslen-Wilson 1980; Marslen-Wilson and Tyler 1980, 1981) according to which the incoming sound-segments activate a decreasing cohort of candidate words: Id activates all the words that begin with /e/, then /!/ reduces the cohort to those that have initial /el/, and so on, until only one candidate remains—which is why a word like elephant can be recognised long before the end.

54

Richard Hudson

This evidence suggests, then, that when we hear a word it is embedded in a network of phonological and semantic links to other words. If this is so in speech perception, we should expect the same to be true of production—as it seems indeed to be, to judge by the evidence of speech errors. When a target word is replaced by another, the two are usually similar either in phonology or in meaning (Fromkin 1988), suggesting that these characteristics are "networked" (Emmory and Fromkin 1988: 145). It seems almost beyond dispute that the links from words to meanings and sounds involve a network of links between individual words (or lexemes) and individual semantic and phonological patterns—a far cry from the earlier view of the lexicon as a list of discrete lexical entries. It is less important to decide whether these links are just relationships or "rules" (e.g. the "correspondence rules" of Jackendoff 1997); indeed it is hard to know what the basis of such a distinction might be.

2.3. Default inheritance Large parts of the network are organised hierarchically on the basis of the most important relationship of all, the isa relationship (which is widely recognised in Artificial Intelligence—see Reisberg 1997: 280), Langacker's "categorization" (1990: 266), Goldberg's "instance link" (1995: 79). This is the relationship that underlies all classification, so saying that "A isa B" is the same as saying that A is an example or subclass of B: "Fido isa dog", and "dog isa mammal"; or "dog isa noun" and "noun isa word". The importance of this relationship lies in the fact that it allows "generalisation by inheritance". If "A isa person", and a typical person has a heart, then A must have a heart as well; so this fact can be stored just once, in connection with the general category "person", and inherited each time it is needed. A classic experiment by Collins and Quillian (1969) showed that it takes longer to verify a fact which is inherited than one which is not. For example, subjects take longer to verify

Language as a cognitive network

55

that a canary can fly than that it can sing, because the ability to fly is inherited from "bird" but canaries are kept partly for their singing so the ability to sing is a salient fact which is stored in relation to canaries; and it takes even longer to verify that a canary has skin, because this is inherited via "bird" from "animal". In short, the time taken for verification depends (among other things) on the conceptual distance between the inheriting node and the source of the information (Reisberg 1997: 267-270), as measured simply in terms of the number of intervening links that have to be traversed. Inheritance is clearly one of the effects of activation which is controlled by conceptual distance. One particular consequence is that properties are only inherited by default, i.e. in the absence of a more specific specification to the contrary. In short, the process is more accurately described as default inheritance—"normal mode inheritance" (Goldberg 1995: 73), "schematicity" with either "elaboration" or "extension" (Langacker 1990: 266-267). The overriding of defaults follows from the network model because the inherited property is always the one which involves the shorter route, i.e. the least conceptual distance. Take the canary example just quoted. The ability to sing is directly related (in the network) to "canary", giving a distance of just one link, whereas the ability to fly is separated from "canary" by two links: from "canary" to "bird", and from "bird" to "can fly". As we saw, retrieval is faster over shorter distances. Suppose now that the bird species in question was "ostrich". Unlike the typical bird, the typical ostrich cannot fly, so this fact must be stored directly in relation to "ostrich". If we now assume for simplicity that the link is labelled "flies?", with "yes" and "no" as possible values, it is clear that the value "no" will be retrieved before the value "yes" simply because of the difference in conceptual distance.

56

Richard Hudson whole

flies?

bird

yes

verb

flies?

ostrich

whole

no

'past

Figure 3

The relevant network is shown in Figure 3, alongside a similar network for verb morphology which shows why irregular morphology takes priority over regular. In this case, the relationship whose default value is overridden is "whole", which gives the word's fully-inflected form. The small triangle in the diagram signals the isa relationship iconically, since its (large) base rests on the supercategory and its (small) apex points towards the subcategories; so "ostrich" isa "bird", and GO isa

2A. The lexicogrammar continuum and "temporary competence" in performance If all generalisations are made (as in cognitive grammars) by default inheritance down an isa hierarchy, the facts which other theories express by means of general "rules of grammar" will be expressed as facts about relatively general categories near the top of the hierarchy. As in other cognitive theories, this leaves no place for a boundary between the "lexicon" and the rest of the grammar (Langacker 1990: 102; Goldberg 1995: 7); so if the "lexicon" is a network, the same must be true of "the rules of grammar". Moreover, Word Grammar syntax uses word-word dependencies rather than phrase structure, so the grammar generates just words and their dependencies. This means that all higher-level generalisations apply to single-word

Language as a cognitive network

57

categories—to the category "word" itself, or word-classes such as "noun" or "auxiliary verb". The result is that all our linguistic competence is organised around an isa hierarchy of word-types, with "word" at the top and individual lexemes and word-forms at the bottom—the "lexicogrammar" (Halliday 1985). Indeed the hierarchy is continued even further down during performance by the addition of two extra kinds of word. First, there are word-types which are inherited rather than stored. For example, it is generally agreed that at least some regularly inflected words are not stored; so a newly invented word obviously cannot have been stored by the hearer, and yet we can all understand inflected forms like wugs. Presumably the same is true at least for inflected forms of regular and rare words like disambiguate. In order to understand the form disambiguated we must assign it a new temporary concept which isa the lexeme DISAMBIGUATE and "pasttense verb" but which is also distinct from the word token which we are understanding (whose properties need not be exactly the same—it could be mis-spelt, for example). In Word Grammar notation, this word-type is called DISAMBIGUATED. The second kind of temporary addition to the network is the actual word token as recognised in processing. Suppose I am reading the following sentence: (1)

Pat shuffled uphill and Jo shufled [sic] downhill.

This contains seven word tokens, but two are tokens of the same type: SHUFFLED (a concept which is unlikely to be stored). The Word Grammar name for the tokens is based solely on position, so they are called words 2 and 6, "w2" and "w6", giving three temporary concepts: SHUFFLED and w2 and w6, each of which "isa" the former. The isa relationship allows mismatches, so this analysis is compatible with the mis-spelling shown. The hierarchy needed for these two tokens of shuffled is shown in Figure 4. In this view of language, then, there is no boundary between grammar and lexicon, but neither is there any formal distinction between competence and the structures created in performance.

58 Richard Hudson

Performance is the production of mental representations which "isa" those of competence. Moreover we can extend this theory into the area of learning by assuming that we build our competence by permanently storing some of the representations that we produce during performance (depending on factors such as frequency, salience and so on). In short, Word Grammar is a "usage-based" theory, just like the other two theories (Langacker 1990: 264; Goldberg 1995: 133). word

T

verb v SHUFFLE

past K

PERMANENT TEMPORARY

W1

Pat

W2

shuffled

W3

W4

downhill and

W5

W6

Jo shufled

W7

uphill.

Figure 4

2.5. Non-modularity More generally, there can be no modularity, in the sense of Chomsky (1986) or Fodor (1983). A network has no natural boundaries, which means that the network must have the same fundamental formal properties throughout. For instance, it could not allow spreading activation, or default inheritance, in some areas but not in others; nor could it accommodate a phonology which was procedural and a syntax which was declarative.

Language as a cognitive network

59

This is not to say that the network is completely uniform throughout. On the contrary, any cognitive network will reflect the boundaries of "nature", so that similar experiences will be more closely related to one another than to more different experiences. The result will be a pattern of subnetworks with more or less vague boundaries—one network for our family members, another for colleagues, a much larger one for people (which includes both these subnetworks), another for telephone numbers, and so on. These subnetworks are by no means isolated from one another—for example, many telephone numbers belong to some person, so numerous links go from individual numbers in the telephone-number subnetwork to individual people in the people subnetwork. The same organisation must surely be true of language, especially given the importance of isa hierarchies for bringing together units that are similar. For example, the relationships of syntax comprise a hierarchy of dependency-types (e.g. "subject", "complement", "adjunct") while those of morphology belong to a different hierarchy which includes "stem", "suffix" and "whole"; there is no direct link between these two hierarchies, and they probably each define a different subnetwork of the total grammar. Words belong to numerous overlapping subnetworks, but these need not correspond to the traditional divisions into linguistic modules. For example, brain damage can impair very specific areas of vocabulary such as all words which name things found inside the house (Pinker 1994: 314); no linguistic theory recognises a module for such words or their meanings, but such cases are to be expected if knowledge is indeed clustered into subnetworks as suggested above. Given our ability to see connections in the world, knowledge must be compartmentalised to the extent that our experience of the world itself is.

2.6. Accessibility Another well-established finding of psychology (and of ordinary experience) is that some concepts are more accessible than others. This has been particularly well documented for word retrieval; for

60

Richard Hudson

example, accessibility depends on frequency and recency (Yelland 1994: 5012). These notions can surely be accommodated in some way in a network model, but they again make very little sense in terms of rules and lexical entries. However, in a network-based theory we have to answer an important question: is it the nodes themselves, or their links, that vary in accessibility? In Cognitive Grammar, it is nodes (e.g. structural Schemas) that are described as having various degrees of "entrenchment" (Langacker 1987a: 59), but my own view is that at least some variation applies to the links between nodes, with some links being more easily activated than others. For example, when we hear a familiar piece of music it is often hard to recall its composer, regardless of how "entrenched" that composer may be; what is inaccessible is not the composer, as a person, but the link from the tune to that person. This is easy to explain in terms of a network in which each connection between two nodes has a "weight" which decides how much activation it passes from one node to the other (Luger and Stubblefield 1993: 517). If this is the explanation for all kinds of quantitative variation, then entrenchment will have been replaced by weighted links, but of course it may turn out that in some sense both nodes and links need to be weighted. This is clearly an important question for research. There is an interesting connection between these ideas and the findings of quantitative dialectology. This work (largely led by William Labov, cf. Labov 1972a, b, 1994) has uncovered a mass of quantitative links between speakers' social characteristics (sex, age, social class and so on) and their use of selected linguistic variables (e.g. the pronunciation of a particular phoneme or word). These findings have so far not been successfully integrated into any mainstream theory of language structure, but the network approach promises to be more helpful. Suppose a speaker S alternates between overt III and no /r/ in words \\kefourth ana floor, but regularly uses /r/ in about 35% of all tokens where /r/ is possible—a very typical finding of Labovian research (Hudson 1996, chapter 5). This behaviour must be based on S's knowledge of the local social structures, so let's assume that S believes /r/ is typically used by

Language as a cognitive network 61

members of social group G. We now have a simple subnetwork in S's mind which links the following concepts: speaker S, group G and pattern /r/. We can now explain S's behaviour as the result of the weightings of these links—for example, a weighting of 35% on the S-G link and 100% on the /r/-G link. It remains to be seen whether it is possible to build a single network model of this kind which integrates the figures for a larger number of linguistic variables (Hudson 1996: 248-252).

2.7. Relationships A concept's characteristics—its "definition"—all involve its relationships to other concepts, so there are no "one-place predicates". This follows from the claim that a network expresses all of knowledge, and not just those bits of knowledge that involve traditional two-place predicates. It is possible to imagine a cognitive system in which two-place predicates are expressed by network links while one-place predicates are attached directly to nodes as some kind of label; but this is not compatible with a pure-network theory such as the present one. This conclusion applies of course throughout the grammar, but its consequences are probably most controversial in semantics. In traditional logic, a verb such as yawn is a one-place predicate, not a relationship, so its argument structure is "yawn (x)" and the logical structure of a sentence such as John yawned, is "yawn (John)" where a fundamental distinction is made between "predicates" (yawn) and their arguments (John). This distinction is in any case undermined by various facts such as the possibility of nominalisations (John's yawning, or even John's yawn) which turn the putative predicate into an argument; but in any case it is incompatible with network models since "yawn" and "John" are both concepts which need to be represented in the same way, as nodes in the network. (To emphasise their similarity we shall use the nominalisation "yawning" as the name for the node which corresponds to yawn.) In this semantic analysis of John yawned, the

62

Richard Hudson

verb refers to a particular "act of yawning", which we may name simply "y", and John's role in the yawning is shown by a link from the "John" node to the "y" node. Similarly, traditional logic allows a general category such as "linguist" to act as a predicate with a single argument, so John is a linguist expresses a proposition based on a one-place predicate: "linguist (John)". In a network, however, it involves two distinct concepts linked by the "isa" relationship: "isa (John, linguist)". Adjectives and adverbs raise rather special problems because there is no well-developed Word Grammar analysis of them, but in principle the same remarks apply to them. Thus an adjective like bald cannot be treated as a one-place predicate but defines a relationship. In this case the argument is supplied by the noun to which the subject is applied—e.g. by John in John is bald—and the value is some degree of baldness, namely a degree which is above average which we can call simply "bald". In traditional notions, then, the meaning of John is bald is "baldness (John, bald)", and that of bald linguist is "x, isa (x, linguist) & baldness (x, bald)". Similarly for adverbs like loudly, whose argument is the verb to which they are applied as adjuncts. Take the following example: (2)

The bald linguist yawned loudly.

By assumption the semantic structure of this sentence involves the pairs of concepts and relationships shown in Table 1. Table 1 argument x (= the yawner) y (= the yawning) y x y

value linguist yawning yawn-er baldness loudness

relationship isa isa x bald loud

These relationships are shown in Word Grammar notation in Figure 5, which is a partial structure for the sentence. It does show part of the syntactic structure (below the dotted line), but it gives no

Language as a cognitive network

63

morphology and the semantics is incomplete. The main point is to show how apparent one-place predicates like "linguist", "yawning", "bald" and "loudly" can be shown as relationships between distinct concepts. (For more details see Hudson 1990, chapter 7, 1995.) yawn-er

t.b.l. yawning loudly .baldness . .

i

x

I

loudness

bald linguist\ yawning loudly A 1 linguist / yawning

loudly.

2.8. Conclusion We have reached two general conclusions about the network view of language. First, contrary to what I myself once thought (Hudson 1990: 15), it is not just a matter of notation, an alternative way of presenting information which could have been presented as a list of propositions. It is true that a network can be "unpacked" as a set of propositions, but the network model places specific constraints on propositions such as the fact that they have to be binary. Furthermore, networks have properties that a set of propositions does not have, including the possibility of measuring conceptual distances in terms of intervening links. Above all, only a network can serve as the database for a spreading-activation processor. The second general conclusion is that there is considerable evidence to support the network model—evidence from experimentation (e.g. priming), from speech errors, from neuropathology and even from quantitative sociolinguistics. All these sources of

64

Richard Hudson

evidence converge on a view of language which is very different from the more familiar "box" model, in which cognition contains a box of language which in turn contains distinct boxes of syntax, morphology, lexicon and so on. Instead, we can see language as part of the "sprawling network" of cognition, even though it may be separated from other parts by regions of relatively low-density connections; within this subnetwork we may be able to find further sub-subnetworks and so on, but every part of the total network is organised along the same lines (around isa hierarchies) and shares nodes with other subnetworks. In the rest of the paper I shall spell out some consequences of this view for specific areas of language analysis.

3. Polysemy and homonymy What light does the network view throw on the traditional distinction between polysemy and homonymy? This distinction is recognised as a major problem in lexical semantics, and indeed as a practical problem for lexicographers, because it requires a binary distinction where everyone agrees there is in fact a continuum. The distinction has to be made wherever we find a range of alternative meanings which can be expressed by a word of the same form, because we have to decide whether or not the meanings all belong to the same lexeme. According to the traditional view the two extremes of the continuum are occupied by polysemy (one lexeme, many meanings) and by homonymy (two lexemes, one form). A standard example of polysemy is BOOK, whose sense may be either concrete (volume) or abstract (book-contents); and a standard example of homonymy is BANK^^ and BANKriver, whose senses are unconnected. It is clear in these cases whether the senses are closely related (polysemy) or extremely distant (homonymy), but in between such examples lies a continuum of less clear cases: DRY (applied to wine or soil), CYCLE (bicycle or abstract repetition), THROW (stone or party), and so on. The assumption that lies behind all these discussions is that one of the functions of a lexeme is to show semantic relatedness. By

Language as a cognitive network

65

assigning two senses to the same lexeme, the analyst is guaranteeing that these senses are closely related. But as we saw earlier a network analysis of meaning already shows the degree of semantic relatedness in terms of "cognitive distance", which is a continuum. Presumably there is just one link between "volume" and "bookcontents", contrasting with perhaps ten links in the chain from "money-bank" to "river-bank", and some intermediate number for the chain from the dryness of wine to the opposite of wet. When the network already shows semantic relatedness so sensitively, it seems perverse to use the very blunt instrument of lexeme contrasts to impose a binary cut in the same information. Once freed from this obligation to show semantic relatedness, the lexeme can be used for the job that it is uniquely suited to perform: that of expressing the coincidence of syntactic, morphological and phonological characteristics. More strongly, lexemes must never be redundant—they must always "do some work" in the grammar. For example, suppose we recognise two BANK lexemes; they would be grammatically identical—same morphology (plural: banks), same word class, same lack of syntactic valency. Therefore one of them must be redundant because its meaning could be assigned to the other without losing any information at all in the grammar; so they should be collapsed into a single lexeme. In contrast, the verb BANK (as in He banked the money) must be a distinct lexeme in spite of its close semantic links to the "money-bank" sense of the noun. The result of this analysis is shown in Figure 6, where the division into lexemes follows the grammar and not the semantics. The crucial point about this diagram is that the meaning "money-bank" is much nearer, in terms of links, to the meaning "banking" than it is to the meaning "river-bank". The chain which links "money-bank" to "river-bank" is just a place-holder for a properly justified analysis, but it is hard to imagine a shorter chain between these two concepts; whereas the one from "banking" to "money-bank" is both short and well founded (banking something means putting it into a moneybank). This analysis shows exactly the same relationships as a different analysis in which the two traditional BANK noun lexemes are distinguished, so this distinction must be redundant, and wrong.

66

Richard Hudson

entity 'thing

place

T

institution terrain-feature

shop

τ/

T river-bank

money-bank

putting place

result

banking verb

BANK.

N

BANF

Figure 6

The conclusion, then, is that a network is a very good way of showing degrees of relatedness between words, whether this relatedness is a matter of sound, of syntax or of meaning. There is no point in trying to force a binary distinction between "similar" and "different" meanings in terms of one lexeme versus two.

4. Regular and irregular morphology Another area where the network view is relevant is in the treatment of inflectional morphology (Creider and Hudson 1999; Hudson 2000). The widely held "Dual mechanism" model makes a fundamental distinction between regular and irregular morphology on the grounds that regular morphology is handled by rule, while irregular morphology is expressed as default inheritance in an inheritance hierarchy (Marcus et al. 1992; Pinker 1999). Such a view is incompatible with network analysis because there is no formal distinction between rules and default inheritance: "rules" are expressed as defaults which are inherited unless blocked. Depending

Language as a cognitive network

67

on how we describe the information in a grammar, either everything is rules, or everything is network links; but there is no division into rules and links. The network view is therefore supported by the research which supports a model in which a single mechanism handles all morphology (Plunkett and Marchman 1993; Matcovich 1998). It remains to be seen exactly what kind of network model is needed—a connectionist model, or a more "symbolic" one such as I am proposing. As we saw in section 2.1-2.3, Word Grammar morphology is based on two functions, "stem" and "whole". A word's stem is the form that it inherits from its lexeme, and its whole is its fully inflected form; and inflectional morphology handles all differences between the two, while derivational morphology deals with relations between the stems of different lexemes. (A language in which every word's form is that of its lexeme has no inflectional morphology, and needs no distinction between "stem" and "whole".) A Word Grammar morphology consists of a set of "equations" each of which defines a different function: (3)

a. b.

The whole of a verbpast is its stem + its suffix, The suffix of a vert)past. is ed.

Equation (a) is the one diagrammed in fig. 1, where "verb^," is the name of a subclass of verb. If we assume that the past tense of WALK isa both WALK and verbpast, it follows by default inheritance that its whole must be walked. Diagram notation is particularly helpful when the right-hand side of the equation contains a functional definition (e.g. "its stem"). For present purposes, the most important fact about Word Grammar morphology is that exceptions, general defaults and subgeneralisations all have the same formal properties. We can illustrate this point through a tentative analysis of a fragment of the Norwegian verb system, focussing on past-tense morphology (Simonsen this volume). Most forms in a regular verb paradigm have a suffix, so by default a verb consists of its stem followed by some suffix, and in a past-tense verb, the most widely used suffix is et (e.g. KAST, 'throw',

68

Richard Hudson

past kästet). This defines the so-called "Large weak" class which we can call simply "verb". Like English, Norwegian also has a class of "Strong" verbs, which have no suffix and generally have a different stem-vowel in the past (e.g. DRIKK, 'drink', past drakk)\ for these we suppress the default suffix, and introduce another function, "stemvowel" which maps onto the last vowel of the stem. (We must take this mapping for granted here.) The stem's default vowel is replaced in the past tense by another which varies from verb to verb (with subregularities that we shall ignore). This prose description shows how both regular and irregular verbs are analysed in terms of the same combination of functions and equations. These rules are diagrammed in Figure 7. whole

verbstron stemvowel

ROP

DRIKK v

stemvowel

x verb t vero strong

past

verbsw: past suffix

Figure 7

Unlike English, however, Norwegian also has an important subclass of weak verbs which have a number of peculiarities, one of which is that their past-tense suffix is te or de instead of the default et (e.g. ROP, 'call', past ropte). These verbs are also shown in Figure 7 as the "Small weak" class (abbreviated here to SW). This sub-regularity is treated in just the same way as the regular and irregular patterns: we recognise a sub-class of verb and assign to it characteristics which

Language as a cognitive network

69

override the defaults. Sub-regularities such as this are a serious problem for the dual-mechanism theory because the "rules" for them are themselves exceptions. Needless to say, in spite of its apparent complexity Figure 7 is a considerable simplification. One area of analysis which it does not try to show is the experiential basis of the generalisations. As mentioned earlier, the total network is assumed to include not only these stable high-level generalisations but also at least some of the individual verb-tokens which were held in memory long enough to yield the general patterns consisting of lexemes and general wordtypes such as "verb:past". Moreover it is intended to accommodate the relative frequency and salience of the various links, though at present it is not clear exactly how this information can be formalised.

5. Conclusion: Symbolic networks The main purpose of this paper has been to argue that the whole of language (not just the lexicon) is a network with quite wellunderstood and consistent formal properties based on default inheritance and binary functions. This conclusion is completely compatible with the Cognitive-Linguistics assumption that language is part of the much larger network of general cognition. The one issue which we have not considered is how these networks fit into the contrast that Pinker (among others) draws between "mentalese", the symbolic language of thought, and "connectoplasm", thought as defined by a connectionist version of a neural network (Pinker 1997: 128). Pinker believes that the mind is divided between the two, though it may be unclear where the boundary lies: "Where do the rules and representations in mentalese leave off and the neural networks begin?" (1997: 112). The two systems have quite different formal properties: mentalese is symbolic, "digital", rule-governed, and uses concepts that are "crisply" defined, whereas connectoplasm is an "analog" network in which concepts are not defined (as such) at all, but related only by degrees of similarity induced from experience. Given these formal

70

Richard Hudson

differences, it is unclear how the two parts communicate, so alternatives should still be considered. One alternative is that it is "networks all the way down", as claimed in connectionism, but that at least some of the networks are much more like Pinker's mentalese. We can call them symbolic networks. Symbolic networks are widely used in cognitive modelling, and indeed Pinker himself recognises the role of so-called "semantic networks" within mentalese (1997: 87) as a way of displaying a prepositional database. The same is true of the networks of cognitive linguistics, including those that I have advocated in this paper. A symbolic network is symbolic, like mentalese, in that each concept has just one node and a finite and relatively fixed list of properties; but it is a network because the properties are defined by links to other concepts. It can be called the "language of thought" because it has the same formal characteristics as human language— namely, those of a network; thus it is the network that allows compositionality and recursion, which Pinker picks out as especially problematic for connectoplasm. As we have seen, it is merely a matter of terminology whether we call the links in a network "relations" or "rules", so we can say that a symbolic network is rulegoverned like mentalese. In short, symbolic networks have all the properties that Pinker attributes to mentalese, so we may be able to conclude that mentalese is a symbolic network.

Notes 1. I should like to thank Chet Creider, Joe Hilferty and an anonymous referee for comments on an earlier version of this paper. 2. The books about Word Grammar are Hudson (1984, 1990, 1995, 1996, 1998). The following web site holds downloadable material, including a 30-page introduction and a 100-page encyclopedia which is updated annually: http://www.phon.ucl.ac.uk/home/dick/wg.htm

Section II Morphological aspects of the verb

The Norwegian verb Rolf Theil Endresen and Hanne Gram Simonsen

This chapter aims at providing an overview of central aspects concerning Norwegian verbs. Starting with a short description of the Norwegian language situation and basic aspects of South East Norwegian phonology, a survey of the main simple and periphrastic verbal constructions and their use is given. Then follows an overview of the system of verb classes, focussing on their prototypical structure, and giving a description of their phonological patterning and frequency characteristics. As will be argued in several subsequent chapters in this book, it is necessary to take such features of a language into account when evaluating different theoretical approaches to morphology.

1. Introduction Since not all readers can be assumed to know Norwegian, this paper gives a short survey of the Norwegian verb, which plays a central part in the majority of the papers in this book. We shall begin with a few notes on Norwegian phonology, since morphological generalisations often mention phonological conditions; the description is based on South East Norwegian pronunciation (§2). We then give an outline of morphological and periphrastic constructions and their use (§3), followed by an overview of the verb classes (§4). The Norwegian verbal system is often presented in a way that clearly reflects the classical Latin background of our grammatical tradition. We want to present the system in a way that more clearly reflects its prototypical structure.

74

RolfTheil Endresen and Hanne Gram Simonsen

1.1. Nynorsk and Bokmäl But first of all we shall say a few words about the Norwegian language situation. Our language has two official written forms, called Nynorsk 'New Norwegian' ana Bokmäl 'Book Language'. The official Norwegian way of classifying them is not as two languages but as two 'language forms' (mälformer). Nynorsk is written by 80%. Let us emphasise that we are talking about written languages. It would be very misleading to say that 20% of the population speak Nynorsk and 80% speak Bokmäl. In fact most people speak their local dialects, which correspond in differing degrees to the Nynorsk and Bokmäl written standards. Bokmäl and Nynorsk are very similar; when they differ, Nynorsk words are preceded by a superscript N; "vere 'be' / \ori 'been', while Bokmäl words are preceded by a superscript B; \cere 'be' / \(Ert 'been'. Words that are the same in both written forms have no mark: er 'am, is, are' / var 'was, were'. Our last comment on the Norwegian language situation is that the names of these language forms—Nynorsk and Bokmäl—are both quite misleading. 'New Norwegian' is slightly more bookish than the 'Book Language' and the 'Book Language' is slightly newer than 'New Norwegian'.

2. A few notes on South East Norwegian phonology 2.1. Consonants Most South East Norwegian dialects have the 23 consonant phonemes depicted in Table 1. Note that the relationship between /!/, /[/, and l\l is slightly more complex than indicated in the table; a distinctive opposition between /!/ and l\l is found only after the vowels /a(:)/ and /o(:)/. In other contexts, [1] and [|J have a complementary distribution: [1] is found after the laminals /t d s n/, and [|J is found elsewhere, i.e., word

The Norwegian verb

75

initially, after all consonants but /t d s n/, and after all vowels but /a(:)/ and /o(:)/. In these "other contexts", laterals will also be written as/1/. Traditionally, l\l has a marginal status in the standard dialect, where IM is often used instead. In non-standard dialects, on the other hand, it is more stable. Table 1. The consonant phonemes of South East Norwegian dialects Consonants

labial

laminal

apical

dorsal

P b

t d n s 1

t et

k

Unvoiced plosive Voiced plosive Nasal Fricative Approximant Flap

m f υ

9 D 9 j

π. J

l

r

glottal

h

r

Tap

2.2. Vowels Most South East Norwegian dialects have the nine short and the nine long vowel phonemes in Table 2. Table 2. The vowel phonemes of South East Norwegian dialects Short Vowels

i e as

y

« 0

a

Long Vowels

u

i;

O

e: ae:

y:

«: 0:

u: o:

a:

Eleven vowels are front, /i i: y y: u «: e: 0 0: ae ae:/; four vowels are back, /u u: ο οι/; two vowels vary between front and back (socially, geographically, and depending on the quantity), /a a:/; and one vowel, Id varies between front and central (depending on stress).

76

RolfTheil Endresen and Hanne Gram Simonsen

2.3. Word accent A Norwegian word normally has only one accented syllable, but there are two distinctive word accents (or tönernes), in South East Norwegian dialects; a low accent and a high-low accent. They are not marked orthographically, but in our transcriptions we mark the low accent (L) with a grave accent £} and the high-low accent (HL) with a circumflex accent Q; the accent marks are placed above the vowel of the accented syllable.1 E.g., in Bokmäl, the two present indicative forms gjelder 'applie(s)' and gjelder 'castrate(s)' have different words accents—consider (1). (1)

gjelder 'applie(s)' gjelder 'castrate(s)'

/jeler/ /jeler/

When the last syllable of a word carries the accent, we can only have an L accent, with the exception of some verb-plus-particle constructions not discussed in this paper.

3. Morphological and periphrastic constructions 3.1. The three finite simple forms Norwegian has three finite simple verbal forms, where simple means non-periphrastic. In Table 3, these forms are illustrated with the verb D0MME 'judge, deem'. Table 3. The three finite simple forms of D0MME 'judge, deem' Meaning Mood Tense Indicative Indicative Imperative

Present Past —

Orthography

South East Norwegian pronunciation

dimmer d0mte d0m

/domer/ /dornte/

/dam/

The Norwegian verb

77

The finite forms have no person and number inflection. Notice furthermore that the indicative forms are aspectually neutral, as seen in (2) and (3). (2)

Present indicative ofARBElDE 'work' a. Arbeider Eva her? work.PRES Eve here 'Does Eve work here?' (Habitual) b. Ser du "ikkje/'ikke at vi arbeider? see.PRES you not that we work.PRES 'Don't you see that we are working?' (Progressive)

(3)

Past indicative ofBARBERE 'shave' a. Han barberte seg i gär. he shave.PAST himself yesterday. 'He shaved yesterday.' (Perfective) b. Eva barberte Adam da vi kom. Eve shave.PAST Adam when we arrive.PAST 'Eve was shaving Adam when we arrived.' (Progressive)

3.2. The three non-finite simple forms Norwegian also has three non-finite simple forms—cf. Table 4: Table 4. The three non-finite simple forms of D0MME 'judge, deem' Meaning Mood Tense — — —

— Present Past

Orthography

South East Norwegian pronunciation

d0mme ^mmande, *d0mmende d0mt

/dome/ 7d0mane/, Vd0mene/ /d0mt/

The present participle is a marginal form which not everybody agrees to including in the verbal paradigm. The infinitive and the past participle will be discussed in connection with the periphrastic forms.

78

RolfTheil Endresen and Hanne Gram Simonsen

3.3.

Types

offormatives

Now, take a look at Table 5, where we have tried to show the major types of formatives that are used in Norwegian verb inflection, by means of two verbs that both are pronounced /jele/ in the infinitive. The past indicative and past participle of 'castrate' has two alternative forms in Bokmäl: /jela/ gjelda and /jelet/ gjeldet (Nynorsk has only /jela/ gjelda). The choice between them is sociolinguistically or stylistically conditioned. In urban areas, /jela/ gjelda is roughly a low prestige working class form and /jelet/ gjeldet roughly a high prestige middle and upper class form. Both forms can be used in the written language—but not surprisingly, the high prestige forms dominate.2 Table 5. Two GJELDE verbs Infinitive /jlie/ HL,-e gjelde

Infinitive /jele/ HL, -e gjelde

GJELDE 'castrate' Present indicative Past indicative 7jelar/, Vjeler/ /jela/, Vjelet/ HL, -ar/-er HL,-a/-et

Past participle /jela/, Vjelet/ HL, -a/-et

"gjeldar, "gjelder

gjelda, "gjeldet

gjelda, "gjeldet

GJELDE 'apply (to)' Present indicative Past indicative 7jel/, Vjeler/ 7galt/, /jält/ L, -er L, -t, e ~ a

Past participle 7goli/, /jelt/ L, -t, e -

"gjeld, "gjelder

"galdt, gjaldt

"goldi, gjeldt

Just like in English, both suffixes and vowel alternations are used as formatives in Norwegian. In addition, as mentioned in section 2.3 above, Norwegian has a phenomenon not found in English: different word accents. Word accent also plays a role as a formative: in the present indicative, HL accent is typical of a weak verb, while L accent is typical of strong verbs. In §4 below we shall show more systematically how the formatives are employed in the different conjugational classes.

The Norwegian verb

79

3.4. Periphrastic forms I: The perfect Norwegian has many periphrastic forms, and no exact number can be given. Some highly grammatical!sed ones are those using the auxiliary verb HA 'have', which has the paradigm shown in Table 6. Table 6. HA 'have' Meaning

Orthography

S.E. Norw. pronunciation

Infinitive Present indicative Past indicative Past participle

ha har hadde hatt

/haV /ha:r/ /hade/ /hat/

This auxiliary combines with the past participle in the perfect forms, as in Table 7. The perfect forms are used roughly as in English. For some types of intransitive verbs—for example, motion verbs—the auxiliary "VEREI *v/ERE 'be' can be used instead of HA—cf. Tables 8 and 9. Table 7. Periphrastic forms: the perfect of KJ0PE 'buy' Meaning Orthography S.E. Norw. pronunciation Present perfect 'has bought'

har kj0pt have.PRES buy.PAST.PCP

/ha:r 90pt/

Past perfect 'had bought'

hadde kj0pt have.PAST buy.PAST.PCP

/hade ςα

Table 8. "VERE 1 *V£RE 'be' Meaning

Orthography

S.E. Norw. pronunciation

Infinitive Present indicative Past indicative Past participle

"vere, "vcere er var "vori, "van

/uae:re/ /ae:r/ /ua:r/ 7uo:ri/, Vuaet/

80

RolfTheil Endresen and Hanne Gram Simonsen

Table 9. Periphrastic forms: the perfect of REISE 'travel, go, leave' Meaning

Orthography

Present perfect 'has gone'

Past perfect 'had gone'

S.E. Norw. pronunciation

har reist have.PRES go.PAST.PCP

/hair raejst/

er

reist

/KIT raejst/

be.PRES

go.PAST.PCP

hadde reist have.PAST go.PAST.PCP var reist

be.PAST

/hade raejst/ /ua:r raejst/

go.PAST.PCP

3.5. Periphrastic forms II: the future The future can be expressed in several ways in Norwegian; the three presented in (4) are the most common ones in South East Norwegian. (4)

a. Using the simple present indicative "Dei/BDe "kjem / "kommer snart. they come.PRES soon They'll come soon.' b. Periphrastically; SKULLE 'shall' and VILLE 'will' as aux. "Eg/"Jeg skal "gjere / Bgj0re del i "morgon / "morgen. I shall do.lNF it tomorrow shall do it tomorrow.' Det vil bli regn ikveld. it will become.lNF rain tonight There will be rain tonight.' c. Periphrastically, with KOMME 'come' Det "kjem / "kommer til ä it

come.PRES

tO.PREP tO.INF.MARKER

bli "ein/'en fin dag. become.lNF a fine day 'It will be a fine day.'

The Norwegian verb

81

3.6. Periphrastic forms HI: the progressive As stated in section 3.1 above, the simple finite forms are aspectually neutral in Norwegian. This does of course not mean that there are no ways of expressing different aspects. For example, there are several ways of expressing a progressive meaning. One common way is found in (5). (5)

The progressive expressed by a position verb Han Nligg/ligger og Nh0yrer /h0rer pa ei / "en ny plate. he lie.PRES and hear.PRES on a new record 'He is (lying) listening to a new record.' N Jenta/"Piken "sit/sitter og les/leser. girl.the sit.PRES and read.PRES The girl is (sitting) reading.' Mennene star og diskuterer problemet. men.the stand.PRES and discuss.PRES problem.the The men are (standing) discussing the problem.'

From a grammaticalisation perspective, the use of position verbs to express progressive meaning has evident connections with the fact that position verbs are used much more in Norwegian than in English. Consider the sentences in (6), where position verbs are used in describing the location of inanimates. In English, it is more natural to use the verb BE in such cases, while this often sounds rather artificial in Norwegian. (6)

Position verbs used with inanimate subjects Boka/'Boken "Hgg/ligger pa hylla / "hyllen. book.the lie.PRES on shelf.the The book is (lying) on the shelf.' Koppen star pä bordet. cup.the stand.PRES on table.the The cup is (standing) on the table.'

82

RolfTheil Endresen and Hanne Gram Simonsen

3.7. Passive and related meanings A few words should be said about the expression of passive. There are three constructions that are traditionally classified as passive, one simple and two periphrastic ones—cf. (7): (7)

a. Simple passive: S-verb Boka/"Boken skal "omsetjas / "oversettes book.the shall.PRES translate.lNF.PASS til norsk. to Norwegian The book shall be translated into Norwegian' b. Periphrastic action passive: BLI 'become' Boka/'Boken blei/"ble "omsett / "oversatt book.the become.PAST translate.PAST.PCP til norsk. to Norwegian The book was translated into Norwegian' c. Periphrastic state passive: "VEREIBV/ERE 'be' Boka/BBoken er "omsett / Oversatt til norsk. book.the is translate.PAST.PCP to Norwegian The book is translated into Norwegian'

The so-called s-verbs do not always have a passive meaning. They may roughly be divided into three types: passive, reciprocal, and deponent, see Table 10. Table 10. S-verbs

Passive 'be bought' Infinitive Present indicative Past indicative Past participle

N

kj0pas / *kj0pes kj0pes (kjtfptes)

Deponent 'think'

m0tas/Bm0tes m0tes m0ttes m0ttes

"synas/Bsynes synes syntes syntes

N

Reciprocal 'meet'

The Norwegian verb

83

Deponent verbs are verbs with no counterpart without the -s suffix. Notice also that passive s-verbs generally are not much used outside the infinitive and the present indicative. In Bokmäl, passive s-verbs are used quite generally in the infinitive and the present indicative, while the past indicative form is marginal. In Nynorsk (where the sverbs may end in -st instead of -s) and in most non-standard dialects the infinitive is in general use, while the present indicative is marginal and the past indicative non-existent. Furthermore, it should be noticed that in Bokmäl (but not in Nynorsk), the infinitive and the present indicative of the s-verbs are practically always formally identical. As argued by Enger (1998b), they should not necessarily be analysed as two different forms of the verb in Bokmäl.

4. More on Norwegian verbal morphology 4.1. Nynorsk and Bokmäl revisited Although Nynorsk and Bokmäl are very similar, there are some morphological differences between them. Generally, Nynorsk has a slightly more complex morphology than Bokmäl, with more inflectional classes and a larger variation of forms. In the verbal morphology, the differences are clearest in the indicative forms, in particular in the present indicative, as illustrated in Table 11 on the next page.

4.2. Strong and weak verbs Norwegian verbal morphology is in many respects similar to that of other Germanic languages—for instance, it shares the basic distinction between strong and -weak verb classes. (We use these classical Germanic terms strong and weak, instead of the terms irregular and regular, since we find those a bit misleading: using the

84

RolfTheil Endresen and Hanne Gram Simonsen

term irregular about strong verbs precludes somehow the possibility that there may be important regularities among these verbs.) Table 11. Paradigms for some Norwegian verbs KASTE 'throw' Infinitive Present indicative Past indicative Past participle Present participle Imperative

KJ0PE 'buy'

/käste/ /kästar/ B /käster/ N

/kästa/ "/kästet/ /kästa/ "/kästet/ Tkästane/ Vkästene/ /käst/

/90:per/ /90pte/ /9 Weak (smaller class) /fait/ —» Strong (no change class)

It also implies that verbs with both a vowel change and a syllabic suffix are counted as weak despite the vowel change, as exemplified in Table 13, but they are irregular weak verbs, as compared to those

The Norwegian verb

85

without a vowel change (and there are few such irregular weak verbs, about 10 in Bokmäl and 20 in Nynorsk). Table 13 *VELGE "VELJE "SP0RRE "SP0RJE

'choose'

'ask'

Infinitive

Past ind.

Vuelge/ 7uelje/ Vsp0re/ "/spdrje/

Vualkte/ 7ualte/ /sp«:te/

—> Weak (smaller class) —> Weak (smaller class)

These examples also serve to illustrate the prototypical structure of the verb classes. We will return to this point shortly. For ease of exposition, we will restrict ourselves to Bokmäl in the rest of this paper. The inflectional patterns are based on the standard dialect with a South East Norwegian pronunciation, as spoken in Oslo. The basis for the description of Bokmäl verbs is The Bokmäl Dictionary (Bokmälsordboka, BO) from 1986. In BO are listed a total of approximately 4800 verbs, with a distribution of strong and weak verbs illustrated in Figure 1. (All numbers are approximations.) THE NORWEGIAN BOKMÄL DICTIONARY Total number of verbs: 4800 Strong:

200 (4%) V-class 2700 / (WL) (56%)

Weak: 4600 (96%)

./de/ \

C-class (WS)

150

1900 (40%) .

' Bother

^

-/te/

1750

(1 5%)

'

80 (1,5%)

^"ere verbs 1500 V.

other 250 (5%)

Figure 1. The distribution of weak and strong verbs according to Bokmalsordboka

86

RolfTheil Endresen and Hanne Gram Simonsen

The terminology V-class and C-class is used to refer to the suffixes in the two weak classes: The vocalic conjugation or V-class has past suffixes which are vowel initial, i.e., /-a/ or /-et/, while the consonantal conjugation or the C-class has past suffixes which are consonant initial, i.e., /-de/ and /-te/ in the past indicative and /-d/ and /-t/ in the past participle. This terminology was first introduced in Endresen (1992). Focussing instead on the number of members in the conjugational classes, Simonsen (1993, this volume) calls the vocalic conjugation the larger weak class (WL class) and the consonantic conjugation the smaller weak class (WS class). There is no general agreement in Norwegian grammars on what to call these classes. Faarlund, Lie, and Vannebo (1997: 492) refer to the vocalic conjugation as the 1st class of weak verbs and the consonantal conjugation as the 2nd class of weak verbs, while more traditional grammars, like Beito (1986: 302-308); Naes (1972: 181187) refer to the vocalic conjugation as the 1st class of weak verbs, while the consonantal conjugation is split up into the 2nd, 3rd, and 4th classes of weak verbs; this is in accordance with grammars of Old Norse, like Noreen (1923). Aasen (1864) divides the consonantal conjugation into the 1st and 2nd classes of weak verbs, while he refers to the vocalic conjugation as the 3rd class of weak verbs, while Western (1921) on the other hand refers to the vocalic conjugation as the 1st class of weak verbs and divides the consonantic conjugation into the 2nd and 3rd classes of weak verbs.

4.2.1. Strong verbs As already mentioned, the absence of a syllabic suffix is chosen as the basic feature singling out strong verbs. This entails that the strong past indicative form ends in an accented syllable—this is a salient phonological characteristic, and might be a good basis to form generalisations over, as indicated in the schema in Figure 2, where the symbol 'σ indicates an accented syllable.3

The Norwegian verb

87

' \\ [[PROCESS, PAST, INDICATIVE] / [...v...]] Figure 2. Strong verb schema

In addition, there are several other features characterising strong verbs, to a larger or smaller degree. Most salient and most often cited is what is often called a vowel change, i.e., a different vowel in the infinitive and past indicative stems. (These verbs differ as to which vowel they have in the past participle—the infinitive stem vowel, the past indicative stem vowel or a third vowel.) Another salient feature of strong verbs is that they carry the L accent (low tone) in the present indicative, as opposed to most weak verbs, which carry the HL accent (falling tone).4 This feature is exemplified in Table 14: Table 14 Infinitive

Weak Pres. ind.

Past ind.

Infinitive

Strong Pres. Ind.

7Y7TE 'look'

/titer/

/titet/, /tita/

SITTE 'sit'

/siter/

/sät/

LYSE 'shine'

/ly:ser/

/ly:ste/

FRYSE 'freeze'

/fry:ser/

/fr0:s/, /fr0js/

Past ind.

Lastly, as indicated in Table 12 above, a few verbs (some with a vowel change and some without a vowel change) form the past tense with a non-syllabic suffix. These four features characterising strong verbs are listed in Table 15, along with the numbers of verbs covered by each. Table 15. Features characterising strong verbs 1. L accent in the present indicative 2. Vowel change 3. Past indicative form ending in an accented syllable5 4. Past indicative with a non-syllabic suffix

151 143 137 13

88 Rolf Theil Endresen and Hanne Gram Simonsen

The three first features, which are all phonologically salient, are shared by a lot of the strong verbs, while the last (and possibly less salient) feature covers only a few. Features 1 and 2 are also shared by some weak verbs—this is why the number of verbs covered by these two features exceeds the number of verbs covered by feature 3, which was chosen as the defining feature for strong verbs. Finally, there is a small set of verbs sharing the three first features, but each having an irregularity of its own in the past tense inflection—the past indicative forms of these verbs may be characterised as more or less suppletive (e.g., gi 'give': /ji:/ - /g :/, g 'go': /go:/ - /jik/, ligge 'lie down': /lige/ - /lo:/). Thus, the category of strong verbs has a prototypical structure, illustrated in Figure 3 below. (b)

(a)



I

l\

[[PROC, PRES, IND] / [...V,...r]]

[[PROC, PAST, IND] / [...V...]]

(c)

(d)

[[PROC, PAST, IND],

'σ l·

I l\



[[PROC, PAST, IND] / [...VCt]]

Figure 3

In Figure 3, the square (a) represents the present indicative schema of the strong verbs, with a L tone on the accented syllable and an -r suffix. The square (b) represents the most schematic or general past indicative schema for strong verbs (identical to Figure 2 above), ending in an accented syllable with an L tone. The square (c) is the most prototypical instantiation of (b), from which it differs by having a different vowel (Vb) than the one in the present tense (Va), i.e., socalled vowel change. Finally, the square (d) represents the schema for past indicative forms with a non-syllabic suffix.

The Norwegian verb

89

Verbs likesitte 'sii'-satt.fryse ' freeze'-fr0s/fr0ys, stä 'stand'-sto, bite 'bite'-bet/beit share the 3 first features, and are typical strong verbs. Slightly less typical strong verbs are verbs like brenne 'burn'brant, sharing the three first features, but in addition having feature 4. Some verbs (so-called No Change strong verbs) like komme 'come'-fcom, I0pe 'run'-/0p share features 1 and 3, but not 2—a few of these, e.g. falle 'fair-fait, holde 'hold'-holdt also have feature 4. Table 16. Strong verb subclasses in Bokmäl Subclass 1

Inf. stem vowel

PT stem vowel

lil

/a/

lei lyl li'J

2

/ey

Examples

dritte 'drink' dette 'fall' jyn^e 'sing'

/drike/ /dete/ /syrje/

/dräk/ /dät/ /säry

bite 'bite'

/birte/ /fly:te/ ~/fty:te/

/be:t/ ~/bajf* /f!0:t/ ~/ft0jt/*

/dray

/drüy

/jloy

/jiuy

/suaSrge/

/suu:r/

14

/baSire/ /bey

9

9

~/aej/*

3

/yy

4

/ay

/ey

/0-J /u-J

lo-J /ae/ /aey

5

/ey /iy

/ay

/e/

6

/oy /o/ /a/

/ey /0y The Rest *

No change

No. of Vs

./fyte 'flow' rfra 'pull' le 'laugh' j/ 'hit' sverge 'swear' bcere 'carry' be 'ask' si 'say' /egge 'put'

/lege/

/bä:r/ /bä-7 /say /läy

save 'sleep' komme 'come' falle 'fall' hete 'be called' /0/?e 'run'

/sö:ue/ /körne/ /fäle/ /he:te/ /10:pe/

/sö:u/ /körn/ /fält/ /he:t/ /10:p/

se 'see' £« 'go' etc.

/sey

/söy /jik/

/g*/

40 33 27

5

With a few exceptions, the choice between a long vowel and a diphthong in these verbs is sociolinguistically or stylistically conditioned.

90

RolfTheil Endresen and Hanne Gram Simonsen

Subclasses 1, 2, and 3 all have a relatively high number of members. The phonologically most coherent classes among these are classes 2 and 3, with one single vowel in the infinitive stem, while class 1 is slightly less coherent since it has three stem vowels in the infinitive. As for the classes 4 onwards, they have a clearly lower type frequency and are much less phonologically coherent—thus the psychological plausibility of their status as subclasses may be questionable. To group the different strong verbs into subclasses, we have chosen a classification based essentially on the stem vowel in the past indicative form, i.e., a product-oriented definition. One subclass, however, is partly source-oriented—the No Change class, which is characterised as having the same stem vowel in the infinitive and the past indicative (i.e., verbs lacking feature 2 above). This classification is not uncontroversial (see for example the discussion in Enger 1998a);6 however, it has the advantage of being strictly synchronic, in addition to having a potential psychological plausibility as providing identifiable patterns. The strong verb subclasses in Norwegian are listed in Table 16, in an order of descending type frequency.

4.2.2. Weak verbs Unlike English, Norwegian has not only one class of weak verbs, but two. The two main classes are distinguished through the different suffixes they carry. As indicated in Figure 1 above, the larger of the weak classes in Bokmäl (the V(owel) class or the larger weak (WL) class) has approximately 2700 members. This class takes a past indicative suffix beginning with a vowel (-/et/ or -/a/), where the use of either of these is sociolinguistically or stylistically determined (cf. §3.3 above). The smaller one (the C(onsonant) class or the smaller weak (WS) class) has approximately 1900 members and takes two suffixes, both beginning with a dental consonant (-/te/ and -/de/). The use of these two suffixes is lexicalised for each verb.

The Norwegian verb

91

The WL class is clearly the most phonologically open class. With a few possible exceptions which we will mention below, stems with any form can take this suffix. It is also the class which is generally productive, in the sense that new verbs entering the language normally fall into this class. The WS class is phonologically less open. Two subclasses of this class can possibly be said to be phonologically defined (although the second subclass needs morphological information for its proper definition)—in any case each is phonologically very coherent. The first subclass includes verbs with a stem ending in the derivational affix -er /e:r/, which are inflected with the -te suffix, as exemplified in Table 17: Table 17. Verbs with a stem ending in -er. WS with -te suffix reparere 'repair'

/repareire/



/repar&te/

sortere 'sort'

/sot&re/



/so\e:\e/

markers 'mark'

/marke:re/



/marke^e/

This subgroup of verbs is actually very large, with approximately 1500 members (thus constituting nearly 80% of all the WS verbs, and 37 % of all verbs in Bokmal!), and since suffixation with -er is a productive process for verb derivation in Norwegian, the class is growing. However, generally these verbs have a very low token frequency. (Aspects of these verbs are treated by Gundersen this volume.) The second subclass, exemplified in Table 18, contains all weak verbs with a monosyllabic infinitive form. These count approximately 70 verbs, and all take the -de suffix. Table 18. Weak verbs with a monosyllabic inf. form: WS with -de suffix

sy 'sew' na 'reach' tree 'thread'

Isy'J /no:/ /treV

— — —

/syde/ /node/ /trede/

92

Rolf Theil Endresen and Hanne Gram Simonsen

A third, irregular subclass of WS verbs are verbs with a vowel change in addition to the syllabic suffix in the past indicative (cf. Table 13 above). The rest of the WS verbs, approximately 350 verbs, form a phonologically relatively open class, but less so than the WL class. E.g., there are restrictions on the set of stem final consonants for WS class verbs: verbs with stem final /b {ιχ J il do not occur in this class, and verbs with stem final consonant clusters are very rare in the WS class (limited to the two verbs hilse 'greet' - hilste and tenke 'think' - lenkte). But there are also some more subtle phonological distinctions governing the distribution of weak verb stems, helping in determining their class membership. As pointed out by Hagen (1994) and Bjerkan (2000), there is a strong correlation between the length of the infinitive stem vowel and verb class. In the WL class, approximately 80% of the verbs have a short stem vowel, and only 20% have a long stem vowel, while the opposite is the case for the WS class (these counts exclude the -ere verbs). In addition, in some cases vowel quality may be a good indicator of class membership: In the WL class, the short /a/ is the most frequent stem vowel, occurring in 20% of the verbs, while there is only one single verb with a short /a/ in the WS class (kalle 'call' - kalte). In the WS class, on the other hand, /0:/ is the most frequent stem vowel, while there are very few WL verbs with this vowel. A summary of these phonological characteristics of the two weak classes is given in Table 19 on the opposite page. More research needs to be done for us to be able to characterise the phonological patterning of the weak classes properly. Many of the phonological characteristics we have identified so far are only tendencies, thus not supplying clear-cut class boundaries. However, these tendencies may have implications for processing and for class formation during acquisition—questions addressed in the papers by Bjerkan, by N0klestad, and by Simonsen (all this volume).

The Norwegian verb

93

Table 19. Phonological characteristics of the two weak classes WL class

WS class

Essentially open

less open

Generally productive

subclasses, possibly phonologically defined: • weak verbs with a monosyllabic infinitive • -ere verbs (with stem ending in -/e:r/)

80% short stem vowel 20% long stem vowel

80% long stem vowel (excluding -ere verbs) 20% short stem vowel

/a/ most frequent stem vowel: (20% of verbs)

/a/ extremely rare as stem vowel: (one verb only)

I0'J very rare as stem vowel

I0'J most frequent stem vowel

non-restricted set of stem final consonants

restricted set of stem final consonants

wide range of stem final consonant clusters

very few verbs with stem final consonant clusters (two verbs only)

5. Summary In this paper, we first gave a short description of the Norwegian language situation, with the two official language forms, Bokmäl and Nynorsk, followed by a short summary of South East Norwegian phonology. Then followed a survey of the main simple and periphrastic verbal constructions and their use. There are three simple finite forms (present indicative, past indicative, and imperative), three simple infinite forms (infinitive, present participle, and past participle), and various periphrastic forms, expressing, i.a., perfect, future, progressive and passive. Finally, we gave an overview of the system of verb classes. We focused on the phonological patterning of the classes, taking into account their frequency characteristics and their prototypical structure. Norwegian verbs can be divided into strong and weak classes. Strong verbs have no suffix or a non-syllabic suffix in the past indicative, while weak verbs have a syllabic suffix. We divided the strong verbs into six subclasses of varying regularity, plus one

94

RolfTheil Endresen and Hanne Gram Simonsen

subclass of irregular strong verbs; the classification is based essentially on the stem vowel in he past indicative, i.e., a productoriented definition. A salient feature of the Norwegian verb class system is the fact that it has not only one, but two main weak classes: 1) The V(owel) class or The larger weak class (WL), and 2) The C(onsonant) class or The smaller weak class (WS), the first of which is phonologically more open than the second. This fact plays a central role in several of the papers in this volume.

Notes 1. In the literature on Norwegian, the word accents are often called Tönerne l and Tönerne 2, and marked accordingly with a raised number ' or 2 in front of the accented syllable. In South East Norwegian, Toneme 1 corresponds to the L accent (low tone), while Toneme 2 corresponds to the HL accent (falling tone). 2 According to Bokmälsordboka, the verb meaning 'castrate' also has the past indicative form gjeldte /jelte/, but we have ignored forms that to our knowledge are not current in South East Norwegian dialects. Both Nynorsk and Bokmäl accept considerable variation. 3. In accordance with Norsk referansegrammatikk The Norwegian Reference Grammar' (Faarlund, Lie, and Vannebo 1997: 480), approximately 10 very irregular verbs (modals plus a few other irregular verbs) are counted as different from from both weak and strong verbs and form a separate group—socalled preterito-presentic verbs. 4. A few weak verbs have the L accent in the present tense, and there are even a couple of strong verbs which have the HL accent in the present tense. However, these discrepancies are due to diachronic movement between inflectional classes. 5. The total number of strong verbs counted here is 137. This number is lower than the approximate number of strong verbs cited from Bokmälsordboka to be approximately 200—cf. the figure in (21). The reason for this is that our verb classification differs slightly from the one assumed to be underlying the verb class assignment in Bokmälsordboka; furthermore we have only included noncompound strong verbs, and we have excluded a few verbs that we do not consider strong in Bokmäl any longer. 6. Interestingly, until recently, Western (1921) was the only description of this type in the available literature on Norwegian verb classes. However Faarlund, Lie, and Vannebo (1997) have adopted a similar solution.

Building blocks or network relations: Problems of morphological segmentation Helge Gundersen

Traditionally, morphological analysis is identified with an exhaustive segmentation into discrete "chunks". This building-block model of morphology is contrasted with a cognitive-linguistic network conception of morphology, which may be viewed as its diametrical opposite. The building-block model induces several unnecessary problems, and at the same time overlooks certain morphological phenomena. Three segmentation issues are presented: phonological residues (cranberry morphs), morphological ambiguity (overlapping items), and submorphemic relations (morphologically complex items lacking constituents). The remainder of the paper focuses on problems of synchronic morphological segmentation in one part of the Norwegian language, namely in a group of verbs derived by the suffix-er.1

1. Two conceptions of morphological analysis 1.1. The building-block model In grammatical theory, it is common to separate grammar from general skills and knowledge. A morphological analysis will then not be a study of how the language user handles the words in question. Instead, it may be seen as a study of how grammar constructs its own output. Morphology is only seen as a matter of objective composition independently of the language user. The foundation of this way of thinking is objectivist philosophy (cf. Lakoff and Johnson 1980: 195209; Lakoff 1987: 157-184; Gundersen 1997). Lakoff and Johnson put it this way (1980: 204):

96

Helge Gundersen According to the myth of objectivism, objects have properties in and of themselves and they stand in relationships to one another independently of any being who understands them. When words and sentences are written down, they can be readily looked upon as objects. This has been the premise of objectivist linguistics from its origins in antiquity to the present: Linguistic expressions are objects that have properties in and of themselves and stand in fixed relationships to one another independently of any person who speaks them or understands them. As objects, they have parts—they are made up of building blocks: words are made up of roots, prefixes, suffixes, infixes; sentences are made up of words and phrases; discourses are made up of sentences. Within a language, the parts can stand in various relationships to one another, depending upon their buildingblock structure and their inherent properties. The study of the buildingblock structure, the inherent properties of the parts, and the relationships among them has traditionally been called grammar.

Langacker (1987a: 452) speaks of a building-block metaphor, "which sees the meaning of a composite expression as being constructed out of the meaning of its parts simply by stacking them together in some appropriate fashion". It seems clear that the notion of a buildingblock metaphor may be extended from the semantic compositionality of expressions (which is Langacker's point) to the morphological structure in general, cf. Lakoff and Johnson's mention of roots and affixes in the quote above. In structuralist and generative linguistics morphology is to a large extent treated as an exhaustive segmentation into discrete constituents seen as "building blocks". This approach is depicted by the simple algebra in (1), in which A and B are the building blocks of the composite structure C. (1)

A+B=C

A complex word is then seen as made up of its constituents, and the aim is to find these building blocks and the rules that combine them. When the building-block model is made serious use of in morphological theory it must be viewed in the light of underlying objectivist assumptions. Complex words are seen as constructed independently of the language user, and their morphological constituents then naturally become the building blocks.

Problems of morphological segmentation

97

One important manifestation of this way of thinking is the morpheme-based structuralism of Bloomfield and his followers. This tradition is still taken as a point of departure in much of contemporary morphology. Matthews (1993: 98-110) has examined not only standard introductory textbooks in linguistics, such as Fromkin and Rodman (1988), but also influential generative treatments, such as those by Halle (1973), Williams (1981), McCarthy (1982), and Lieber (1989): [T]he theory of strict lexicalism is the most important new development in morpheme-based morphology. Of the deeper hypotheses on which it rests, one was an innovation: that derived words must pass through a lexical filter. The other can be traced back to the origins of structuralism, half a century earlier. This second assumption seems to have been automatic. When I read Lieber's paper on the percolation of morphosyntactic features (1989), I pencilled in my copy of the journal: OK but why should the features originate in morphemes in the first place?' There might have been substantive reasons, but Lieber does not give them. It was simply taken for granted from the outset. Sixteen years earlier, Halle had begun his 'Prolegomena' by pointing out that speakers know the words of their language. They also know' that a word like trans-form-at-ion-al 'is composed of the morphemes shown and that these five morphemes cannot be concatenated in most other orders'. Such knowledge must be represented in a grammar: 'hence', he goes on, 'the assumption has been made quite generally that a grammar must include a list of morphemes as well as rules of word formation' (1973: 3). It had been made indeed, first by Bloomfield and then by those who followed him. (Matthews 1993: 110)

Jensen, in his introduction to generative morphology (1990), bases much of his exposition on Nida's (1949) principles for the identification of morphemes, which was an important contribution to bloomfieldian morphology. Before quoting Nida's principles, Jensen himself (1990: 21) states a principle not supplied by Nida which, I believe, appears to give much of the essence of the building-block model of morphology. Jensen is not necessarily representative of generative morphology (cf. Marantz 1992), but I mention his principle here, as I think it will serve well as an introduction to Nida's principles:

98

Helge Gundersen

(2) A word must be exhaustively divided into morphemes. Nida's (1949) principles are quoted here without commentary: Principle 1 Forms which have a common semantic distinctiveness and an identical phonemic form in all their occurrences constitute a single morpheme. Principle 2 Forms which have a common semantic distinctiveness but which differ in phonemic form (i.e. the phonemes or order of the phonemes) may constitute a morpheme provided the distribution of formal differences is phonologically definable. Principle 3 Forms which have a common semantic distinctiveness but which differ in phonemic form in such a way that their distribution cannot be phonologically defined constitute a single morpheme if the forms are in complementary distribution in accordance with the following restrictions: 1. Occurrence in the same structural series has precedence over occurrence in different structural series in the determination of morphemic status. 2. Complementary distribution in different structural series constitutes a basis for combining possible allomorphs into one morpheme only if there also occurs in these different structural series a morpheme which belongs to the same distribution class as the allomorphic series in question and which itself has only one allomorph or phonologically defined allomorphs. 3. Immediate tactical environments have precedence over non-immediate tactical environments in determining morphemic status. 4. Contrast in identical distributional environments may be treated as submorphemic if the difference in meaning of the allomorphs reflects the distribution of these forms.

Principle 3 is simplified by Nesset (1998b: 74) in this way: "Sets of morphs realizing the same content belong to the same morpheme if they are in complementary distribution." Principle 4 An overt formal difference in a structural series constitutes a morpheme if in any member of such a series, the overt formal difference and a zero structural difference are the only significant features for distinguishing a minimal unit of phonetic-semantic distinctiveness.

Problems of morphological segmentation

99

Principle 5 Homophonous forms are identifiable as the same or different morphemes on the basis of the following conditions: 1. Homophonous forms with distinctly different meanings constitute different morphemes. 2. Homophonous forms with related meanings constitute a single morpheme if the meaning classes are paralleled by distributional differences, but they constitute multiple morphemes if the meaning classes are not paralleled by distributional differences. Principle 6 A morpheme is isolatable if it occurs under the following conditions: 1. In isolation 2. In multiple combinations in at least one of which the unit with which it is combined occurs in isolation or in other combinations.

Jensen is probably correct in stating that his principle seems crucial to an understanding of Nida's principles. However, while the principle (2) says that the segmentation should be exhaustive it does not per se say that this segmentation is the only way of analysing a word morphologically. Yet, another consequence of the building-block model is that analysing a complex word amounts to finding its constituents. Let us formulate this as another principle: (3)

There is no other way that a word can be morphologically complex than by having discrete components.

1.2. Cognitive morphology There are several alternatives to a rigourously implemented buildingblock approach, and a spectrum of positions may be envisaged. The alternative view presented here will be found at the quite opposite end of this spectrum. According to this view morphological structure is considered as a set of relationships, or lexical connections, within the mental lexicon, and morphological elements as linguistic units activated through the lexical connections when the words are being used (Bybee 1985, 1988, 1995b). In this mode of thinking, the word

100

Helge Gundersen

as a whole, the composite structure, has its own semantic value, instead of this being solely derivable from the meanings of the morphological components. (The latter view is what Langacker says is based on the building-block metaphor, see the previous section.) The morphological components motivate, or highlight, certain aspects of the composite structure (Langacker 1987a: 453,1997a: 11). THROW/kast secondary activations THREW/kasta

primary activation

Ο /\

A Figure 1. Kasta 'threw'

Figure 1 is meant to depict this viewpoint with the Norwegian verb form kasta, the simple past of kaste "throw", as the example. Since kasta is a relatively frequent verb form, let us assume that the word as a whole is activated as a unit. Additionally, its components kast (the stem) and -a (the past tense suffix) are activated. Kasta is not seen as built up by kast- and -a; its composition is rather seen as a case of primary and secondary activation.

Problems of morphological segmentation

101

There are several types of common morphological phenomena that the building-block model unnecessarily makes into problems, or overlooks, including the following: • Phonological2 residues (cranberry morphs) • Morphological ambiguity (overlapping items) • Submorphemic relations (an item is morphologically complex without having constituents) These problems will be discussed in the following paragraphs.

1.2.1. Phonological"residues" One type of problem is phonological residues, so-called cranberry morphs, residual morphs, or unique morphs, such as Tues- in Tuesday or the Norwegian examples in (4). (4)

durkdreven 'cunning' (dreven = 'skilled') uhyre 'monster' (u- is a negative derivational prefix)

deüig 'delightful, delicious' (-ig is an adjectival derivational suffix) rädere 'etch' (-er is a verbal derivational suffix, see §2; -e is the infinitive suffix)

Phonological residues is the more apt name in the sense that these items are not really morphs. But "residues" may suggest either that they are leftovers when we have found the building blocks of the synchronic analysis, or that they always are historical residues of what once were meaningful components. The latter is not entirely correct, since cranberry morphs may also come into being in other ways. This is at least the case when some other part of the word is the result of analogical reinterpretation (misleadingly called "folk etymology"), see Gundersen (1995). An example is when many Norwegians perceive the otherwise unanalysable noun f0ljetong 'serial' (< French feuilleion) to contain the verb f0lge, or f0lgje, 'follow', thus making tong a "residue". (The variant f0lgje is

102

Helge Gundersen

pronounced in the same way asffilje- inf0ljetong.) The polysyllabic, opaque, and relatively infrequent word as a whole, f0ljetong, becomes more cognitively functional by being "anchored" to a more entrenched item, f0lge. The fact that cranberry morphs may be created in the process of making words more functional, suggests that these items are more of a problem for theoretical linguists than for language users. The solution offered in a morpheme-based model has been to extend the definition of morphemes to cover residues that are combined with recurrent morphological constituents, such as berry in cranberry. On the basis of the fact that a cranberry is a specific kind of berry, Bloomfield (1935: 160), in a subordinate clause ("since its meaning is constant"), takes the meaningfulness of cran- for granted. Hockett states (1958: 126), for the same reason, that "cran- clearly carries a meaning". Bauer (1988: 40-41) explains: 'The meaning associated with a unique morph is determined by subtracting the meanings associated with the known morphs in the construction from the meaning of the construction as a whole. So the meaning associated with cran is precisely what makes cranberries a subset of all berries." Aronoff (1976: 10) succinctly points out that "there is no noncircular way to assign meanings" to unique morphs. He makes use of this observation as one of several arguments for not defining the morpheme in terms of meaning. However, neither ascribing meaning to the "residuals" nor making this a case against a semantic understanding of the minimal morphological component is a necessity in a morphological theory. Let us take a look at the Norwegian noun shredder 'tailor', as depicted in Figure 2, a parallel to Figure 1. Cognitive grammar enables one to say that the derivational suffix -er makes explicit some of the meaning of skredder, and in this way motivating it, while skredd- does not. What makes skredder a subset of all the things designated by the suffix -er, is already included in the meaning of the composite structure as a whole, which is primarily activated. As for the phonological pole of the word skredder, the suffix -er is certainly dependent on the stem skredd-, so that skreddcontributes phonological content for -er to "work on" (see §2.3 for

Problems of morphological segmentation

103

this kind of analysis). As for the semantic pole, however, it is not necessary to say that skredd- contributes anything, as this is unproblematic from the standpoint of a network conception of morphology (see also Bybee 1988: 128). Skredder has lexical connections to the morph -er (specifically, skredder is, in Langackerian terms, an elaboration of the schema [NOMEN AGENTIS/...er]), but we are not required to ascribe meaning to skredd-, which remains a part of skredder as a whole without having connections to other lexical items. NOMEN AGENTISAer

secondary activation

TAILOR/skredder

primary activation

Figure 2. Skredder 'tailor'

104

Helge Gundersen

The premise for this position is that "the composite semantic structure has an essential kind of priority vis-ä-vis the component structures" (Langacker 1997a: 11). The word may secondarily contain one or more components which support the meaning of the whole, highlighting certain aspects of it. Among the considerations underlying this view is a notion of the language user associating elements with each other (finding similarities and differences) in accordance with general cognitive principles. In the building-block model, on the other hand, grammar operates independently of general cognitive principles, and the parts of a composite expression are objects stacked on top of one another, with the morphemes making up the bottom foundation.

1.2.2. Overlapping items A rigid interpretation of the building-block model excludes, or makes problematic, the possibility of ambiguous or overlapping analyses, in which one part of the word belongs to more than one constituent simultaneously. Well-formed tree structures are desirable in the building-block model. An example is the simple past of the Norwegian verb sende 'send' in some western dialects and in the Nynorsk variety of written Norwegian; the verb is inflected and pronounced as in (5). (5)

Infinitive: senda /senda/ (-a is the infinitive suffix) Present: sender /sender/ (-er is the present tense suffix) Past: sende /sende/ (-de is evidently the past tense suffix, as in levde, ringde, and many other word-forms)

The speech sound /d/ in the past, and actually also the letter , may be seen as having a double function: as a part of the stem send, and as a part of the past tense suffix -de. In a generative analysis, this example would be handled by a degemination rule which converts an underlying d + d into one surface d. This analysis involves constructs not allowed in cognitive grammar, but the point is that cognitive

Problems of morphological segmentation

105

grammar offers alternatives. For phonetic reasons, only one /d/ is pronounced, but this does not preclude a morphological analysis in which sende has connections to both the stem send /send/ and the suffix -de /de/. The word as a whole is primary to its constituents, and segmentation into unambiguous components is not a necessary requirement. Letting the components overlap falls naturally into the framework without it being a "slackening" of the model in order to accomodate some special type of occurrence. On the contrary, it makes it possible to catch connections that would otherwise be difficult to see. In the realm of compounding we have a phenomenon that would be very difficult to handle by rules (cf. Bauer 1983: 232-237, 293296), namely telescoped words (blends, clipping compounds, portmanteau words), such as English slanguage, brunch, motel, guestimate, infotainment, and bit (< binary digit', presumably, this example is synchronically telescoped for some speakers only). A new word is formed by combining the beginning of one word and the end of another, and the two parts of the new word are overlapping, and/or one or both parts are "incomplete". Slanguage has overlapping only, while brunch has clipping only. In motel, both clipping and overlapping are present, as may be seen in Figure 3. motor

hotel

motel Figure 3. Motel

The compounding elements in the clipped words would not be enough to represent the morphs in a morpheme-based model, despite the fact that the speaker/hearer is dependent on the connections to the full morphs to apprehend the full impact of the word. The language user will recognize the "original" words even if their phonological/ orthographical expressions are incomplete. This can happen because grammar is not something that is autonomously generated, but rather

106

Helge Gundersen

something that the language user does. An experiment by Cutler (1980), see Prideaux (1984: 90), may be taken as further evidence for the strategy. On the basis of adjectives like splendid, the subjects tended to make verbs like splendify, on the analogy of liquid/liquify, rather than splendidijy. The elimination of the final consonant (-d) was less of a loss than to move the stress away from the base form (splendid). Telescoped words have a flair of something eccentric and funny about them, seemingly not deserving much interest from the serious grammarian. However, overlapping items may be posited in noncompounding morphology as well, such as in sende above (inflectional morphology) and in some of the derivational examples in §2. Second, it may be argued, following Bauer (1983: 232), that blends, clippings, and other unpredictable formations are too common to be considered out of the ordinary, although they appear as oddities in the context of generative grammar. Third, according to the principles of cognitive linguistics a model should accomodate any kind of occurrence in an equally "organic" way, illuminatingly describing how it differs from, and is similar to, other kinds of occurrences.

1.2.3. Submorphemic morphology Both in Langacker's cognitive grammar and in Bybee's morphology, the locus of morphology in grammar is the middle part of a continuum from lexicon to syntax, in which a common denominator is that sounds are used to symbolize meaning. This continuum of symbolic units can be handled along several dimensions, such as how specific (or schematic) the item is, how morphologically (or syntactically) complex it is, to what degree it is an entrenched unit in memory, how productive a symbolic pattern is, etc. These considerations imply a gradient view of morphological structure. In the building-block model, morphology is seen as segmentation into separate building blocks. If such a segmentation is unworkable, there is no morphological relationship at all. We would then have either

Problems of morphological segmentation

107

fully segmentable words or fully simple words. This view is expressed rather clearly by Hockett (1954): "A linguistic form is either simple or composite" (with reference to item and arrangement), and "[a] linguistic form is either simple or derived" (with reference to item and process). This stance may be viewed in the context of similarity and identity (see also Enger 1998a: 140-152 with respect to inflectional morphology). In the tradition I am criticizing similarity is expressed by the presence of identical elements and is a matter of either-or. Cognitive linguists, on the other hand, can draw on the ability of the language user to observe similarities and differences between elements, and they can thus make use of degrees of relatedness. There are no clear boundaries between different types of morphological structure, but in Table 1, six groups are selected with the aim of arguing that words may often have symbolic complexity without being easily segmentable into discrete components. In particular, it will be argued that the two "submorphemic" types D and E are not entirely morphologically simple, despite the fact that they do not contain traditional morphemic component structures. Table 1. Some types of word-structure with varying degrees of complexity A Fully segmentable words (words [un][work][able], [black][board], analysable into symbolic segments) [bank] [er] B Partially segmentable words (words Tues[day], possible], butch [er] with in part symbolic components, in part purely phonological portions) C Inflections/derivations with modified sit-sat, begin-began, etc. stems/bases, instantiating a pattern relief-relieve, strife-strive, etc. D Words with root-forming morphs glare, gleam, glimmer, glow, etc. (roots with meaningful portions) sneer, sneeze, sniff, snort, etc. E Morphologically related wholes east-west, north-south (semantically and phonologically here-there similar words) male-female F Morphologically simple words hair, town, tree

108

Helge Gundersen

The basis for the grouping is to be found in more than one dimension, and it is not my claim that all words may be put unambiguously into one place in the table. The words at the bottom, in F, are morphologically simple, and, roughly speaking, the higher a word is found in the table, the more easily it is viewed as morphologically complex. Type F and type B (see §1.2.1) will not be treated further, nor will type C in any detail (although fully irregular alternations will be discussed below). It is well-known that this and other non-agglutinative phenomena are notoriously difficult to handle convincingly in a morpheme-based grammar (see especially Matthews 1972). Numerous devices have been launched, including the rather large apparatus of nonconcatenative morphology (e.g., McCarthy 1982), which Matthews (1993: 110) finds hard to see as "much more than an elaboration, in modern dress, of part of [Zellig] Harris's account of discontinuous morphemes". It is possible for a strict building-block approach to analyse quite a few nonagglutinative phenomena, but at the cost of introducing the extra machinery in, e.g., nonconcatenative morphology. Type A may be termed the ideal type, consisting of inflection and derivation by means of affixes (prefixes and suffixes are certainly the best representatives), in addition to compounding. The conceptual structure is symbolized by the phonological structure in a neat way. However, this type is not that "simple". One reason is that some words are perceived as more transparent, or analysable, than others (unworkable is probably the most analysable of the words in the table). A novel expression will be fully analysable, but when the word becomes more entrenched, the linguist may be in doubt whether to segment it. These matters will not be elaborated further here (see, e.g., Langacker 1987a: 457-466). However, this gradient analysability should be incorporated in a grammatical model in a consistent manner, with independently motivated constructs. As regards affixes, I shall hold that they may show submorphemic characteristics; that is, they may display parallel phonological and semantic connections below the level of the morph. In the Nynorsk variety, many nouns (many feminines and some masculines) end in -er and -ene in the plural, while many others (a large minority of

Problems of morphological segmentation

109

feminines and most masculines) end in -ar and -ane. When ending in -enel-ane they are at the same time definite, while the word-forms in -erl-ar are not inflected for definiteness. The lexemes vegg 'wall' and hest 'horse' (both masculine) thus go like this in the plural: (6)

vegger 'walls' veggene 'the walls'

hestar 'horses' hestane 'the horses'

As Endresen (1982: 105-106) points out, the -e in -er may easily be perceived as the same as the first -e in -ene, and the -a in -ar as the same as the one in -ane. The endings -enel-ane have been viewed in two ways: The Norwegian Reference Grammar (Faarlund, Lie, and Vannebo 1997: 161, 193) sees them as a sequence of the plural -er/ -ar (in a form without the -r) and a definite suffix -ne. The segmentation is then vegg-e-ne, hest-a-ne. The other view (cf. Vangsnes 1999) is to see -ene and -ane as morphs with two functions; "plural" and "definite". (The exact motivations for the two analyses may vary between scholars.) It is tempting to suggest that both views have their advantages; the analyses vegg-e-ne and hest-ane appear counter-intuitive and exaggerate the segmentation (the missing -r is possibly also problematic), while they still catch a connection between -er and -ene, and -ar and -ane, respectively. A permissible analysis in cognitive morphology is to regard -er and -ene (and -ar and -ane) as two morphs (suffixes) indeed, but to ascribe a secondary morphological connection between them based on semantic commonality (plural) and phonological similarity (both begin with an -e, or an -a). This move is allowed in a model which deals with network relations and which does not equate analysis with segmentation. Turning to type D, we have a situation where one portion (traditionally called a phonestheme) of the root is meaningful, but, putatively, still not a discrete component that takes part in the hierarchical organization of the word. A number of linguists, such as Bloomfield (1935: 245) and Rhodes and Lawler (1981), have seen a gl- in English pertaining to the field of light, occurring initially in a number of words, such as glare, glimmer, and glow. English words

110 Helge Gundersen

like squash, squeeze, and squelch have in common both the squ- and the meaning "compressed". Many of those doing research in the field, sometimes with the aid of experimental evidence (e.g. Abelin 1999), consider phonesthemes fairly widespread. For example, Rhodes and Lawler (1981: 328) believe that more than 50 per cent of English monosyllabic words are analysable, and Abelin found that 36 of the 37 word-initial consonant clusters in Swedish may function as phonesthemes. The phonestheme, such as g/-, may be viewed as "halfway" separated from the rest of the word, simultaneously being a meaningful unit and a part of the root. It is perhaps relevant that the remaining parts of the words (-are, -immer, etc.) can be very dissimilar from each other phonologically, so that it is the phonestheme alone that symbolizes the shared semantics in the group of words, in the case of gl- something concerning light. In this way, these units are similar to affixes. As Rhodes (1994: 286-287) points out, phonesthemes are often allotted a special status below the morphological level, but Rhodes and Lawler (1981: 326) and Rhodes (1994: 288-290) argue for a case of derivational morphology, with the word being either wholly or partially analysable. We may follow Bloomfield (1935) and call them "root-forming morphs", thereby indicating that they are minimal symbolic (morphological) units and, at the same time, part of the root. (The latter point implies that they may not have all the characteristics of standard morphs, and that the words in which they are found may not be as easily segmentable as in "type A".) Because the building-block model identifies morphology with exhaustive segmentation into morphs, as conventionally understood, the whole phenomenon then disappears from morphology. In a cognitive-linguistic network conception, any instance of phonological and semantic connections that are recognized as running in parallel in words, may be treated as a case of morphology. This does not mean that cognitive morphology cannot make use of morphs; Langacker does, and defines the morph ("morpheme" in his terminology) as an "expression of minimal symbolic complexity, i.e. one not analyzable into smaller symbolic components" (1991: 550). But it is not a decisive question whether

Problems of morphological segmentation

111

we can single out morphs in particular cases. Bybee (1995b: 233) characterizes morphemes as epiphenomenal or "emergent". (See also Langacker 1987a: 400-401; Bybee 1988: 128-129, and Davis 1992 for discussions of submorphemic units in relation to morphological segmentation.) In type E we have a different situation than in D. Certain words may have a morphological relationship to one another that cannot be ascribed to any specific part of the whole. It is the words as wholes that are morphologically related. There are phonological and semantic connections running in parallel, but the phonological connections are in the form of a similarity relation, not an identity relation. Examples may be English east and west, north and south, which are most likely perceived by the language user as similar, both semantically and phonologically. (The pairs are discussed by Matthews 1991: 18, while Plank 1981: 187-189 discusses the German counterparts.) From the standpoint of cognitive linguistics, it is natural to draw attention to the fact that all the words are monosyllabic (prosodic factors may be included in a morphological relation), that the two members of each pair are similar (not identical) in their CV-structures, and, notably, that both end in -stl-th. It is not difficult to concur with the opinion that it is very doubtful to see -st and -th as suffixes, but it is still possible to regard the words as morphologically related. The Norwegian words her /hae:r/ 'here' and der /dar.r/ 'there' illustrate the point even more clearly, since der has been changed from dar, in my opinion largely because of analogical influence from her (Gundersen 1995: 68, 132), while no sane linguist (including me) would even consider positing the part -er as a morph. As shown in Figure 4, these two words start with one consonant and continue with /se:r/. A broken line indicates shared features, while the solid lines stand for identity relations. The present model allows us to consider the relation between her and der morphological. The alternative in a strict building-block approach would be to ignore the contribution of the phonological similarity altogether, and to treat her and der as fully on a par with purely semantically related words like floor and ceiling.

112 Helge Gundersen

her ι

der Figure 4. Her 'here' and der 'there'

The relationship between the two words these and those as wholes looks like a clear case of type E: The two words are monosyllabic, and both the th- and the -s- appear in the same positions. The vowel is long in these, and a diphtong in those, making up a similarity relation (just like the initial consonants in her and der). The word this is also morphologically related to these words, especially these. As for the th- by itself, this portion of the words perhaps rather points to type D, but bordering on type E; the th- is tied up to a very limited set of words with schematic meanings (also including they, that, and others), but they may nevertheless form the basis of a schema (pattern). In this way th- would be similar to a phonestheme, thus a root-forming morph (see also Bloomfield 1935: 244). A clearer candi-date may be the wh- that appears initially in most English interro-gatives (what, where, etc.). The English demonstratives display in this way the gradual transition existing between words with root-forming morphs (type D) and singular relations between whole words (type E). Pairs, or small groups, of words with close-knit semantic relations frequently resemble each other phonologically, and occurrences like her and der, east and west are examples of that phenomenon. Over time, semantically related words may also achieve a phonological similarity originally not present or become more similar than to begin with, as has been noted in studies of analogical change (see, e.g., Osthoff 1878 and Winter 1969 on, mainly, numerals). Some other types of words than those noted in the previous paragraphs are personal pronouns, auxiliary verbs, numerals, kinship terms, names of months and the days of the week, basic colour terms, basic antonymies like "up-down", "left-right", "heavy-light", "male-female", and also idiosyncratic pairs of words like Norwegian ridder 'knight'

Problems of morphological segmentation

113

(from Low German, cf. English ride) and rytter 'rider' (originally probably from Latin rutarius, ruptuarius 'highwayman', according to Bokmalsordboka). We often have a singular relation between two or three words, but at other times a morph-like portion of the words seems to emerge. The above will fall outside of what has traditionally been considered morphology. This cannot be said so easily about der-ivational pairs like English pope and papal, which show alternations of the base that do not follow a pattern. However, this phenomenon can fruitfully be treated in line with the above discussion. (This may also be said about the suggested analysis of the Norwegian suffixes -ene and -one above, although these units are affixes, not roots.) While -al is a morphological component (type A above), the alternation between -o- and -a- is more open to debate. The alternation is not regular, but pap still appears to be derived from pope. In a structuralist or generative framework, the linguist is forced either to list the two as unrelated forms in the lexicon, or to make use of some special device, as in the form of Jackendoffs (1975) "redundancy rule" or some later, equivalent development. Since the alternation is unique, a "rule" like this will, fairly obviously, not tell us anything more than listing the two forms in the lexicon as unrelated. (This actually follows from Jackendoffs own formulations, cf. Carstairs-McCarthy 1992: 42, and it seems that devices like morphological redundancy rules are intended to apply to occurrences of type C in Table 1.) Rules will not work satisfactorily, and the phonological similarities between the alternating forms are too apparent to be dismissed. If the fundamental assumptions are revised, however, we can simply assume that the items pope and pap- are related by shared semantics and similar phonology, and take this to be a morphological statement. The solution will be based on theoretical concepts needed independently in the model. (See further treatment in §2.3.) This step cannot fully be taken if we entertain the view that a grammatical analysis largely consists of listing building blocks and stating how these are combined. In fact, I think it would be fair to say that it is necessary to adopt a view of morphology as a network in the mental lexicon.

114 Helge Gundersen

Now, even when all the principles behind the discussion in the previous paragraphs are accepted, it will still be possible to argue that the term "morphology" is best reserved for "parallel phonological and semantic connections, if they represent a pattern found in multiple sets of items", which is what Bybee (1995b: 232) holds, as a cor-rection to statements in her earlier work. "If there were only one pair of items with such connections this would not be regarded as consti-tuting a morphological relation." The proposed connections be-tween, say, Norwegian her and der, or ridder and rytter, would then not belong to morphology, and this is most likely in accordance with standard conceptions. As mentioned at the beginning of §1.2.3, there is a continuum between lexicon and morphology along several di-mensions, and we may see what belongs to what as partly a matter of terminology. Nevertheless, the transition from a relation between two singular items to a set of items conforming to a pattern is gradual with room for considerable indeterminacy. In particular, morphologists are sometimes confronted with a doubt as to whether an inflectional or derivational alternation instantiates a pattern or is a fully irregular similarity relation (such as between pope and pap- in papal). Thus, it seems profitable to be able to sort the study of parallel phonological and semantic connections in singular pairs of items under the heading "morphology", as long as we have to do with phonologically different words. It is in any event desirable to make explicit the continuum from complex to simple items. When the cases in question are excluded from both morphology (because they do not represent a pattern) and lexical semantics (because there are differences in the phonological, not only semantic, structure of the items), they may easily fall through the cracks without receiving attention. To sum up, there is no clear boundary between what ordinarily is considered morphology on the one hand and submorphemic (yet morphological) occurrences on the other. This is not a problem in a cognitive-linguistic network approach, because segmentation is not required. Morphology is under all circumstances seen as a network in the mental lexicon, and a morphological viewpoint may easily be

Problems of morphological segmentation

115

adopted with respect to various "difficult" cases, including submorphemic relations.

2. The Norwegian verbs in -ere In §1.2 I argued that cognitive morphology does not suffer from the shortcomings imposed by a model that invites to exhaustive (§1.2.1), unambiguous (§1.2.2) segmentation into separable constituents (§1.2.3). In this section, I shall exemplify my arguments further, but this time from a different angle. Instead of going through types of phenomena, and exemplifying them, I shall concentrate on a limited area in the Norwegian language, and show how various segmentation problems will crop up there. My intention is not to propose solutions to all of these problems. My aim is rather to argue that a cognitivelinguistic network conception of morphology partly opens up fields of study necessarily ignored in a building-block framework, and partly suggests ways out of segmentation problems that otherwise seem inextricable. The area in question is the synchronic derivational morphology of a group of Norwegian verbs with stems ending in —er (-ere in the infinitive). These make up about 1500 of the 4800 verbs in Bokmalsordboka (cf. Bjerkan and Simonsen 1996: 193; see also Endresen and Simonsen this volume). The verbal derivational element in these words is the suffix -er, which also exists in Swedish, Danish, Dutch, and (in the incarnation of -ier) in German. A number of ere-verbs were borrowed as wholes into Norwegian from Low German during the Hanseatic period (1200-1550), and became productive (Leira 1992: 37-38, 147-148), although zeroderivation of verbs occurs more frequently. As is often the case with loans, irregularities arise in the course of the morphological adaptation into the receiving language. (7) shows some well-behaved introductory examples of the verbs, with the derivational bases to the left. I cite the verbs in the infinitive, with the infinitive suffix -e included at the end.

116

Helge Gundersen

(7)

respekt 'respect' steril 'sterile' jazz 'jazz'

respektere 'respect' sterilisere 'sterilize' jazziflsere 'jazzify'

-iser (as in sterilisere), and the longer -(i)fiser (as in jazziflsere) are very often taken as variants of -er in Norwegian (cf. Leira 1992, Johannessen 1993, Faarlund, Lie, and Vannebo 1997: 121; see also Söderbergh 1967, Thorell 1981 for Swedish), -iser is the most productive of these (Leira 1992: 38). The -er, including the part -er in the longer variants -iser and -ifiser, is always stressed. The base for -er-derivation as well as zero-derivation of verbs in Norwegian is ordinarily a noun or an adjective. Both types of derivation (with -er and zero) appear as ways of making verbs in general, not restricted to certain verb meanings. (However, there is a strong tendency to prefer the one over the other on phonological, and probably also semantic, grounds, and the same may be said for the choice between the variants of-er.) There are a number of morphological problems on the boundary between -er and what comes before it. One type of problem is where we have one or more speech sounds before -er that are difficult to account for, in that it is not clear which morphological components they are parts of, if any at all. Verbs with three frequent terminations will be used as examples: -tere, -isere and -ulere.

2.1. The termination -tere The verb handtere 'handle' illustrates the problem nicely. Handtere, or handtere (depending on the variety of Norwegian), stems from French (it is related to English haunt), and is borrowed into Scandinavian from Low German hanteren, which, according to Lübben (1993), meant 'visit frequently', 'go hither and thither', 'trade' (the latter meaning originating in the traffic of pedlars). It is associated with hand /hon/ or hand /han/ 'hand' in Norwegian, as shown by the spelling with -d, and the fact that it may be pronounced

Problems of morphological segmentation

117

with an /o/. Here we may imagine the following possible segmentations in modern Norwegian: (8)

a. b. c. d.

händt-ere /hont-e:re/ händ-tere /hon-te:re/ both händt-ere /hont-e:re/ and händ-tere /hon-te:re/ hand-t-ere /hon-t-e:re/

In (8a), händt /hont/ 'hand' can be seen as a variant of hand /hon/ 'hand'. This segmentation would most likely be considered the least controversial, since it introduces neither any unorthodox suffix, nor any residuals. However, the relationship between hand and händt remains a bit problematic from a theoretical point of view, as these lexical items are clearly related, but as singular items, not via any rule. It is perhaps relevant that the pronunciation /hont/ for hand can be found, although not in all dialects, in a few other loans from Low German: händverk 'craft', händverker 'craftsman', and händlanger 'helper, assistant'. In (8b), -ter is seen as the suffix. This solution would be considered very original, since no one has ever posited a suffix -ter in the Norwegian grammar. But there are in fact quite a few verb stems ending in -ter, about 230 out of the 1500 ere-verbs in Bokmalsordboka, and one might think that this massive number could be an additional source of reanalysis. Instead of legitimasjon 'identification card' one can surprisingly often hear people say legimitasjon, and, correspondingly, legimitere seg 'prove one's identity' instead of legitimere seg (Gundersen 1995: 76). However, one may convincingly argue that this is a case of purely phonological analogy. In (8c), we have an ambiguous structure; that is: (8a) and (8b) are both valid, the speech sound I\J playing an ambiguous role. In (8d), the /t/ is not part of the morphs hand and -er. This solution retains hand as the base and -er as the suffix, but gives us the -/- as something that, in a building-block framework, would be considered a residue. It solves the problem of (8a), but evokes another theoretically difficult point instead.

118 Helge Gundersen

In theory, all four options should be considered equally permissible and unproblematic, and it is an empirical question which one to choose.

2.2.

The termination -isere

Verbs ending in -isere include those with the variant suffixes -iser and -ifiser, like the two last examples in (7), sterilisere and jazzifisere. -is- is a phonological string that may also appear in other environments, such as in (9): (9)

organisere Organize' (approx. 300 -isere) organisasjon Organization' (23 -isasjon)

Excluding compounds, 23 nouns in Bokmälsordboka end in -isasjon. If the derivational suffix in organisere is -iser, how is the form organisasjon to be accounted for? The morphological status of -isis therefore somewhat disputed. Some have regarded -is- as a purely formal "concatenator" with the function of linking morphemes to one another (cf. van Stek 1983: 51). This solution seems to be at odds with the principles of cognitive grammar, in which grammatical functions are described in semantic terms. One possible solution within this framework, however (there may be other possibilities, too), is to regard -iser and -ifiser indeed as variants of -er, but still to retain a connection between the -is- appearing here and the -isappearing before -asjon in nouns like organisasjon, regardless of whether —is- in the latter case belongs to a nominal suffix -isasjon, or to the derivational base before -asjon. While speakers do seem to analyse words into components (although these components do not have the status of building blocks), this is only one facet of the description, and morphological relations need not always follow the "componentiality" of the words. The string -is- may be seen as problematic in two more ways. First, some of the verbs with -iser correspond to adjectives with the

Problems of morphological segmentation

119

derivational suffix -isk, to nouns with the derivational suffix -z, and to nouns without any of these suffixes: (10) psykologisere 'psychologize' psykologisk 'psychological' psykologi 'psychology' psykolog 'psychologist' The verb psykologisere corresponds to psykologisk, to psykologi, and to psykolog. As depicted in Figure 5, we have a difficulty deciding on whether we have psykologisk + -er 'make psychological' (without the /k/), psykologi + -iser 'make into psychology' or psykolog + iser 'reason like a psychologist'. One possibility allowed by the network model is that all (or two) of these analyses are valid at the same time, i.e. that we have a case of simultaneous multiple analyses. Another possibility is that each of the analyses is correct at different times in different contexts.

p s y k o l o g i s e r e 'makepsychological'

//\\ -er

-isk

p s y k o l o g i s e r e 'make into psychology' -ι

-iser

p s y k o l o g i s e r e 'reason like a psychologist' -iser Figure 5. Interpretations of psykologisere

120

Helge Gundersen

Another problem is that -er seemingly is a part of -iser, while -ifiser may be seen as an extension of -iser. Figure 6 illustrates the formal similarities. If these items are seen as three more or less equivalent suffixes, it is hard to believe that their phonological similarity is strictly coincidental, unrelated to their semantic similarity from a synchronic point of view. One hypothesis might be that there is a morphological relationship between them that is below the morphemic level. If this is so, they are three different morphs, but each of them carries phonological similarities to the other two that are connected to their semantic similarities.

-er

-iser -ifiser Figure 6. The suffixes -er, -iser, and -ifiser

2.3. The termination -ulere Approximately 120 verbs in Bokmälsordboka end in -lere; 44 of those end in -ulere. The sequence -ul- that appears in front of some derivational suffixes, never appears at the end of independent words. 27 of almost 500 nouns with the suffix -asjon end in -ulasjon in Bokmälsordboka, while 6 of the 71 adjectives ending in -ativ also end in -ulativ. Some of the verbs in -ulere correspond to other derivations, like spekulasjon and spekulativ, while others do not. (11) spekulere 'speculate' (44 -ulere) spekulasjon 'speculation' (27 -ulasjon) spekulativ 'speculative' (6 -ulativ)

Problems of morphological segmentation

121

In the literature (cf. references in the introduction to §2), -ul- is not treated as part of a suffix, but my question is: do language users in some way connect -ul- to the suffix coming after it? One interesting formation is spankulere 'strut', 'stroll', which, historically, is created from the shorter (now obsolete) verb spanke 'strut', in accordance with verbs ending in -ulere, such as spekulere 'speculate'. On the other hand, if we say that -ul belongs to -er, how shall we then handle the fact that -ul appears before other derivational suffixes, such as before -asjon and -ativ in (11)? This type of problem was discussed in section 2.2 with examples among verbs in -isere (organisere, etc.). If we do not see -uler as a suffix, we shall still have to deal with some morphological problems concerning all those derivational bases ending in -ul. Are they really—synchronically speaking— coincidental, unrelated phonological elements, or do they perhaps take part in a network of submorphemic relations within morphology? Furthermore, a handful of the verbs ending in -ulere bear both phonological and semantic similarities to nouns ending in -el. Four of them are common, and are shown in (12). (12) fabel 'fable' fabulere 'give one's imagination a free rein' regel 'rule' regulere 'regulate' sirkel 'circle' sirkulere 'circulate' tittel 'title' titulere 'address' (i.e. 'address somebody as...') These correspondences seem to create a type of problem that we find in many other ere-verbs, such as the relationship between the noun kalkyle 'calculation, estimate' and the verb kalkulere 'calculate', and between the noun applaus 'applause' and the verb applaudere 'applaud'. The first of the segmentations in §2.1 considered handt-

122

Helge Gundersen

in kandiere 'handle' a variant of hand 'hand', and this is another instance of the phenomenon. Similar to kandiere are verbs like flambere 'flambe, flame (a dish)', cf. the noun flamme 'flame', and griljere 'grill', cf. the noun grill 'grill'. In §1.2.3 the equivalent example was pope and papal. If we posit that the semantic and phonological connections are perceived as running in parallel in these pairs of items, we may view that as morphology. This type of morphology, however, cannot be captured in a general statement, such as by means of rules. In the case of the verbs in (12), there is still a certain possibility that speakers have extracted a schema (pattern, "rule"), so that we should have a case of what I labelled type C in Table 1, equivalent to belief-believe, calf-calve. If so, this case would be accounted for in cognitive grammar with analytic tools conforming to general cognitive principles, such as Schemas and comparison (in a similar, but not quite identical, way as in Figures 8 and 9 below). In a building-block approach, I assume that special devices (considered artifactual in cognitive grammar) like the tiers of nonconcatenative morphology might come into play. For unique correspondences like those in the other examples above (kalkulere, etc.), such devices would not be feasible in any event. Unique alternations concern submorphemic phenomena of what I labelled type E in Table 1. If we choose to see the correspondences in (12) as unique, sirkul in sirkulere is neither the noun sirkel in some unclear sort of disguise, nor completely unrelated to sirkel. Instead it may be viewed as a unique item that is morphologically related to sirkel below the level of constituency. A simple diagram in line with Bybee's morphology may look like Figure 7.

'circle'

s i r keI ι I

'circle' Figure 7. Sirkel and sirkulere

I

s i r ku l e re

Problems of morphological segmentation

123

There are lexical connections between the semantics of the expressions and between their phonological structures, and these run in parallel. The broken line indicates similarity (both phonemes are vocals), the others phonemic identity. Likewise, connections could have been drawn from -er to other verbs with the same derivation. In Langacker's more comprehensive cognitive grammar, the notation will be more complex, taking more factors into consideration. There are several ways to go about it, but some central points will be noted here. For the sake of simplicity, I shall restrict myself to the stem sirkuler. The composition of the phonological pole of the expression (only the relevant aspects) is shown in Figure 8. For the sake of convenience, orthographical representation is employed.

• sirkel

hΨ sirkul • ^

'4ffffl> er [[SIRKULER]/[sirkuler]]

tr

Im

Figure 9. The semantic pole of the stem sirkuler

Sirkuler is here categorized by the more schematic unit -er. A derivational suffix may be seen as a schematic word (or stem), and the more specific word, in this case the stem sirkuler, fully instantiates, or elaborates, this schematic word. It does that semantically by having a more specific meaning that is compatible with the meaning of the suffix, and phonologically by having a more specific phonology (notationally by filling in the three dots).

126 Helge Gundersen

3. Summary and conclusions This paper has shown that cognitive grammar, or the cognitivelinguistic framework, has something to contribute to morphological analysis, including the seemingly problematic cases. I have contrasted two very different conceptions: In the buildingblock approach, segmentation into discrete components is a defining characteristic of morphology. These components, and the means to assemble them, are conceived of as being independent of the language user. The cognitivist approach sees grammar as a reservoir of linguistic units structured in a mental network and conforming to general cognitive principles. More specifically, I have made use of a strictly symbolic, bipolar view of morphology, in which the discipline becomes the study of the relation between sounds and meaning in (complex) words. Humans employ speech sounds to symbolize meaning, and morphological structure constitutes an often quite complex web of parallel phonological and semantic connections. In this picture, the segmentation of words into separate components is merely one dimension, important as it may be. I do believe that it is psychologically valid to say that the English word falsehood may be divided into two components: false and hood. But firstly, as argued in §1.1 (residuals) and §1.2 (overlapping items), the status of such components is not that of building blocks. Rather, they are secondary to the word as a whole, and can only be singled out on the basis of relations to other items in the web that makes up morphology. Secondly, as discussed in §1.3, a language may contain symbolic structures that are below the level of the morph (as this is commonly conceived), and cognitive morphology allows us to describe these structures in a coherent manner. Although the "component status" of portions of words—to what degree the language user processes these portions as separate morphological components—is an important aspect of morphology, there is more to the morphological structure in languages like English and Norwegian than the building-block model implies. One example is the so-called residual "morphemes", which are a normal part of word structure, as opposed to what the awkward evoking of

Problems of morphological segmentation

127

cranberries suggests. Another example is the meaningful submorphemic relations by which many words are knit together, also in conjunction with inflection and derivation. It should indeed be possible to account much more precisely for those relations, given that the requisite fundamental assumptions are in place. Fully analysable words may well represent the morphological ideal type, but linguists should strive for a description covering all facets of symbolic structures in words. When a type of structure is covered, it should be laid down on the linguistic map in an illuminating way, without resort to extra theoretical machinery. I hope to have demonstrated that models like cognitive grammar offer a framework that has room for an undertaking of that type.

Notes 1. I wish to thank Rolf Theil Endresen and Hanne Gram Simonsen for their support and for their invaluable help in the improvement of this paper. The paper has also benefitted greatly from the input of two anonymous referees. 2. Admittedly, this paper falls into the ubiquitous "phono-centric" view in linguistics to the extent that phonology receives most of the attention among the means to express meaning. However, there is nothing in the framework presented that precludes an inclusion of orthography and paralinguistic features, although I am not sure that the relationship between speaking, writing, and non-verbal elements is a straightforward and obvious one.

Past tense acquisition in Norwegian: Experimental evidence Hanne Gram Simonsen

This paper reports an experimental study of Norwegian children's acquisition of past tense (PT) forms. Using a method similar to that ofBybee and Slobin (1982), past tense forms were elicited from three groups of children, at 4, 6, and 8 years, with approximately 30 children in each age group. Each child was tested individually on 60 verbs, half of them strong and half of them weak. The verbs were carefully chosen according to type frequency (number of verbs belonging to each inflectional class), token frequency (frequency of use of each verb), and phonological properties of the classes. The results indicate that input factors, such as phonological properties of the verbs and their type and token frequencies, play an important role in the acquisition of Norwegian PT forms, both regarding correct performance and regarding the errors children make, and for all verb classes. Children generalise over all the main inflectional patterns, but to a varying degree, and the error patterns change over time. The differences between verb classes seem to be of a gradual nature, and not a question of either/or, thus lending support to models assuming that the mechanisms involved are not dichotomous, but rather of a homogeneous nature.1 1. Introduction In this article the results from an experiment concerning past tense acquisition in Norwegian are presented and discussed in the light of different models for acquisition. The experiment was constructed by Hrafnhildur Ragnarsdottir, Kim Plunkett and the present author to test and compare past tense acquisition in Icelandic and Norwegian children, and a full description of the whole experiment with its

130

Hanne Gram Simonsen

results are presented in the main report (Ragnarsdottir, Simonsen, and Plunkett 1999). However, since many of the articles in the present volume are closely connected with and refer directly to this experiment (Bjerkan; Matcovich; N0klestad; all this volume), the Norwegian data from the experiment are presented again here. Also presented are some additional results for Norwegian which were not included in the above mentioned comparison with Icelandic. The past tense, with its combination of regular and irregular patterns, provides a particularly instructive window into how children gradually master the morphology of their language, and into how morphological patterns are stored and processed. Thus, the question of how children acquire the past tense of verbs has been discussed within different research paradigms and on the basis on different types of data over the last 50 years, with an overwhelming majority of studies based on English-speaking children. For example, there are naturalistic studies based on longitudinal data bases, (Brown and Bellugi 1964; Brown 1973; Marcus et al. 1992), experimental studies based on nonsense forms (Berko 1958; Derwing and Baker 1986) or real forms (Bybee and Slobin 1982; Marchman 1997), simulations of past tense learning in connectionist nets (Rumelhart and McClelland 1986; Plunkett and Marchman 1991, 1993, 1996), and PET imaging studies addressing the same questions in adult processing (Jaeger et al.1996)—to mention a few. Although the bias towards English has changed somewhat with more crosslinguistic work during the last 10-15 years, the perspective has remained relatively English-centered. This seems also to have influenced some of the models put forward to account for the acquisition process.

1.1. Models for past tense acquisition and processing There are currently two main models which aim to explain the acquisition and processing of past tense, which may be called the dual mechanism account and the single mechanism account.

Past tense acquisition in Norwegian

131

The dual mechanism account is related to a generative or symbolic framework (cf. Pinker and Prince 1988; Marcus et al. 1992, 1995; Prasada and Pinker 1993). Here, the basic assumption is that there are minimally two different mechanisms at work: on the one hand a rule-based mechanism which operates on an all-or-none basis, taking care of the regular (weak) inflections including overgeneralisations of them; and on the other hand a lexical retrieval mechanism which is sensitive to input factors like frequency and phonological factors, and which is responsible for all irregular (strong) forms. The retrieval of an irregular form from the lexicon blocks the application of the general rule. The later developments of this model allow for a second split in the lexical retrieval mechanism, where the most suppletive forms are listed separately, but where an associative memory component takes care of all subregularities and similarities among the irregular verbs. However, the basic, dichotomous split is still there—between the "rule-governed" regulars and the "exceptional" irregulars, treated by two qualitatively different mechanisms. The single mechanism account is related to a connectionist framework (Rumelhart and McClelland 1986; Plunkett and Marchman 1991, 1993, 1996), and an analogical (Skousen 1989, 1992; Derwing and Skousen 1989) or cognitive linguistics approach (Langacker 1987a; Bybee 1988, 1995a). This account assumes that there is only one mechanism at work, a mechanism which is sensitive to input factors like type and token frequencies of verb forms, and to their phonological properties. All past tense forms (with the connection to their base forms) are memorised to a certain extent. But given a certain amount of phonologically similar forms, generalisations are formed across these forms. Thus, the dual mechanism account assumes that there is one default rule which is applied to weak verbs, while the strong verbs are handled by a qualitatively different, more rote-like mechanism. In the processing of a strong verb, if the mechanism does not manage to retrieve a strong form from the lexicon, there is nothing to block the general rule from applying, and we get errors of overgeneralisation of the default weak inflection, like goed and silled instead of the correct

132 Hanne Gram Simonsen

went and sat. Such overgeneralisations of weak verb inflections should not be influenced by input factors—nor should the correct acquisition of weak verbs. On the other hand, since the strong verbs are handled by a different mechanism, overgeneralisations of strong inflections are expected to be more rare, and to apply only where phonological similarity and frequency factors allow for them. The single mechanism account, on the other hand, holds that the same mechanism regulates both strong and weak verbs. This implies that overgeneralisation errors take place in essentially the same way both for weak and strong inflectional patterns, once there are enough phonologically similar forms to generalise over. It also implies that input factors will be important for both acquisition and overgeneralisations of all kinds of verbs.

1.2. Norwegian verb morphology as compared to English An outline of Norwegian verb morphology is given in Endresen and Simonsen (this volume). Here, only a reminder will be given of a few interesting points where Norwegian verb inflection differs from that of English. The variety of Norwegian described in this paper is the dialect spoken in and around the capital of Oslo, a variety close to the Bokmäl standard. To a large extent it also covers other dialects spoken in Eastern Norway. One interesting difference between Norwegian and English is that the former has two distinctive word accents or tones. Word tone is usually lexically contrastive, but there are also morphological tone effects: in verbal inflection, word tone in the present tense is a reliable cue as to whether the verb is strong or weak. With a few exceptions, strong verbs have a Low Word Tone in the present tense, while weak verbs have a High-Low Word Tone in the present tense (while the infinitive has a High-Low Word Tone in both weak and strong verbs). For example, the strong verb sitte /site/ 'sit'with the past tense form satt /sät/ 'sat' has sitter /siter/ in the present tense, while the weak verb titte /tite/ 'look' with tittet /titet/ 'looked' in the past tense has titter /titer/ in the present tense.

Past tense acquisition in Norwegian

133

The basic split between weak and strong verbs found in most Germanic languages is similar in Norwegian and English. However, as illustrated in Figure 1, Norwegian has a slightly more complex inflection than English, with 6 word forms in the weak paradigm as compared to 4 in English. And in addition there is a difference in the number of weak inflectional classes. English has only one weak inflectional class, while Norwegian has two, one of which (represented with the verb kaste 'throw') is more productive than the other (represented with the verb rope 'call'). Thus, there is a less clearcut division into two different groups of verbs ("regular" vs. "irregular") in Norwegian than in English. ENGLISH stem (= inf./pres./imp.) -ing-form -5-form (3.p. sg. pres.) -erf-form (past/past part.)

One single weak inflectional class dance call dancing calling dances calls danced called

NORWEGIAN Two weak inflectional classes stem (= imperative) kast( 'throw') rop('call') infinitive kaste rope present kaster roper past kästet ropte past participle kästet ropt pres. participle kastende ropende Figure 1. Weak verb paradigms of English and Norwegian

This way, the testing of verbal inflection in Norwegian as compared to English may provide a new angle to the understanding of past tense acquisition, and possibly shed new light on the suitability of the different models mentioned above.

134 Hanne Gram Simonsen

2. The experiment To test the possible effect of input factors on acquisition, parallel experiments were designed for Icelandic and Norwegian, eliciting past tense forms of cognate verbs in these languages. A full description of this study including the details of the statistical analyses is found in the main report (Ragnarsdottir, Simonsen and Plunkett 1999). As mentioned above only the Norwegian data from the experiment will be reported and discussed here.

2.1, Elicitation task The elicitation task follows the pattern used in Berko (1958), Bybee and Slobin (1982), where the child is shown a picture of someone performing an action. The experimenter says: "Here is a boy who knows how to VERB. He is VERBing. He did the same thing yesterday. What did he do yesterday?" And the child fills in (presumably with the past tense form of the verb): "He VERBed (or VARB)." Since the word tone in the present tense is a reliable indicator as to whether the verb has a strong or a weak past tense inflection, approximately half of the children were given the infinitive as a cue (e.g., "Her ser du en gutt LIGGE i sengen" = Here you see a boy LIE in the bed), while the other half were given the present tense as a cue (e.g., "Her ser du en gutt som LIGGER i sengen" = Here you see a boy who LIES in the bed). (The actual phrasing of the rest of the cue in Norwegian corresponds closely to the English version.)

2.2. Subjects Three groups of children were tested, with approximately 30 children in each group, at 4, 6, and 8 years of age, half of them boys, half of them girls. The children were recruited from preschools and schools in middle class areas in Oslo, and chosen from families where both parents spoke an East Norwegian dialect.

Past tense acquisition in Norwegian

135

Table 1. Subjects Subjects 4-year-olds 6-year-olds 8-year-olds

n 28 29 39

Mean Age

4; 4. 4 6; 3. 15 8; 3. 0

SD 0;3.5 0; 2. 10 0;1.27

2.3. Verbs Each subject was tested individually, on 60 verbs, 27 strong and 33 weak, presented in a single random order. (To control for fatigue during testing, half of the children were presented with the last thirty verbs first. Before testing, the children were presented with a few warm-up verbs to check that they understood the test and that they managed past tense inflection.) The verbs in the test were carefully chosen, taking into account input properties assumed to influence acquisition—both frequency factors and phonological factors. Concerning frequency factors, both type frequency (i.e., the number of verbs belonging to each verb class) and token frequency (the frequency of use of each verb) were taken into consideration. The three main verb classes in Norwegian vary in type frequency, as illustrated in Figure 2 (cf. also Endresen and Simonsen, this volume, section 4.2). Type frequency three main classes \\feak

Strong

LowTyfi· (4%)

Smaller 0*8)

Larger (WL)

MdTyft (40%)

HghTyfi(56%)

Figure 2. Relative type frequency of the main verb classes in Norwegian

136

Hanne Gram Simonsen

Verbs from all the three main classes were included in the test. For the weak verbs, half were taken from the larger weak class (WL) and half from the smaller weak class (WS). For the strong verbs, subclasses and their relative type frequency were also taken into account (cf. Endresen and Simonsen this volume, section 4.2.1): two groups of verbs were chosen from two of the subclasses with a relatively higher type frequency among the strong verbs (the drikkedrakk class with 40 members, [SI], and \h& flyte-fl0t class with 27 members, [S2]). The third group of strong verbs was chosen as a mixture of different subclasses with few members each (less than 9; [S3]). For each group of verbs, items were included to represent both high token frequency and low token frequency in an approximately equal amount, excepting items from the third subgroup of strong verbs (S3), which all had a relatively high token frequency. (Good token frequency measures are hard to find for Norwegian, in particular for spoken language. See Ragnarsdottir, Simonsen, and Plunkett 1999 for a description and discussion of the token frequency counts used.) As for phonological factors, both phonological coherence/ openness within each inflectional class and phonological similarity/ difference between different inflectional classes were taken into account. Phonological coherence/openness is a measure of the degree of phonological homogeneity of the members within each verb class and as such an indicator of how predictive the phonological properties of the infinitive or present tense in each class are of the past tense form. The weak verb classes in Norwegian differ with respect to phonological coherence/openness, in that the WL class is more open than the WS class (cf. Endresen and Simonsen this volume, section 4.2.2). The strong verbs also differ in this respect, as exemplified by the number of different vowels in their infinitive stems. The flyte-fl0t subclass is phonologically very coherent with one single infinitive stem vowel (/y:/) for all verbs, while the drikkedrakk subclass is less coherent with three different infinitive stem vowels (III, /e/, or /y/).

Past tense acquisition in Norwegian

137

Measures of phonological similarity/difference may for instance be developed on the basis of whether the verbs from different classes have infinitive stems which rhyme or not. In the test, this factor was built in for the weak verbs through splitting each of the main weak classes into two subgroups—one containing verbs rhyming with strong verbs, and one containing verbs not rhyming with any strong verbs. The test contained the seven groups of verbs listed in Table 2. Table 2. Verb classes included in the test • 3 groups of strong verbs (9 verbs in each group): 1 SI (HiTyFr; LoPhoC) 2 S2 (HiTyFr; HiPhoC) 3 S3 (LoTyFr; LoPhoC)

Nldlyl ~ /a/ sitte-satt /y:/ ~ /0:/(/0y/) fryse-fr0(y)s Strong verbs of different, I0pe-I0p, le-lo small classes • 4 groups of weak verbs (8 verbs in groups 4-6; 9 verbs in group 7): 4 WLR (Larger weak class rhyming with strong) 5 WSR (Smaller weak class rhyming with strong) 6 WLN (Larger weak class not rhyming with strong) 7 WSN (Smaller weak class not rhyming with strong)

-/et/or-/a/

titte-tittet/titta

-/te/, -lad

lyse-lyste, gre-gredde

-/et/ or -/a/

kaste-kastet/kasta

-/te/, -/de/

kj0re-kj0rte

Results were calculated on the basis of these seven verb groups, and they will occasionally be referred to in the presentation below. However, since only a few of the differences between these subgroups turned out to be significant, a less fine-grained division will most often be sufficient, i.e., a distinction into the three main verb classes: S (strong verbs), WS (the smaller weak class), WL (the larger weak class).

138

Hanne Gram Simonsen

3. Results 3.1. Correct performance There was a clear developmental progression in the overall correct performance, as shown in Table 3. The mean percentage of correct answers increased and the standard deviation decreased substantially with age. (Of course, age is a notoriously poor indicator of development, but we have tried to eliminate the worst features of overlap by using 2-year gaps.) All differences in performance between age groups were significant.2 Table 3. Correct answers for all verbs by age group Mean SD

4-year-olds 51% 19%

6-year-olds 72% 12%

8-year-olds 90% 8%

However, the differences in performance were not the same across the three main verb types, as shown in Table 4: Table 4. Correct answers by verb type and age group

Age 4 years 6 years 8 years

WL class Mean SD

WS class Mean SD

S class Mean SD

85 % 94 % 96 %

47 % 71% 91 %

33 % 60 % 86 %

20 % 8% 5%

27 % 18% 10 %

24 % 22 % 14 %

The children scored highest on the verbs from the larger weak class (WL), followed by the smaller weak class (WS), and lowest on the strong verbs (S). The differences in performance were significant between all age groups for all main verb types—except for the WL verbs, which reached a ceiling at age 6 (i.e., the difference between ages 6 and 8 was not significant). For the WS verbs, performance increased significantly from 47% at age 4 to 71% at age 6, but it was still significantly behind the

Past tense acquisition in Norwegian

139

performance for the WL verbs at this age, clustering with the strong verbs. It was not until age 8 that the children performed equally well on the verbs from the two weak classes. Performance for the strong verbs lagged significantly behind that for the WL verbs in all age groups. At age 4, it was significantly behind that for both weak classes; at age 6, it clustered with that for the WS class significantly behind that for the WL class, and at age 8 it was still significantly behind that for the WL class verbs. As for the subgroups of strong verbs, the performance for subgroup S2 was the lowest throughout the period tested, and significantly lower than for most other groups. The effect of getting the infinitive or the present tense as a cue seemed to be important for the younger children, but not for the 8year-olds, as illustrated in Table 5—the children who got the present tense as a cue (where the word tone provides information about whether the verb is weak or strong) performed better than those who got the infinitive (where the word tone does not provide such information): Table 5. Correct answers by cue and age: Percentage correct with either infinitive or present tense as a cue. Infinitive Age

Mean

SD

4 years 6 years 8 years

47% 69% 90%

19% 12% 9%

Present tense Mean SD

56% 78% 90%

18% 8% 6%

Statistical analyses revealed a significant main effect for cue, but when each age group was considered separately, this effect was significant only for the 6-year-olds (for the 4-year-olds, there was probably too much variation for the difference in performance to be significant).3 As mentioned earlier, the weak verbs were split into two groups, rhyming and non-rhyming. There were no significant differences in correct performance between rhyming and non-rhyming groups of weak verbs. However, for the 8-year-olds the rhyme effect was close

140

Hanne Gram Simonsen

to being significant, and interestingly, in a replication of this experiment performed on adult Norwegian subjects by Kirsten Meyer Bjerkan, there was a significant rhyme effect, with consistently and significantly better performance on the non-rhyming weak verbs (Simonsen and Bjerkan 1998).4 On the other hand, looking at effects of token frequency on correct performance, this performance was significantly better on high token frequency verbs than on low token frequency verbs, for all the children at all ages. As for the different verb classes, token frequency was significant for strong verbs and for WL class verbs, but not for WS class verbs. However, the token frequency effect seems to diminish with age: for adult Norwegian subjects no main effect of token frequency was found (Simonsen and Bjerkan 1998).

3.2. Errors The errors made by the children were classified into 6 main types: 1. Generalisation of the weak large pattern 2. Generalisation of another weak pattern 3. Generalisation of a strong pattern 4. No change of stem 5. No answer 6. Non-past form

(GEN>WL) (GEN>WS) (GEN>S) (NO CHANGE) (NO ANSWER) (NON PAST)

As was the case with correct performance, the types of errors made by the children also changed with age, as illustrated in Table 6. (Note that the percentages of errors in this table add up to 100% of the errors, not of the total performance. The general error level in each age group is indicated in the column "average number of errors", referring to the average number of incorrect forms out of the 60 verbs contained in the test.) By far the most frequent error types were the types 1-3, overgeneralisations of one of the main inflectional patterns. One exception was found in the 4-year-olds, who made quite a few NON

Past tense acquisition in Norwegian

141

PAST errors, where infinitive or present tense responses were given instead of a past tense form. These errors occurred in all verb classes and can be interpreted as a general error strategy of "repeating the cue"; this error type more or less disappears at age 6 and will not be discussed further here. Table 6. Error types: Percentages of all errors across verb groups

Age

Average number of errors (max 60)

E #1 GEN >WL

R #2 GEN >WS

R #3 GEN >S 2% 9% 13%

4 years

29.4

44%

12%

6 years

16.7 6

52% 37%

27% 44%

8 years

Ο #4 NO CHANGE 1% 2%

R #5 #6 NO NONANSPAST WER FORM 4% | 38% | 2% 7% 2% 4%

Focussing on the overgeneralisation errors, we see that at ages 4 and 6, GEN>WL was clearly the dominant overgeneralisation error type. However, GEN>WS errors increased steadily, and already at age 6 they accounted for more than 1/4 of the errors. At age 8, this tendency continued, and as the percentage for GEN>WL decreased, that for GEN>WS increased, so that the error type GEN>WS surpassed GEN>WL in the 8-year-olds. And again, interestingly, in the replication of this experiment made on adult subjects, GEN>WS was found to be the dominant error type, accounting for 45% of the errors as compared to 10% of GEN>WL and 20% GEN>S (Simonsen and Bjerkan 1998). So-called "irregularisation" errors, i.e., the overgeneralisation of certain strong verb patterns (GEN>S), were a much less frequent error type, but they did occur, increasing slowly with age. The developmental pattern of overgeneralisation errors, here including the adults, is illustrated in Figure 3. As for the distribution of error types across verb classes, GEN>WL errors were found in all verb types, with one exception: verbs with a monosyllabic infinitive showed no such errors. Among the strong verbs, the subgroup SI in which all verbs have a short

142

Hanne Gram Simonsen

infinitive stem vowel, overwhelmingly got GEN>WL errors, while subgroup S2, in which all verbs have a long infinitive stem vowel (as well as the verbs from subgroup S3 with a long infinitive stem vowel) showed a more balanced distribution between GEN>WL and GEN>WS errors, with an increasing amount of GEN>WS errors with age. GEN>WS were also an important source of errors in WL verbs. GEN>S errors occurred both in weak and strong verbs, but in a relatively small set of verbs, and only in verbs which rhyme with strong ones—thus there was an effect of rhyme here after all. In the strong verbs, such errors typically occurred either in verbs which have an irregularity within their own class (yielding, for example, lyve 'tell a lie' -I0yv instead of Iyve-I0y), or they were inflected according to another strong pattern with a higher type frequency than their own (for example, ligge 'lie down' -lagg instead of ligge-la.) •GEN>WL •GEN>WS -GEN>S

0% Age 4

Age 6

Age8

Adults

Figure 3. Development of overgeneralisation errors: Percentage of total errors

4. Discussion It seems evident from these results that input factors play an important role for the acquisition of past tense in Norwegian.

Past tense acquisition in Norwegian

143

4.1, Type frequency Type frequency seems to play a major role in correct performance. All the three main inflectional classes were acquired in an order reflecting their type frequency: WL>WS>S. Thus, the larger the number of verbs in an inflectional class, the earlier it was acquired. However, for the different subgroups of strong verbs tested, we did not find any positive effect of type frequency on correct performance. Subgroup S2, which has a relatively high type frequency among the strong verbs, was acquired later than subgroup S3, which has a lower type frequency. It is possible that the differences in type frequency between these subclasses are not large enough to have an effect. In addition, token frequency may play a role here: overall, the S2 verbs have a lower token frequency than the S3 verbs. Type frequency is clearly also relevant for overgeneralisations. The overgeneralisation errors are relatively well synchronised with correct performance: all the three main classes form bases for overgeneralisations, and to a large extent their relative importance as sources of overgeneralisations reflect their type frequency. But since the WS class becomes increasingly more important as a source of overgeneralisation with age, other factors (such as phonological factors) must also contribute to the pattern of performance.

4.2. Phonological factors In fact, to a large extent the phonological properties of the verb classes seem to line up with the pattern of type frequency in Norwegian. For instance, if we put the three main inflectional classes on a scale of segmentability and salience of their past tense forms, the ranking will match a type frequency ranking. The larger weak class is inflected with a syllabic suffix, it has one single past tense suffix, starting with a vowel, and there is no change in the stem when the suffix is added. (Examples: haste 'throw' kast-et; tine 'peep' - titt-et, see also Table 2 above.) The smaller weak class is also inflected through suffixation, but there are two

144

Hanne Gram Simonsen

different suffixes to choose between; the suffixes start with a consonant, possibly making them less easily segmentable from the stem; and often assimilations and other smaller stem changes like vowel shortening also occur. (Examples: lyse /ly:se/ 'light' - lys-te /ly:ste/; leve /le:ve/ 'live' - lev-de /levde/; kj0re /90:re/ 'drive' - kj0rte /90:te/; see also Table 2 above.) The strong verbs are found at the lowest end of the segmentability scale, being inflected without a syllabic suffix, but most often with a change in the stem vowel. (Examples: sine 'sit'-sa/ί; sove 'sleep'-sov.) These characteristics are summed up in Table 6 below. Table 6. Type frequency and salience/segmentability Type frequency Salience/ segmentability

WL class High •56% High •suffix: syllabic, vowel initial •stem: no change

WS class Mid •40% Mid •suffix: syllabic, consonant initial •stem: often consonant assimilation and vowel shortening

Strong verbs Low •4% Low •no syllabic suffix

• stem: vowel change

As already mentioned, the verb classes also differ by way of degree of phonological openness and coherence. The WL class is clearly the most phonologically open, while the WS class has subclasses with a high phonological coherence, as well as a more restricted set of phonological patterns involving prosody (vowel length), vowel quality, and consonants. Since these patterns are not absolute, but only tendencies, they may take more time (and more verbs) to discover than more clearcut class patterns. This way, we seem to have evidence of a critical mass effect: before the child is able to generalise over these subtle patterns, a substantial number of verbs have to be acquired which illustrate them. This may help explain the developmental profile of the weak verbs: the higher segmentability and the higher type frequency of the WL class, combined with its

Past tense acquisition in Norwegian

145

phonological openness, makes this class the easiest one to acquire and generalise from. However, given more time and more verb forms, the higher phonological coherence of the WS class makes it possible to establish phonological patterns for children to generalise over - both correctly and incorrectly. As for the strong verbs, phonological coherence does not seem to have a positive effect on correct performance: in fact, the subclass with high coherence (and high type frequency), S2 (the fryse-fr0s class), was the one acquired latest of all. As indicated above, token frequency may play a role here, since this class has a lower global token frequency than the other strong classes. However, phonological coherence does seem to have an effect on the overgeneralisation errors made: the strong subclasses with the highest phonological coherence served as the most important bases of GEN>S errors. Effects of phonological similarity were found only in the older children. The rhyme effect—i.e., a poorer performance on weak verbs rhyming with strong verbs than on those not rhyming with strong verbs—did not even reach significance in the 8-year-olds, but was significant in the adults. This indicates that the strong verbs and their phonological patterning need to be mastered for this effect to emerge. The error patterns corroborate this assumption: for the strong verbs, all the verbs generalised to strong classes rhyme with strong verbs, showing a similarity effect, but again, such errors do not occur before the strong verbs are mastered well enough to be generalised over. Similarity effects were also found in the weak verbs: overgeneralisation errors of verbs with a monosyllabic infinitive were never of the GEN>WL type. Furthermore the distribution between GEN>WL and GEN>WS, in particular in the older children, was to a large extent predicted by the predominant vowel length in the weak classes. The information provided in the cue by the word tone concerning verb class membership, helping to distinguish between strong and weak verbs, turned out to be important for the younger children but diminished with age, having no effect for the 8-year-olds. This cue is salient, but can only help the child distinguish between a strong and a weak verb; it does not indicate which of the two weak inflections or

146 Hanne Gram Simonsen

which one among the strong inflections is needed. Thus, to inflect the verbs correctly, the child needs to discover other, more detailed patterns for identifying each of the verb classes, and at that point the tone cue loses some of its importance. Interestingly, in a connectionist network simulation of past tense acquisition in Norwegian, a similar pattern was found: word tone information was most important for correct performance early in the training period (N0klestad 1996). And in a similar experiment to the present one for Italian children, a parallel cue effect was found to be significant only for the 4-year-old children (Matcovich 1998).

4.3. Token frequency Finally, the results showed a clear token frequency effect on correct performance in all age groups, and notably this effect was found not only in the strong verbs, but in the WL class as well—the class which would otherwise be the closest candidate to functioning as a default class. (I have no explanation for the lack of token frequency effect in the WS class). This indicates a similarity in the processing of verbs irrespective of their status as weak or strong. As opposed to the phonological similarity effects, the token frequency effect seems to diminish with age, with the adults showing no such effect. A token frequency pattern with a similar developmental profile was found for Italian children (Matcovich 1998, see also Matcovich this volume). Again it seems possible that in the course of development, when enough verbs are acquired for patterns to be established, the strength of the patterns for the individual verbs may be overridden by the strength of the pattern representative of the class as a whole.

5. Conclusions In sum, the results support the position that input factors such as phonological and frequency factors play an important role in the acquisition of past tense morphology in Norwegian, both in terms of

Past tense acquisition in Norwegian

147

correct performance and the types of errors made. Importantly, these factors also seem to play a role for all verb classes, not only for the strong verbs. All the three main verb classes are used as bases for generalisation: the weak classes are used more often, but there is no single class which is used as a default across all age groups. In the course of development, children gradually use more classes productively. The more verbs and verb classes the child acquires, the larger the basis for generalisation, and the relative importance of each class in this respect changes during development. There is little evidence of clear-cut differences between the classes of an either-or type, but rather the differences seem to be of a gradual (and dynamic) nature. How do these facts fit with the two different models mentioned in the introduction? For the dual mechanism model, the fact that there are three main classes (WL, WS, and S) and not only two (W and S) already presents a problem: where should the third (WS) class be placed—in the rule mechanism, or in the lexical retrieval mechanism? At the outset, the children seem to treat verbs from this class in the same way as the strong verbs, indicating that the WS verbs are retrieved from the lexicon. However, in the course of development children gradually come to master these verbs better, and also start overgeneralising increasingly more to this pattern—and at the adult level the WS class becomes the clearly dominant source of overgeneralisation errors. This shift complicates matters for the dual mechanism model: although an analogical component in the lexical retrieval mechanism might be posited to allow for generalisation effects to occur, the fact that such effects become predominant seems to contradict or at least strongly stretch the idea of the lexical retrieval mechanism as a "minor" mechanism. The larger weak (WL) class is the most obvious candidate for the default, rule-governed class in Norwegian, and at the beginning of the development it does seem to function as a default for the children, being clearly the class they master best, as well as the dominant source of overgeneralisation errors. However, according to the dual mechanism model neither frequency factors nor phonological factors should influence the default class, yet both the correct

148 Hanne Gram Simonsen

performance and the overgeneralisation errors in our Norwegian data seem to contradict this claim. As for type frequency and phonological openness, these factors to a large extent coincide in the different verb classes, with the WL class as the one which is both phonologically most open and the class with the highest type frequency. The way the Norwegian verb classes are structured, then, it is difficult to single out the effects of each of these factors and to separate those again from the possible effects of the WL class as a default. Since the WL class is so open as to allow nearly any phonological pattern, phonological similarity effects in the errors are also hard to identify. However, the few phonological restrictions found in this class are in fact mirrored in the errors. Verbs with monosyllabic infinitives are not found in this class, and accordingly there are no errors of overgeneralisation of such verbs to this class. In addition, the distribution of vowel length in the two weak classes is to a large extent mirrored in the distribution of overgeneralisation errors to these two classes, in particular in the older children. Furthermore, the results show token frequency to have a significant effect on performance already from the beginning of the development registered in the Norwegian children, not only for the strong verbs, but also for the verbs from the larger weak class— contradicting the claims made by the dual mechanism model. The results show that the relative importance of the different verb classes as bases for generalisation changes during development, indicating that the differences between the classes are gradual rather than of a dichotomous nature. This seems poorly explained through a model with a strict partition into two qualitatively different mechanisms. Furthermore, the manner in which the relative role of the different input factors also changes in the course of development makes it worth while turning to the other model, the single mechanism model. This model sees all inflectional patterns built up in essentially the same way, through usage of the different verb forms. Generalisations are formed over verb forms with perceived similarities, and the larger the type frequency, the stronger the schema. Thus input factors

Past tense acquisition in Norwegian

149

should play a similar role for all verb classes. The developmental pattern found here, with type frequency as an important predictor of correct performance, a relatively well synchronised development between correct performance and overgeneralisation patterns, and token frequency playing a role for both strong and weak verbs,5 are all in accordance with such a model. The same goes for the evidence of changing importance of different input factors during development: While token frequency, salience, and the effect of word tone play an important role earlier in development but become less important with age, other factors like the effects of phonological patterning of the different classes and rhyme effects do not show up at the beginning, but become increasingly important with age. These results can all be reasonably explained in a model based on increased entrenchment of verb forms and Schemas formed across them through usage. When enough verbs from a verb class have been acquired to form a pattern or schema, the importance of the token frequency of each verb may become less important than the strength of the schema. The information carried by the word tone is salient, but can only partly help the child in identifying class membership; more subtle phonological information is needed to assign the verbs to specific strong or weak inflectional patterns. However, since many of the phonological patterns found are indeed not absolute, but only tendencies, a substantial amount of verbs have to be acquired for these patterns to be identified, and well enough entrenched for the child to start generalising over them. This may explain the late but increasing importance of phonological patterning during development. Thus, this evidence of a dynamic interaction between different input factors is consistent with an exemplar- or usage-based, single mechanism model of acquisition, storage, and processing. To what extent these results for Norwegian find a parallel in other languages, is of course an empirical question. Ragnarsdottir, Simonsen, and Plunkett (1999) show that a similar acquisition pattern applies to Icelandic—another Scandinavian language. Experimental studies of Italian, still another language with less of a dichotomous distinction into regular and irregular patterns than

150

Hanne Gram Simonsen

English, have shown that irregular and regular verbs have the same priming patterns, and that phonological factors and frequency factors seem to play a similar role for regular and irregular verbs—these results have been found both for adults (Orsolini and Marslen-Wilson 1997) and for children (Matcovich 1998, this volume). This might indicate that English, with its clearcut division which has served as the starting point for the dual mechanism model, is more of an extreme case, and as such less suited as a universal template. However, even for English, recent experimental studies have yielded parallel results to the ones found here. Marchman (1997), using the same elicitation task as the present one for past tense forms in English-speaking children, found that both regular and different irregular patterns were used productively, and that their use was influenced by sets of item-based properties, notably frequency and phonological factors, which all contributed through competing and converging pressures to the productivity patterns found.

Notes 1. I am indebted to Kim Plunkett and Hrafnhildur Ragnarsdottir for their contribution to the research behind this paper. I also wish to thank two anonymous reviewers for helpful comments. 2. Statistical significance was tested through ANOVAs and Tukey post hoc tests for multiple comparisons, with a significance level of pWL is the productive use of the larger weak class, etc), and Non-Past errors (that is, to give the infinitive or present tense instead of the past tense form). The 4-year-olds generally either overgeneralised to the larger weak class, or they gave a NonPast response. The 6-year-olds almost never used a Non-Past form. Their most frequent error type was Gen>WL, but they also had a number of Gen>WS. The 8-year-olds overgeneralised slightly more to the WS than to the WL, and both these weak classes were used productively to a large extent. There was also a slight increase in the Gen>S (productive use of strong patterns) with age. Cf. Table 2.

Past tense inflection in SL1 children

157

Table 2. Error types, ND children

AGE

ND4 ND6 ND8

GEN >WL

GEN >WS

GEN

44% 52% 37%

12% 27% 44%

2% 9% 13%

>S

NO CHANGE 1% 2%

NO ANSWER 4% 2% 2%

NONPAST FORM 38% 7% 4%

2.3.2. SLI children The SLI children generally had a poorer performance than the ND children, and they often showed particular problems with one of the verb classes, either WS or S. Table 3 shows the correct performance of all the SLI children. Since they are too few to constitute a representative group of the SLI population, they are presented individually. The variation is large, both with respect to the total correct scores and the scores on the separate verb classes. Generally, the SLI children have a lower correct score than the ND children of the same chronological age. Table 3. Correct answers, SLI children Subject Age Total 6;1;4 53% Ida 40% Alex 6;6;20 6;7;21 Tim 63% 6;9;14 72% Rita 78% Henry Michael 67% 8;2;0 68% Albert 8;2;7 Karen 8;8;28 90%

WL class 88% 63% 75% 88% 88% 94% 88% 100%

WS class 35% 53% 71% 94% 82% 29% 88% 82%

S class 44% 19% 52% 48% 70% 74% 44% 89%

The types of errors they make, shown in Table 4, are to a large extent synchronised with the correct performance, in that they overgeneralise the most to the class on which they have the highest correct score. All of them are able to use verb patterns productively,

158

Kirsten Meyer Bjerkan

but most of them also have a number of Non-Past responses, as was the case for the ND 4-year-olds. Table 4. Error types: numbers across all verb groups, SLI children Subject

Total no. of errors (max 60)

GEN >WL

Ida

28 36 22 17 13 20 19 6

8 16 5

Alex

Tim Rita Henry Michael Albert Karen

5 8 5 2

GEN >WS

8 6 6 2 2 7 3

GEN >S

NO CHANGE

NO ANSWER

NONPAST FORM

1

2

3 2 1

14 1 6 9 6 5 5

4 1 1

1

WRONG VERB

9 4 2

1

3. Presentation of two SLI children, Ida and Alex In this article, I will focus on two of the six year old SLI children, Ida and Alex. I have chosen these two because they differ the most in their approach to this task.

3.1. Child 1, Ida Ida (age 6;1;4) is the youngest of the 6-year-olds. Her overall score is 53% correct answers, which is below the mean for the ND 6-yearolds (72%), but approximately the same as the ND 4-year-olds (51%). Considering the different verb classes separately she performs fairly well on the WL class, but not on the WS and S classes (Table 3). When compared to ND children of her own age, she is late in acquiring all verb classes. As for the ND 4-year-olds, she is better than them on the strong verbs and at about the same level on the WL class, but is behind their mean on the WS class (although she is within one standard deviation below). Her particular problem is the

Past tense inflection in SLI children

159

WS class. It seems like she has noticed the regularities of the large weak class, but not of the small one. This is confirmed by her errors: She does not have one single overgeneralisation to the WS class. As is shown in Table 4, her most frequent error type is the use of a Non-Past verb form. Half of her errors are of this type. This strategy is used both for strong and weak verbs, but most of the weak verbs are from the WS class. All the answers are in the present tense, not the infinitive; she only repeats the input form. The fairly high number of GEN>WL shows that this pattern is acquired and that she is able to use it productively. Both strong verbs and verbs from the WS class are overgeneralised to this class. She is, however, not totally unable to see any other patterns, she has one GEN>S; a verb from the small weak class, spise 'eat' is inflected as /spasis/ according to one of the most frequent strong patterns. We can, however, not draw any conclusions on the basis of one occurrence. Ida is the only SLI child who makes No Change-errors, that is the use of the verb stem as the past tense form. This error occurred quite frequently in Bybee and Slobin's (1982) English data, but is rare among all the Norwegian children, which may be due to morphological differences between Norwegian and English. Ida inflects two verbs according to the No Change pattern. Both of them are strong, and both belong to the same strong class: /fry:s/ instead of fr0s 'froze' and /ly:v/ instead of I0y 'lied'. The high number of Non-Past errors (the use of the present tense) shows that she does not have an obligatory past tense marking on all verbs, although the grammatical concept of past tense is mastered to a large extent.

3.2. Child 2, Alex Alex (age 6;6;20) has the lowest percent of correct answers of all the SLI children, with a total score of only 24 of the 60 verbs correct, or 40%. As is shown in Table 3, his main problem is the strong verbs. Like all the other children, he performs best on the WL class. He is

160 Kirsten Meyer Bjerkan

well below the mean for ND 6-year-olds on all the verb classes, and also below the ND 4-year-olds on the strong verbs and the WL class. On the WS class his performance is at about the same level as the ND 4-year-olds with 53% vs their 48%. What is interesting about Alex, is that even though his score for correct answers is fairly low, he does mark past tense obligatorily. He makes only one Non-Past form error, and in this respect he is the best in his SLI age group, and also better than three of the four SLI 8year-olds. This one error is for the verb reparere 'repair' which belongs to a subclass of the WS class which has very low frequency in language spoken to children. The majority of his errors are overgeneralisations to the WL class. He has a very productive past tense pattern which says that past tense ends in -a, and this pattern is applied both to strong verbs and to verbs from the WS class (also in his spontaneous speech, not just in the test). He overgeneralises to the WS class, showing that he has acquired this pattern, too, at least partly. Mainly strong verbs are overgeneralised to this class. Of the eight overgeneralisations, there are six strong and two weak verbs. Another frequently occurring error is the use of a wrong verb. The verbs he uses are always synonymous to the target verbs, and they are always in the past tense. For Alex, morphological marking of past tense is obligatory.

4. Discussion The only similarity between Ida and Alex is that both perform best on the large weak class. For Ida, the main difficulty is the smaller weak class, and for Alex it is the strong verbs. They are both able to overgeneralise to the larger weak class, but only Alex uses the smaller weak pattern productively as well. Ida's default answer is to repeat the input present tense; Alex always uses the past tense. We will now turn to a discussion of the data related to the concepts of symbols and Schemas (see Langacker 1987a). A

Past tense inflection in SLI children

161

linguistic symbol consists of a phonological and a semantic pole, and on the basis of all the symbols we hear, we create Schemas. Possible Schemas for weak past tense inflections in Norwegian (adapted from Endresen, Simonsen and Sveen 1996: 242) are: WL: [/2...a/ "~, past tense"] WS: [/2...te/ "~, past tense"]

[A..de/ "~, past tense"]

The schema is to be read as follows: The slashes indicate the phonological transcription of the verb. The upper case 2 shows the word tone (Norwegian being a pitch-accent language with two word tones). The three dots serve as place holders for the stem of a verb, whereas /a/, /te/ and /de/ are the past tense suffixes in the weak verb classes. The quotation marks indicate the verb's meaning. The tilde is an abbreviation of its lexical meaning, whereas its grammatical meaning (here, past tense) is specified. Once such a schema is created, it can be instantiated to new verbs. Some instantiations are conventional in a language community, whereas others are idiosyncratic or unconventional, occurring only in one or a few individuals. Examples of conventional and unconventional uses of weak past tense are: WL: [/2...a/ "~, past tense"] -> [/2hopa/ "jump, past tense"] (conventional) WL: [/2...a/ "~, past tense"] -> (/2drika/ "drink, past tense") (unconventional) WS: [/2...te/ "~, past tense"] -» [/2lekte/ "play, past tense"] (conventional) WS: [/2...de/ "~, past tense"] -> (/2lede/ "laugh, past tense") (unconventional) The conventional past tense forms for 'drink' and 'laugh' are drakk and lo, respectively. In order to use a pattern productively, one must not only have created a schema, but the schema must also be entrenched to a certain extent. A symbol is entrenched by encounter and by use, and in order

162

Kirsten Meyer Bjerkan

for it to be entrenched, one must pay attention to it. A lack of attention may, according to some researchers, be an explanation for the problems of the SLI children. But as we have seen in the data presented here, both children are able to create Schemas and use them productively. SLI children are often hypothesised to have particular problems in discovering patterns and regularities, and are thought to learn all inflected forms of words by rote. E.g., Gopnik (1994) and van der Lely and Ullman (in press) found indications of this in their data. Their subjects performed better on strong than on weak verbs. They claim that SLI children's main problem is the formation of rules. This hypothesis is contradicted by my data. Ida does perform better on the strong verbs than on the WS class, but she is much better on the WL class, and she is able to use this pattern productively. Alex's performance is an even clearer counterexample. He does not master the strong verbs at all, he has a low score for correct performance and no overgeneralisation to the strong class. The weak patterns, on the other hand, are used productively to a large extent. For Alex, these Schemas for weak inflection have high degrees of entrenchment, particularly the WL class, which is his default pattern. On this basis, we would expect his correct performance on the WL class to be high, but it is not. This is mainly due to his use of wrong verbs, which is probably not caused by lack of knowledge about the inflection of the target verb, but rather by lack of concentration or attention; he pays more attention to the semantic than to the phonological pole of the symbol (see Kirchner and Klatzky 1985). He even sometimes inflects the new verb according to the schema of the target: e.g. for the target kästet 'threw' he answers hiva, a strong verb inflected according to the WL schema. The concept of morphological rules is problematic in accounting both for the individual children's performance and for the difference between them. A rule is expected to be either present or absent, and not only partly acquired. Schemas, on the other hand, can have different degrees of entrenchment. The schema for the WL class is well entrenched in both children, but more in Alex than in Ida. Alex

Past tense inflection in SLI children

163

also has a fairly well entrenched schema for the WS class, whereas Ida has not. She may have created a schema, but that is impossible to know as long as she has a low score and does not overgeneralise. If it exists, it is not well entrenched. Similar Schemas may, of course, be created for the different strong subclasses. The only indication we find in these data of a schema for strong inflection is Ida's one overgeneralisation, and that is not enough to make any claims in either direction. The two other SLI 6year-olds I have not presented here did not overgeneralise strong patterns, either, but the SLI 8-year-olds did, showing that also SLI children are able to create different Schemas, and extend them to new verbs.

5. Conclusion SLI children, even though they do not perform as well as their normally developing peers, are able to create Schemas based on verbs they hear, and they are able to extend these Schemas to other verbs. There is no evidence for an impaired ability to form generalisations. Of course, we cannot generalise from two cases to a whole population of SLI children. The variation is vast. And because of this variation, cognitive grammar is a well suited framework, with its Schemas which are often conventional, but which can be extended unconventionally, and the notion of entrenchment which is always individual.

A connectionist model of past tense acquisition in Norwegian Anders N0klestad

In cognitive linguistics theories and in connectionist models, both regular (or weak) and irregular (or strong) inflectional morphology are handled by the same kind of mechanism, a mechanism that is sensitive to the phonological form and the type and token frequencies of items in the linguistic input. In this paper 1 present a connectionist model of past tense acquisition in Norwegian and compare the model's behaviour to results from past tense elicitation experiments with Norwegian speakers. These comparisons, as well as investigations of the model's treatment of novel verbs, show that the model reflects many aspects of past tense inflection among Norwegian speakers, thus lending support to the single mechanism view on the acquisition and processing of inflectional morphology.1

1. Introduction Over the past 40 years, a considerable amount of research has been carried out on the acquisition and processing of inflectional morphology, and in particular on the acquisition of the English past tense. English verbs can be relatively easily divided into two main classes based on past tense inflection: regular, or weak, verbs like jump, whose past tense is formed by adding the suffix -ed to the stem, and irregular, or strong, verbs, which form their past tense either by a bare stem (such as hit), by some modification of the stem (sing-sang; bring-brought) or by suppletivism (go-went). The weak inflection can be applied to any novel or nonsense verb, even to verbs that violate the phonotactics of English, while generalisation of strong verb classes is much more restricted (Prasada and Pinker 1993). Furthermore, when children make overgeneralisation errors,

166

Anders N0klestad

i.e., when they apply the wrong inflectional pattern to an existing English verb, they almost invariably overgeneralise, or overapply, the weak pattern (Marcus et al. 1992). This has lead many researchers to believe that strong and weak inflections are carried out by qualitatively different mechanisms: a rote learning or associative memory for storing the past tense of strong verbs, and a suffixation rule that applies to any verb which does not have a past tense form stored in this memory, or whose past tense form the speaker is unable to access (see e.g. Marcus et al. 1992, 1995; Pinker and Prince 1988; Prasada and Pinker 1993). According to such models, only the storage of strong forms should be affected by input factors like type frequency (the number of verbs in the class), token frequency (the frequency of occurrence of a particular verb form) and the phonological form of the verb. Weak inflection should be insensitive to such factors, since it is carried out by a rule applying to all words that are marked as verbs. An alternative account is offered by models such as Bybee's (1985, 1988, 1995a) Network Model and Langacker's (1987a, 1990) Cognitive Grammar. According to these models, strong and weak inflection are carried out by the same kind of mechanism, one that acquires all inflectional patterns by generalising over the linguistic input. Consequently, inflection of both kinds of verbs should be affected by frequency and phonological factors. Single mechanism accounts for English have also been explored on computers through the use of connectionist models (e.g. Plunkett and Marchman 1991, 1993; MacWhinney and Leinbach 1991; Rumelhart and McClelland 1986). These are general processing mechanisms which do not contain any built-in linguistic knowledge, but which are able to learn regularities like verbal inflections by generalising over the linguistic input and by being sensitive to the kind of input factors mentioned above. Most previous studies of past tense acquisition have focused on English. However, the English verbal system, with its clear dichotomy between a very large group of highly regular verbs and a rather small group of fairly irregular verbs, is not representative of the verbal systems found in the world's languages. Hence, if our aim

Norwegian connectionism

167

is to discover the mechanisms involved in the acquisition and processing of inflectional morphology in any language, the English data need to be complemented by studies of languages with different past tense systems. Here I will present some of the results from a series of connectionist simulations of past tense acquisition in Norwegian (see N0klestad 1996 for a more comprehensive account). Since Norwegian does not exhibit such a clear regular/irregular verb dichotomy as is found in English, studies of past tense acquisition in this language might yield valuable contributions to the debate over the kinds of mechanisms involved in morphological acquisition and processing. The simulation results will be compared to data from past tense elicitation experiments with 4-, 6- and 8-year-olds reported by Ragnarsdottir, Simonsen, and Plunkett (1996, 1999), see also Simonsen (this volume), and similar experiments with adults described by Bjerkan and Simonsen (1995, 1996). In these experiments, the subjects were given sentences which contained infinitive or present tense forms of Norwegian verbs, and they were encouraged to give replies that contained corresponding past tense forms. Note that the adults were asked to produce the past tense form of the verb as quickly as possible, so that unlike the children, the adults were required to respond under time pressure. By comparing the connectionist modelling results with the results from these experiments, we will be able to see to what extent the model supports a hypothesis that the experimental data seem to suggest, i.e., that input factors are important for the acquisition of both strong and weak inflection in Norwegian.

2. The Norwegian past tense system Norwegian weak verbs can be divided into two main classes, characterised by different past tense suffixes. Following Ragnarsdottir, Simonsen, and Plunkett (1999), cf. also Simonsen (this volume), I will refer to these classes as the larger weak, or WL, class and the smaller weak, or WS, class, respectively. As the names

168 A nders N0klestad

imply, the WL class is considerably larger than the WS class, and it is also far more productive, in the sense that the large majority of novel verbs are treated by Norwegian speakers as belonging to this class.2 This is a commonly held view among Norwegian linguists (see e.g. Ragnarsdottir, Simonsen, and Plunkett 1999), and it is supported by application of Baayen's (1991, 1992) measures of productivity on the past tense forms found in the bokmäl part of The Oslo Corpus of Tagged Norwegian Texts (OCTNT) constructed at the Text Laboratory, University of Oslo. This corpus contains material from a wide range of text types written in the bokmäl language standard, which is very close to the spoken language variety used by the subjects in Ragnarsdottir, Simonsen, and Plunkett's and Bjerkan and Simonsen's experiments. Only the 11 million words of newspaper, magazine and fiction material in the corpus were used for the present purposes, since these were the categories considered to be most similar to spoken language. The Norwegian past tense system is similar to the English one in that the majority of Norwegian verbs form their past tense by suffixation. However, while all English weak verbs take the -ed suffix, there are a number of different suffixes used by the group of suffixing verbs in Norwegian. Following Simonsen (1993), I consider Norwegian weak verbs to be those which take some syllabic suffix in the past tense, such as hoppe - hoppet 'jump -jumped' and spise - spiste 'eat - ate'. All other verbs, including those with asyllabic suffixes, like binde - bandt 'bind - bound', are considered to be strong. Thus, although weak verbs tend to show far less modification of the stem than strong verbs, the weak/strong distinction used here for Norwegian is based on the syllabicity of the past tense suffix, and not on the amount of stem change or on the degree of phonological conditioning (i.e., the criteria on which the regular/irregular distinction has traditionally been based). There are a number of classes of strong verbs, which either form their past tense by a bare stem or by one of several possible vowel alternations. Baayen's concept of global productivity for a morphological class is defined along two dimensions: a) extent of use, corresponding to the type frequency of the class, and b) degree of productivity,

Norwegian connectionism

169

obtained by counting tokens from the class in a corpus (the token frequency of the class) and dividing the number of hapax legomena (single occurrences) by the token frequency of the class. Table 1. Type frequencies, token frequencies, number of hapax legomena and degree of productivity of the main verb groups in Norwegian as measured on the OCTNT corpus, including and excluding auxiliary verbs All types

Nonaux. types

AU tokens

Nonaux. tokens

Hapax legomena

Degree of productivity (all verbs)

Degree of productivity (non-aux.)

Strong verbs

519 (14.4%)

514 (14.3%)

254947 (61.0%)

115973 (48.7%)

113

0.0004

0.0009

WS class

1374 (38.1%)

1371 (38.1%)

120911 (28.9%)

80332 (33.7%)

337

0.0028

0.0041

WL class

1716 (47.5%)

1716 (47.7%)

42058 (10.1%)

42058 (17.6%)

407

0.0097

0.0097

Total

3609 (100%)

3601 (100%)

417916 (100%)

238363 (100%)

857





Table 1 shows the type frequencies, the token frequencies, the number of hapax legomena, and the degree of productivity of the two weak classes and the group of strong verbs in the OCTNT corpus, including and excluding auxiliary verbs. Examples of verbs from the different verb groups are shown in Table 2. Table 2. Examples of verbs from the different verb groups Infinitive

Present tense

Past tense

Larger weak class verbs hoppe 'jump' /2hope/

hopper /2hoper/

hoppet /2hopet/

Smaller weak class verbs spille 'play' /2spile/ levere 'deliver' /|e'ue:re/ skje 'happen' /'Je:/ kreve 'demand' /2kre:ue/

spiller /2spi|er/ leverer /|e'ue:rer/ jJt/er/'Jeir/ krever /2kre:uer/

spilte /2spijte/ leverte /le'ue:(;e/ skjedde /2Jede/ krevde /2kreude/

Strong verbs synge 'sing' ligge 'lie' I0pe 'run'

synger /'syrjer/ ligger /'Ijger/ I0per /'j0:per/

sang /'sarj/ la /l\p'J I0p /'lp:p/

/2syrje/ /2Uge/

170

Anders N0klestad

Since both type frequency (extent of use) and degree of productivity are highest for the WL class, Baayen's measures rank this class highest in global productivity, followed by the WS class and with the group of strong verbs at the bottom of the productivity scale. Verbs in the WL class have a past tense suffix beginning with a vowel. This suffix is either -et or -a, the choice between them being sociolinguistically or stylistically conditioned. Verbs in the WS class have past tense suffixes beginning with a consonant. The majority of these verbs take the suffix -te, while the rest get a -de suffix in the past tense. Related to the difference in productivity between the weak classes is the fact that they also differ with respect to the phonological variation of their members. Almost any verb can be inflected as a WL class verb. The WS class, on the other hand, puts more restrictions on the phonological form of the verbs. About 80% of the verbs from this class have a stem ending in -er, and there are also other restrictions on verb stems in this class (Hagen 1994; Bjerkan 2000; Ragnarsdottir, Simonsen, and Plunkett 1999): Verbs that take the -te suffix have stems ending in unvoiced stops, nasals, liquids or /s/, while verbs that take the -de suffix have stems ending in /g/, /υ/, /j/ or a long, stressed vowel. Stem-final consonant clusters are very rare in this class. Finally, it should be noted that the WS class is more irregular than the WL class, in that the verbs in this class more often exhibit some modification of the stem—usually a shortening of the root vowel. Thus, we see that in Norwegian there is no clear dichotomy between regular and irregular verbs, which is why I (like Ragnarsdottir, Simonsen, and Plunkett, and Bjerkan and Simonsen) prefer to use the strong/weak classification described above rather than a classification based on regularity.

3. Connectionist models We will now take a closer look at connectionist models; consider the simple illustration of a model in Figure 1 on the next page.

Norwegian connectionism

171

Such models consist of a number of processing units which are organised into a network. Each unit has a certain activity level, which may be taken to represent, for instance, the value of a phonological feature. This means that we can represent a phonological present tense form by a pattern of activity across a set, or layer, of units, such as the input layer in Figure 1. Activity will then spread from the input layer along the connections and make another pattern of activity emerge on the output layer. This pattern can then be interpreted as the network's proposal for a past tense form. output layer

connection hidden layer processing unit

input layer Figure 1. A simple connectionist model

The flow of activity from the input layer to the output layer is determined by the strength, or weight, of each connection. By modifying the connection weights the network can gradually learn to produce the correct past tense form for each given present tense form. Thus, the network's knowledge about verb inflections will emerge gradually in the form of a specific set of connection weights. If the network has to learn to inflect a large number of verbs, this knowledge will have the form of generalisations over the inflections of the individual verbs in the input, generalisations which are similar to the Schemas found in Bybee's model and the schematic units in Langacker's Cognitive Grammar. Of course, such a model will necessarily be extremely simplified compared to the mechanisms found in the minds of real speakers, and the data it is exposed to will be very limited compared to a real

172

Anders N0klestad

linguistic environment. Still, if the behaviour of such models is compared to data on humans' performance on past tense inflection, they can be valuable tools for investigating the role of input factors in the acquisition and processing of inflectional morphology.

4. The Norwegian past tense model A schematic representation of the specific model that was used in the Norwegian past tense simulations is shown in Figure 2a. a)

101 output units

60 hidden units

100 input units 100 input units

b)

oo ooo ο oo oo oo phoneme

phoneme

suffix

stress

word tone

stem Figure 2. Schematic representations of a) the whole Norwegian past tense network, and b) the input layer of processing units

The model's processing units were organised into three layers. Present tense forms were represented on the input layer, as illustrated in Figure 2b. The activation values of some of the units represented

Norwegian connectionistn

173

particular suffixes, while each of the others encoded a phonologically distinctive feature of Norwegian word forms, i.e., either a feature of one of the stem phonemes or the presence or absence of a stress pattern or word tone. The distinctive feature analysis of the stem phonemes was based on Endresen (1991). From the input layer, activation spread to the output layer, and the output was interpreted as a past tense form, again by interpreting the unit values as the values of phonemic distinctive features or as suffix, stress or tone markers. The middle layer of units and the bias unit were not used for explicit phonological coding; they were only there to increase the computational power of the model.

5. Training the model Network training was performed with the tlearn simulator using the back-propagation learning algorithm. The model was trained on phonological representations of 1709 Norwegian verbs taken from a frequency list based on 30 novels from the 1950s, '60s and 70s.3 These verbs were the ones that were listed with a token frequency of 2 or more; the rest were used for testing the network's generalisation to novel verbs, i.e., verbs it had not been trained on (cf. section 5.3). The model was given the present tense form of the verb as input and learned to produce the correct past tense form. Here I will focus on a set of simulations where the model was first trained to a hundred percent correct performance on the 20 most frequent verbs, and then the size of the vocabulary, or training set, was increased gradually, reflecting the vocabulary growth of a child acquiring the Norwegian past tense (see N0klestad 1996 for descriptions of simulations with other training regimes). At a vocabulary size of 100, the vocabulary growth was accelerated, in an attempt to incorporate the so-called vocabulary spun. The verbs were introduced according to decreasing token frequency. Thus, the training set contained only very high-frequency verbs at first and was gradually extended with verbs of decreasing frequency. Since highfrequency verbs are unequally distributed between the different

174

Anders N0klestad

classes, this had the effect that the proportions of both types and tokens from each of the classes varied throughout the training period. Figure 3 illustrates the changes in relative type frequency. Strong vette WS class vats WL class verbs

100 200 300 400 500 600 703 ΘΟΟ 900 1000 1100 1200 1300 1400 1500 1ΘΟΟ 1700 1800

Vocabdarysize Figure 3. Proportions of types from the different verb groups as a function of vocabulary size

At the point where the vocabulary size started to increase, the training set contained 60% strong verbs and 40% verbs from the WS class, but no verbs from the WL class. The proportion of strong verbs then decreased rapidly, while the proportion of WS class verbs increased equally fast. During the rest of the training period the proportion of verbs from these two classes decreased slowly, and the WL class gradually grew to become the one with the highest type frequency. At the end of training, the relative ordering of the type frequencies was the same as that found in the OCTNT corpus, although the WL class was larger and the WS class smaller than in the corpus (56.6% and 30.7% respectively). The proportion of strong verbs was about the same (12.6%). The low initial type frequency of the WL class had important consequences for the model's behaviour, see section 5.2. Figure 4 shows how the proportions of tokens from each of the verb groups changed during training. The relative token frequencies are very similar to the relative type frequencies shown in Figure 3, contrary to what one would expect from the numbers in Table 1. The

Norwegian connectionism

175

reason for this similarity is that the token frequencies used in the simulations were highly reduced compared to the actual frequencies found in the frequency list. The reduced frequencies were obtained by dividing each frequency in the list by a large constant and rounding upwards to the closest integer, leaving only a very few verbs with token frequencies above one. There were several reasons for this frequency reduction. Strong verbs WSdass verbs \M_dass verts

1 0 0 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 8 0 0 9 0 0 1000 1100 1200 1300 1400 1500 1600 1700 1ΘΟΟ

Vocabulary size Figure 4. Proportions of tokens from the different verb groups as a function of vocabulary size

Firstly, the most frequent verbs typically have short duration and receive relatively low stress in connected speech, and thus are phonetically less salient than many lower-frequency verbs. This is especially relevant for Norwegian present and past tense forms, since most strong verbs—which tend to have high frequency—are monosyllabic, while weak verbs are generally polysyllabic. Secondly, although the current model only learns mappings from present tense forms to past tense forms, this does in no way indicate a belief in a similar modularity within the mental linguistic system of a speaker. A natural extension of the single-mechanism view of past tense morphology is the hypothesis that the mechanism responsible for acquiring verb form associations in the human mind is also involved in the processing of other linguistic—and possibly nonlinguistic—input. Such additional input, if included in the simu-

176

Anders N0klestad

lations, would function as "noise" relative to the past tense mappings, and might prevent the network from moving too far into connection weight configurations that are only compatible with verbs of very high frequency. Thirdly, a relatively large proportion of the verbs in the frequency list had to be removed from the training set because they did not fit into the input representation or because they do not occur in the East Norwegian dialect that was being modelled. Since most of these verbs had low frequencies, a number of the remaining verbs would have been far too high-frequent relative to the size of the training set if the original token frequencies had been kept unchanged.

6. Analysis of network performance While the model was being trained, its correct performance was measured, both on the whole training set and separately on the two weak classes and on the strong verbs. The errors made by the model were also analysed.

6.1. Correct performance Figure 5 shows how the overall percentage of correct past tense forms varied throughout the training period. After the initial period of error-free performance, the percentage of correct forms decreases dramatically as the size of the training set starts to increase. This is to be expected, as the network then begins to receive a continuous influx of new verbs. The model quickly recovers, however, and it increases its performance until the training set reaches a size of about one hundred verbs. This is where acceleration of the vocabulary growth begins, which is likely to be one reason for the sudden drop in correct performance at this point. However, it may also be related to the fact that around this vocabulary size the proportion of strong verb types in the vocabulary has dropped to about 30%, and the proportion of WS class verbs has

Norwegian connectionism

177

stopped growing. At the same time there is an increasing proportion of verbs from the WL class, which is a class that the model has not yet learned how to handle, due to the very low type frequency of this class earlier in training. But again the model recovers, and it gradually improves its performance until it reaches a 94% correct performance at the end of training. 100-n

0

100 200 300 400 500 600 700 800 900 100011001200130014001500160017001800

Vocabulary size Figure 5. The network's correct performance on the entire training set

These results should be compared to the data on correct performance from Ragnarsdottir, Simonsen, and Plunkett's and Bjerkan and Simonsen's psycholinguistic experiments, shown in Figure 6. The 4-year-olds in the experiment got 51% of the past tense forms right. For the 6-year-olds, the score was 72%, and for the 8-year-olds it was 90%. The adults scored only a little bit better than the 8-yearolds, giving 94% correct responses. A problem with comparing this developmental profile to the data from the simulations is to decide which stages of the network's performance to compare to which age groups. This decision might have been facilitated by estimates of typical type frequencies in the vocabularies of Norwegian children at different ages. However, since no such estimates are available, we will have to make a rough approximation to an appropriate starting point for comparison.

178 Anders N0klestad

Since the youngest children were 4 years old, they should be well past the point where the vocabulary spurt begins, so we should compare the experimental data to the performance of the model after it has made its second recovery. Now, there do not seem to be any convincing a priori reasons for choosing a more specific starting point for comparison. However, other results, which will be presented below, provide the best match with the experimental data if comparisons are made from about a vocabulary size of 350, so I will choose that as the starting point for all comparisons.

Age 4

Age 6

Age 8

Adults

Figure 6. The human subjects' correct performance on the entire set of verbs. Adapted from Bjerkan and Simonsen (1996)

At the 350 verb mark, then, the network has already reached 87% correct performance, which is much better than the 4-year-olds in the experiments, and consequently, its subsequent performance increase is considerably smaller than the increase with age that was found in the experiments. Nevertheless, the network's performance curve shows a continuously decreasing slope, and in this respect it reflects the tendency found in the psycholinguistic experiments, where the degree of improvement decreased with age. Figure 7 shows the network's separate performance on strong verbs and on verbs from the larger and smaller weak classes, and

Norwegian connectionism

179

corresponding data from the psycholinguistic experiments are found in Figure 8.

-Strong verbs •WS class verbs •WL class verbs

1 0 0 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 Θ 0 0 9 0 0 1000 1100 1200 1300 1400 1500 1600 1700 1800

Vocabulary size Figure 7. The network's correct performance on the different verb groups

Strong classes WS class WL class

Age 4

Age 6

Age 8

Adults

Figure 8. The human subjects' correct performance on the different verb groups. Adapted from Bjerkan and Simonsen (1996)

If we continue to use a vocabulary size of 350 as our first point of comparison, we see that from this point onwards the relationship between the model's hit rates corresponds to what was found for both

180 Anders N0klestad

the 4-, 6- and 8-year-olds, with the highest score on the WL class, followed by the WS class, and with the worst performance on the strong verbs. However, there is a larger discrepancy with the results obtained for the adults in the experiments, since the network's performance on the WS class never supersedes that on the WL class, and since the slightly better score on strong verbs than on WL class verbs that was found for the adults is not reflected in the model's performance.

6.2. Errors I will now consider the kinds of errors that the model made during training. It should be pointed out that in the psycholinguistic experiments most of the errors were overgeneralisations. By overgeneralisation of a verb class I mean cases where the inflectional pattern of this class is applied to a verb belonging to a different class, e.g. when WL inflection is applied (overgeneralised) to a strong verb. CX/ergeneralisation of the strong dasses Overgeneralisation of the WSdass • Q/ergeneralisation of the WL dass

12

5210 S Φ ο

•δ 8 jje «ο

η Ο 4 -§-·

"ο

2 Ο

1 0 0 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 8 0 0 9 0 0 1000 1100 1200 1300 1400 1500 1600 1703 1ΘΟΟ

Vocabulary size Figure 9. Percentages of the network's total number of errors that were overgeneralisations of the different verb groups

In contrast to what was found in the psycholinguistic experiments, overgeneralisations never constituted more than 15% of the errors

Norwegian connectionism

181

made by the model. The main part of the errors consisted of an incorrect change of stress, tone or a stem phoneme, which are types of errors that human subjects are unlikely to make. Here I will nevertheless concentrate on the overgeneralisation errors, since these were the most common errors in the psycholinguistic experiments. Figure 9 shows the model's overgeneralisation rates on the different verb groups as proportions of the total number of errors. In other words, it shows how many of the errors resulted from inappropriate application of WL inflection, WS inflection or some strong inflection, respectively. Corresponding data from the psycholinguistic experiments are shown in Figure 10.

Overgeneralisation of the strong classes Overgeneralisation of the W S class Overgeneralisation of the WL class

Age 4

Age 6

Age8

Adults

Figure 10. Percentages of the human subjects' total number of errors that were overgeneralisations of the different verb groups. Adapted from Bjerkan and Simonsen (1996)

If we once again concentrate on what happens to the network after the 350 verb mark, we see that overgeneralisation of the strong classes and of the WL class stay relatively constant at a fairly low level throughout training. The low proportion of strong class overgeneralisation is consistent with the experimental results, although the model does not reflect the small and statistically insignificant increase of such errors found in the experiments. As for the rate of WL overgeneralisation, it matches the results for the adults, who

182 Anders N0klestad

made very few errors of this kind, but it does not reflect the high proportion of such errors among the children. This is probably due to the low type frequency of this class early in training; as was shown in Figure 3, the WL class does not actually become the largest class until the vocabulary reaches about 550 verbs. Why, then, did the children in the experiments overgeneralise this class to such a large extent? One reason, suggested by Ragnarsdottir, Simonsen, and Plunkett (1997, 1999), might be that verbs from the WL class are easier to segment than verbs from the WS class. This is because the suffix of the WL class verbs begins with a vowel, which probably makes it more salient than the suffixes of the WS class, since these suffixes begin with a consonant that is also often merged with the final consonant of the stem (cf. the past tense of levere in Table 2). The phonological representation used with the connectionist model, on the other hand, did not reflect differences in salience, nor did it incorporate any merging of consonants. Hence it did not reflect this difference in ease of segmentability, which may be at least part of the reason for the difference in the rate of WL overgeneralisation between the network and the children. I will now turn to the most striking feature in Figure 9, which is the development of overgeneralisation of the WS class. It shows a steady increase throughout the training period, so that at the end of training it constitutes two thirds of the overgeneralisation errors. This development is remarkable, since it corresponds to a clear, but unexpected, tendency found in the psycholinguistic experiments. Figure 10 shows that among the human subjects, overgeneralisation of this class increased dramatically with age. Considering the low productivity of this class compared to that of the WL class, this was a surprising experimental result, and it is therefore interesting to see that the model shows a similar increase in the proportion of WS overgeneralisation. In order to explain the low WS overgeneralisation with their youngest subjects, Ragnarsdottir, Simonsen, and Plunkett (1999) again refer to the lower phonological salience of the WS suffixes and to the fact that many WS verbs show stem-suffix assimilation in the past tense, making them harder to segment than the WL verbs. The

Norwegian connectionism

183

increase in overgeneralisation with age is attributed to the high token frequency of the WS class leading to more robust long-term learning and increasing levels of generalisation. While these factors might well be part of the explanation, they cannot explain the similar tendency found in the simulation results. As mentioned above, the model's way of representing word forms did not reflect differences in ease of segmentability. Moreover, although the WS class did have a relatively high token frequency throughout the training period, during the period when its overgeneralisation rate increased the proportion of WS tokens was lower than that of the WL class and decreasing (Figure 4). Ragnarsdottir, Simonsen, and Plunkett do, however, point to the differences in degree of phonological conditioning of the two weak classes (cf. Section 1) as an explanation for the slower rate of WS acquisition. They propose that it takes a longer period of time and exposure to a larger number of verbs for the language learners to figure out the phonological restrictions on the WS class than it takes to learn the relatively open schema of the WL class. As noted by Simonsen and Bjerkan (1998) and Simonsen (this volume), such a proposal could also be used for explaining the increase in WS overgeneralisation in the experiments, and this is an explanation that also applies to the connectionist simulations. If it takes some time to learn the more restricted WS schema, overgeneralisation of this class should emerge only gradually, but its greater degree of phonological explicitness could eventually lead to a greater force of attraction on verbs from other classes that match the schema than is the case for the relatively unrestricted WL schema. The model's increase in WS overgeneralisation is also relevant for the much debated issue of prerequisites for generalisation in connectionist models (cf. e.g. Plunkett and Marchman 1991, 1993; Marcus et al. 1995). The debate has focussed on the acquisition of default inflectional patterns like English past tense suffixation and German plural formation with the -s suffix, and default inflection in connectionist models has been claimed to depend on superior token frequency of the default class (Marcus et al. 1995) and/or on a situation where the verbs forms from the class that are presented to

184

Anders N0klestad

the model cover the entire "phonological space" of the language (Hare, Elman and Daugherty 1995). In Norwegian the most "defaultlike" class, i.e., the one with the fewest phonological restrictions (Bybee's 1995a definition of "default"), is the WL class, and this class also has the highest type and token frequencies throughout most of the training period. Hence it comes as no surprise that this is the class that the model most often generalises to novel verbs, as we will see in the next section. What the increase in WS overgeneralisation shows is that a non-default class, i.e., one with a number of phonological restrictions on the verb stems, can become the predominant source of overgeneralisations even when it has lower type and token frequencies than another class in the language.

6.3. Generalisation to novel verbs Finally, the network was tested on 200 novel verbs, i.e., verbs that were unknown to the model since they were not included in the training set. The network's treatment of these verbs is shown in Figure 11. (Note that the maximum value on the y axis is 40%; a considerable amount of the network's output did not qualify as a valid past tense form of any inflectional class). 40 ·_ 35

I 25 20··

•Generalisation of the strong classes • Generalisation of the WS class •Generalisation of the WL class

1 0 0 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 8 0 0 9 0 0 1000 1100 1200 1300 1400 1500 1600 1700 1800

Size of training vocabulary Figure 11. Network performance on novel verbs

Norwegian connectionism

185

The subjects in the psycholinguistic experiments were not tested on verbs they did not know, so the model's performance cannot be evaluated against experimental results in this case. However, it is worth noting that the network gradually develops generalisation rates for the different verb groups that correspond to the relative verb class productivity reported in section 1, with the WL class as the most productive one, followed by the WS class, and with the lowest productivity for the strong classes. This result is particularly interesting when we compare it to the overgeneralisation rates that were given in Figures 9 and 10, where it was shown that among errors of overgeneralisation to known verbs, it is the smaller weak class that predominates in the later stages of development, in the simulation as well as for the human subjects.

7. Summary and conclusion In this paper we have seen that it is possible for a connectionist network to learn to produce correct Norwegian past tense forms when it is given the corresponding present tense form, and that it does so without being given explicit information about inflectional patterns in Norwegian. Furthermore, the model's overall proportion of correct past tense forms shows a developmental profile that is similar to the one found in psycholinguistic experiments. Its relative performance on the different verb groups reflects the performance of the children, though not the adults, in the experiments, with the best performance on WL verbs, followed by WS verbs and with the worst performance on strong verbs. With respect to error types, the model makes far less overgeneralisation errors than the human subjects do. Nevertheless, in accordance with the experimental results it overgeneralises all of the verb groups to some extent. The model also reflects the fact that there are few overgeneralisations of strong verb inflection regardless of age, that the overgeneralisation rate for the WS class increases with age, and that the WL class is rarely overgeneralised by adults. What it does not reflect, however, is the high rate of WL overgeneralisation among the youngest children.

186 Anders N0klestad

Finally, most of the novel verbs for which the model produces a valid past tense form are inflected as WL verbs, fewer as WS verbs and only a very few as strong verbs, which shows that the model has developed an appropriate relationship between the degrees of productivity for the different verb groups. These results indicate that a single mechanism that generalises over its input is sufficient for explaining the relationship between the productivity of the verb classes as well as the human subjects' tendency, increasing with age, to overgeneralise the WS class. What the model does not explain is the high WL overgeneralisation rate among the youngest children. This could therefore be taken as an indication that a symbolic rule is needed for WL inflection in the early stages of past tense acquisition. Furthermore, if one finds the reduction of token frequencies used in the simulations unacceptable, it could be argued that such a rule is needed for WL processing later on as well, since pilot simulations showed that use of higher token frequencies prevented the model from showing such a high productivity for the WL class. However, the idea of symbolic WL inflection is seriously weakened by Ragnarsdottir, Simonsen, and Plunkett's (1999) finding that all of their subjects do significantly better on WL verbs with high token frequencies than on those with low token frequencies. In a dual mechanism model of the kind proposed by Marcus et al. (1992, 1995), Pinker and Prince (1998) and Prasada and Pinker (1993), high token frequencies are only needed in order to block application of the default rule. The rule itself, being the default, applies to any verb that is not associated with another inflectional pattern in memory, and is thus not dependent on the verb having a strong mental representation caused by a high token frequency. Ragnarsdottir, Simonsen, and Plunkett's results, then, seem to rule out the possibility that WL inflection is a symbolic default. A better explanation might be offered by the hypothesis outlined in section 6.2, which implies that the high rate of early WL overgeneralisation is due to phonological factors which make the WL class easier to segment than the WS class. Although connectionist models with more realistic input rep-

Norwegian connectionism

187

resentations are needed in order to explore this hypothesis, the present simulation results are at least compatible with it. Interestingly, the only class that did not show token frequency effects in Ragnarsdottir, Simonsen, and Plunkett's experiments was the WS class. However, if WS inflection is carried out by a symbolic rule, it cannot be a default rule of the kind discussed above, due to the phonological restrictions on the verb stems in this class. Furthermore, the present simulations show that both the limited productivity and the overgeneralisation rate of this class can be obtained without a symbolic rule. In sum, when combined with results from psycholinguistic experiments, the results from the connectionist simulations presented here do indeed seem to lend support to the idea that type frequency, token frequency and phonological form are important for the acquisition of both strong and weak past tense inflection in Norwegian.

Notes 1. I want to thank Hanne Gram Simonsen and Kim Plunkett for their guidance during the work on this project, and two anonymous reviewers for valuable comments on an earlier version of this paper. I also thank the Oceania group and the Research Committee at the Department of Linguistics, University of Oslo, for grants which enabled me to visit the Department of Experimental Psychology at Oxford University, where much of this work was carried out. 2. Exceptions are monosyllabic verbs and verbs with stems ending in -er, which are always inflected according to the pattern of the smaller weak class. 3. This list was provided by the Text Laboratory at the University of Oslo and the Norwegian Universities Documentation Project. The novels on which the list is based are now part of the OCTNT corpus, but the corpus itself was not yet available at the time when the simulations were run.

Regular inflection in the mental lexicon: Evidence from Italian Paola F. Matcovich

This paper addresses the competing predictions made by the dual mechanism model and the single mechanism model, concerning the base for productive language use. It reports some of the results from two psycholinguistic experiments, where past participle forms (existing verbs, as well as nonce forms) were elicited from Italian children and adults. The first experiment (adapted from Ragnarsdottir, Simonsen, and Plunkett 1999) was conducted on 130 children aged 4, 6 and 8 and consisted of 61 verbs from different regular and irregular inflectional classes, varying in type and token frequency and in phonological predictability of inflection. The second experiment elicited the past participle forms of 40 nonce items from 60 adult speakers. In my interpretation, the obtained results provide support for the single mechanism model, in that regular and irregular inflection are shown not to differ qualitatively either with regard to the effect of token frequency or with respect to phonological factors.

1. Introduction A much debated issue in the linguistic and psychological literature has been the question of how lexical and morphological features are stored and accessed in the brain. More specifically, the issue addressed has been whether regular and irregular forms should be treated as qualitatively different from each other, as claimed by the dual mechanism model (Marcus et al. 1992, 1995; Prasada and Pinker 1993), or whether they should be treated as the same kind of entities, as advanced by single mechanism models: connectionist models, Bybee and Langacker's network models (Bybee 1985, 1988,

190 Paola F. Matcovich

1995a; Langacker 1987a, 1990). Note that while several other theories address the issue of storage and processing,1 the findings in this paper are exclusively interpreted relative to the single/dual mechanism controversy, this having been hotly debated in the last decade, and hence the discussion refers to these only. The core of the dual mechanism model in its most recent version is the claim that a rule-based symbolic component takes care of the representation and generation of regular forms, while an associative rote-memory system deals with irregular forms. The single mechanism approach maintains that all forms, regardless of whether they are regular or irregular, are processed by one and the same mechanism. By virtue of this fundamental difference between the two approaches, the two models make quite different predictions concerning the role played by factors such as the phonological properties of regular forms, as well as their type and token frequency, in establishing and maintaining lexical representation. Whereas in the dual mechanism model neither token frequency effects nor the phonological properties of word forms are predicted to affect regular inflection (this being rule governed), in the single mechanism model such factors are thought of as playing a pivotal role both within regular and irregular inflection. In this paper I will report some of the results concerning these competing predictions from two psycholinguistic experiments I have conducted, eliciting past participle forms (existing verbs, as well as nonce forms) from Italian children and adults.

2. Past participle inflection in Italian Before moving on to the experiments a very brief description of past participle inflection in Italian is in order. Most textbooks of Italian divide verbs into three basic conjugation classes, identified by the thematic vowel in the verb stem: Conjugation I with the thematic vowel -a (as in parlare 'to talk'), Conjugation II with -e (as in credere 'to believe'), and Conjugation III with -i (as in dormire 'to sleep'). These are exemplified in Figure

Italian regular inflection

191

1, Both historically and currently Conjugation I is the most numerous and productive inflectional class, with over 3000 members. Conjugation II and III are much smaller classes, with approximately 400 and 500 members respectively. Conjugation II is no longer productive; historically it is the most unstable class, with verbs migrating to the other classes. Conjugation III is still partially productive, hosting some deadjectival verbs, especially those derived from adjectives expressing colours, such as the existing verb ingiallire 'to yellow' or the neologism ingrigire 'to become grey'. While Conjugation I differentiates itself from the other two classes in terms of type frequency, Conjugation II differs from Conjugation I and III in terms of distribution of regularity, with irregular forms being mostly Conjugation II verbs.

Infinitive Past participle

Conjugation I

Conjugation II

Conjugation III

/par'lare/ /par'ato/

/'credere/ /cre'duto/

/dor'mire/ /dor'mito/

Figure 1. The past participle of regular verbs

Regular Conjugation I and III verbs form their past participle with the suffix -to attached to the stem ending in the thematic vowel, with stress falling on the thematic vowel. The same applies to regular Conjugation II verbs, with the only difference that the vowel -« is used instead of the thematic vowel -e. Examples are given in Figure 1, where the verbs parlare 'to speak', credere 'to believe' and dormire 'to sleep' are used. Conjugation II

Conjugation III

Infinitive piangere, prendere morire, apparire Past participle pianto, preso morto, apparso Figure 2. The past participle of irregular verbs with the suffixes -(t)to and -(s)so

In the case of irregular verbs, (with the exception of twelve Conjugation II verbs ending in -sistere in the infinitive and forming a little subclass of their own), the past participle is formed with the suffixes -(t)to or -(s)so attached to a stressed stem which has

192

Paola F. Matcovich

undergone the loss of the thematic vowel. This is illustrated in Figure 2 and applies both to Conjugation II irregular verbs such as piangere-pianto 'to cry-cried' and prendere-preso 'to take-taken', and to Conjugation III irregular verbs, such as morire-morto 'to diedied' and apparire-apparso 'to appear-appeared'. Further, the past participle form of nearly all irregular -ere verbs involves a consonant elision or modification of the verb stem, and in some cases also a vowel change. This elision or modification of one (or two) stem final consonants is often governed by phonological factors such as phonotactic and assimilation processes, though not always. Finally, the phonological properties of a given irregular verb often offer a cue as to what subclass the verb belongs to, and hence are only partially predictive of its past participle suffix. Italian past participle inflection, then, opens for two alternative interpretations of the dual mechanism model according to how the different types of past participle formation are to be distributed between the two mechanisms. Under the first interpretation, (which for simplicity reasons I will refer to as the "three-rule interpretation"), the rule component which deals with regular inflection is responsible for attaching the -ato, -uto, and -ito suffixes to regular are, -ere and -ire verbs respectively. Under the second interpretation, (which I will refer to as the "one-rule interpretation"), the symbolic component consists of a single rule (just like in English), concatenating the -ato suffix to regular Conjugation I verbs. Both these interpretations have been advanced in the literature.2

3. The experiments

3.1. Experiment! The first experiment, which I will refer to as experiment 1, consisted of an oral sentence completion task eliciting past participle forms of Italian verbs. The task resembled the pattern designed by Bybee and

Italian regular inflection

193

Slobin (1982) and reelaborated by Ragnarsdottir, Simonsen, and Plunkett (1999) for Norwegian and Icelandic children. The experimenter presented the child with the picture of a person engaged in a particular activity and described the situation using a sentence which contained the present tense of the verb depicted in the picture. She formed a question referring to the past, in order to elicit the past tense form of the same verb.3 Three groups of children aged 4, 6 and 8 were tested, with 40 children in the 4- and 8-year-old groups and 50 in the group of the 6year-olds, for a total of 130 children (as indicated in Figure 3). All subjects used the dialect spoken in the Trieste region, and so did the parents of the children in the experiments.4 Age

Males

Females

Tot. # of children

17 25 21 1

40 50 40

4-year-olds 23 6-year-olds 25 8-year-olds 19 Figure 3. Distribution of subjects in experiment

Each group of children included approximately the same number of males and females. Each child was tested on 61 verbs; 15 were regular -are verbs, 14 were regular -ere verbs and the remaining 32 were irregular -ere verbs belonging to different subclasses. Conjugation III verbs were excluded since all the variables we were interested in manipulating (briefly described in the next paragraph) could be tested on Conjugation I and II verbs alone. The parameters taken into account in choosing the verbs and verb classes to include in the test were essentially the factors on which the models briefly described earlier seemed to make competing predictions: the inflectional classes had to differ in terms of type frequency and phonological coherence, whereas the single verbs had to vary in token frequency and predictability of inflection, in order to test the role of these factors in acquisition of past participle inflection.

194

Paola F. Matcovich

Here, however, we will be concerned only with the token frequency parameter. A few methodological details are in order. In order to investigate the role of token frequency in acquisition, each verb class was further subdivided into two groups (a and b), containing verbs with high and low token frequency respectively and approximately the same number of verbs. Token frequency values for each of the verbs included in the test were drawn from De Mauro et al. (1993) Lessico di frequenza dell 'italiano parlato ("Frequency dictionary of spoken Italian"). The corpus in this dictionary consists of recordings of spoken Italian (half a million word tokens, over 15 thousand lemmas), taken from five different contexts of use. The values High and Low were assigned to each verb according to its token frequency (LoToFrVerb < 50 < HiToFrVerb, where ToFr is the sum total of the five contexts for each verb as in De Mauro et al. 1993). Extremely high token frequency verbs were excluded, as they were not thought of as being interesting test items due to their disproportionate frequency of use, as well as very low token frequency verbs, since the meaning of the verbs that were ultimately included in the test had to be familiar to the children. Table 1. Correct PP answers for all verbs by verb class and token frequency (experiment 1)

HiToFr LoToFr H-L

Regular -are verbs (p.p. -οίο)

Regular-ere verbs (p.p. -uto)

Irregular-ere verbs (p.p. -so)

Irregular-«re verbs (p.p. -to)

99% 87% +12%

96% 83% +13%

84% 50%

82% 58%

+34%

+24%

The results from experiment 1 showed that token frequency had a strong effect on children's ability to inflect regular (as well as irregular) verbs in the task (F,(l,124) = 783.59, ρ < .0001; F2(l,50) = 13.02, ρ = .001).5 This effect was found to be significant (Tukey-b post hoc analyses with ρ < 0.05) in both regular inflectional classes that were tested (results are given in Table 1) and for all age groups (Table 2).

Italian regular inflection

195

Table 2. Correct PP answers for all verbs by age group and token frequency (experiment 1)

Age 4 Age 6 Age 8

HiToFr

LoToFr

84% 92% 97%

51% 69% 83%

Before commenting upon these results, let us turn to experiment 2.

3.2. Experiment 2 Experiment 2 consisted of an oral nonce task eliciting the past participle forms of 40 nonce items from 60 adult speakers of Italian. The aim of the task was to test for effects of phonological similarity within Conjugation I regular inflection. The reason why Conjugation I was chosen, is the fact that if effects of phonological similarity do matter and thus are found to be statistically significant within this class, then, neither the "three-rule interpretation" of the dual mechanism model (according to which the rule component is responsible for -ato, -uto and -ito suffixation) nor the "one-rule interpretation" (according to which the rule component is responsible only for -ato suffixation) can be said to draw support from the data. In designing such a test a major problem was that, due to the high type frequency and very open phonological specification of the class of -are verbs, both a nonce verb modelled on existing -are verbs and a nonce verb devised as having no similarity to any existing -are verb were likely to be inflected in -ato. This is very similar to the situation we have in English, with the default past tense suffix -ed. A way of solving this problem, then, was to model the nonce items simultaneously on one or several existing -are verbs and on a group of -ere verbs. This was possible thanks to the wide range of patterns found within Conjugation I verbs. If the input provided to the subjects consists of a form that obliterates all differences between the three conjugations (for instance the first or the second person singular of the present indicative), then the verb would be susceptible

196

Paola F. Matcovich

to be inflected in -ato, but also to take the suffix which is characteristic for the group of -ere verbs that it has been modelled on. For example, if the input form were to be a nonce verb like appogli (2nd person singular present indicative), designed as similar to both regular Conjugation I verbs like spogli 'you undress', sfogli 'you leaf and irregular Conjugation II verbs such as togli 'you take out', raccogli 'you pick up', the subjects have another plausible alternative to inflecting the verb as appogliato (as in spogliato and sfogliato), namely appolto (as in tolto and raccoho). Similarly, the nonce item affargi, which does not rhyme with any existing -are verb, but which is similar to -ere verbs like spargi 'you spread', is susceptible to being inflected both as affargiato (due to Conjugation I's very high type frequency and extremely open phonological specifications) and as affarso (cf. sparso). This is illustrated in Figure 4. Nonce item

Rhyming -are verbs

Rhyming -«re verbs

appogli

spogli (inf. spogliare, p.p. spogliato) s f°gH (inf. sfogliare, p.p. sfogliato)

togli (inf. togliere, p.p. tolto) raccogli (inf. raccogliere, p.p. raccolto) appolto

appogliato

spargi (inf. spargiere, p.p. sparso) affargi affargiato

affarso

Figure 4. Example of two nonce items (experiment 2)

Further, two conditions were created: in Condition 1 nonce verbs were devised that were similar to existing regular Conjugation I verbs, whereas Condition 2 comprised novel verbs which did not rhyme with any existing -are verb. Moreover, both Condition 1 and Condition 2 verbs, had in turn to be similar to existing irregular Conjugation II verbs. This is summarised in Figure 5.6

Italian regular inflection Test cond.

197

Description Each nonce item was modelled simultaneously on a group of existing -are verbs and on a family of existing -ere verbs. Ex.: Condition 1 item appogli is similar to -are verbs like spogli 'you undress' (as well as to -ere verbs like togli 'you take out'). Each nonce item was devised so as not to rhyme with any existing -are verb, and at the same time was modelled on the same group of existing -ere verbs as its pair counterpart. Ex.: Condition 2 item affargi does not rhyme with any existing -are verb (but is similar to -ere verbs like spargi 'you spread').

Figure 5. Description of Conditions 1 and 2 (experiment 2)

In order to control for possible token frequency effects, the nonce items were devised in pairs, with one item from each of the two conditions outlined above, so that each pair was modelled on one and the same group of irregular -ere verbs. To achieve this, the test exploited a certain property of Italian verb inflection, namely the fact that a number of Conjugation II verbs present a systematic alternation in the verb stem. The results are laid out in greater detail in Matcovich (1998). For our purposes it is enough to remember that in each nonce pair (ex. appogli/appolgo), both items are modelled on the same group of -ere verbs, namely that of verbs such as togli-tolgo ('you take out-I take out'). However, the Condition 1 item appogli rhymes also with existing -are verbs such as spogli 'you undress', while the Condition 2 item appolgo does not rhyme with any existing -are verb. This is illustrated in Figure 6. Nonce pair

Rhyming -are verbs

Rhyming -ere verbs

appogli

spogli (inf. spogliare, p.p. spogliato) sfogli (inf. sfogliare, p.p. sfogliato)

togli (inf. iogliere, p.p. tolto) raccogli (inf. raccogliere, p.p. raccolto)

appolgo

no verb ending in -olgare

tolgo (inf. togliere, p.p. tolto) raccolgo (inf. raccogliere, p.p. raccolto)

Figure 6. Example of a nonce pair (Experiment 2)

198

Paola F. Matcovich

Finally, two versions of the test were created (A and B), each containing one item from all the twenty pairs, so that each test included ten Condition 1 and ten Condition 2 items. The two tests were mirror images of each other, in that for every Condition 1 item included in test A the Condition 2 item of the same pair was assigned to test B and vice versa. Attention was paid to have roughly the same number of items in the first person singular as in the second singular in both tests.

4. Results and discussion Condition 1 items (rhyming with existing Conjugation I verbs) were found to be inflected regularly (-ato suffix) significantly more often than Condition 2 items (which did not rhyme with any existing -are verb) (F(l,36) = 36.916, p < .0001).7 This difference in regular response rate between Condition 1 and 2 remained significant also for (1) all the four alternations employed in the experiment,8 (2) seven of the eight rhymes,9 (3) over three fourths of the nonce pairs. Table 3. Percentage of regular Conjugation I (p.p. -ato) responses as a function of test condition. Test Condition

Gen>ato

Gen>irr'°

1 2

59% 31%

41% 69%

As briefly mentioned earlier, in the dual mechanism model token frequency and phonological similarity effects are predicted to affect children's correct performance only on irregular verbs. This is due to the fact that regular forms are believed to be computed by rule, and rules are insensitive both to token frequency and phonological similarity effects. Conversely, in all single mechanism models, token frequency and phonological similarity effects are expected to be significant for both irregular and regular verbs, since no structural distinction between the two types of inflections is postulated by this approach.

Italian regular inflection

199

Now, the fact that both token frequency (experiment 1) and phonological similarity effects (experiment 2) were found to be statistically significant for regular inflection, seems to be perfectly compatible with the single mechanism account, in that regular and irregular inflection are shown not to differ qualitatively either with regard to the effect of token frequency or with respect to phonological factors. Conversely, these results do not seem to lend any support to the dual mechanism model. Finally, note that this applies not only to the "three-rule-interpretation" of the dual mechanism model, but also to the "one-ruleinterpretation", since Conjugation I (which is handled by the rule component under both interpretations) showed statistically significant effects of both token frequency and phonological similarity.

Notes 1. Cf. for instance the Analogy Model of Royal Skousen, (Skousen 1989; Derwing and Skousen 1994), and psycholinguistic models such as the Augmented Addressed Morphology Model developed by Burani and Caramazza (1987), Caramazza, Laudanna, and Romani (1988), Burani and Laudanna (1992) and the Morphological Race Model formulated by Frauenfelder and Schreuder (1992). 2. For the "three-rule interpretation" see for instance Orsolini and Marslen-Wilson (1997). For the "one-rule interpretation" see Say (1997). 3. The elicited form was passato prossimo, which is a compound tense made up by an auxiliary verb and the past participle of the main verb. 4. The past participle inflection in the variety of Italian spoken in the Trieste area corresponds (with some entirely predictable phonological mappings) to the standard norm. 5. Correct response proportions were entered into a multivariate analysis of variance on subjects and items, with Age (4, 6 and 8 years), Token Frequency (High and Low), Input Form (first and third person singular of the present indicative) and Verb Class (regular -are verbs p.p. -ato, regular -ere verbs p.p. -uto, irregular -ere verbs p.p. -to and irregular -ere verbs p.p. -so) as factors. The results reported here concern the main effect of Token Frequency.

200

Paola F. Matcovich

6. A third condition consisting of rhymes between nonce items and Conjugation I verbs only was not included since it was not believed to result in added significance, all items of such a condition being extremely likely to be inflected in -ato. 7. For each nonce verb the number of regular (p.p. -ato) responses was calculated. These results were entered into a multivariate analysis of variance with Test Condition (1 and 2) and Test Version (A and B) as factors. The results reported here concern the main effect of Test Condition. 8. The four alternations in question were /lg-X/, /k-tj7, /g-dy and /rjg-ji/. Paired samples t-test revealed that regular Conjugation I responses (p.p. -ato) differed significantly between the two conditions for all alternations (df = 59; /g-dy: t = 9.17, p < .0001, /k-tj/: t = 3.77, p < .0001, /lg-/C/: t = 6.12, p < .0001, /rjg-ji/: t = 2.45, p < .017). 9. The eight rhymes employed in the test were the following: Rhyme 1: -ar-g/dj-are, as in the nonce pair affargo-qffargi Rhyme 2: -an-g/dj-are, as in the nonce pair miango-miangi Rhyme 3: -un-g/dj-are, as in the nonce pair pillungo-pillungi Rhyme 4: -in-k/tf-are, as in the nonce pair sfinco—sfinci Rhyme 5: -or-k/ff-are, as in the nonce pair allorco-allorci Rhyme 6: -e-lg/rf-are, as in the nonce pair sfelgo-sfegli Rhyme 7: -o-lg/a-are, as in the nonce pair infolgo-infogli Rhyme 8: -e-rjg/ji-are, as in the nonce pair sfengo—sfegni. Paired samples t-test revealed that regular Conjugation I responses (p.p. -ato) differed significantly between the two conditions for seven of the eight rhymes: df = 59, rhyme 1: t = 5.42, p < .0001; rhyme 2: t = 3.16, p = .002; rhyme 3: t = 5.66, p < .0001; rhyme 5: t = 7.29, p < .0001; rhyme 6: t = 5.60, p < .0001; rhyme 7: t = 3.15, p = .003; rhyme 8: t = 2.45, p = .017. 10. Those are non-Conjugation-I-responses, that is past participle responses other than -ato. Of these, approximately 70% are irregular Conjugation II responses (p.p. in -so οτ-to), 18% are regular Conjugation II responses (p.p. in -uto), and 12% are regular Conjugation III responses (p.p. in -ito).

Section III Constructional aspects of the verb

Malefactive by means of GIVE Öle Torf inn Fagerli

The present study is concerned with two common ways of expressing benefactive and malefactive in some selected languages. Serializing languages—which include among others the Kwa and Gur groups of West-Africa, languages of South-East Asia, and some pidgin and Creole languages—very often use a word identical to the verb GIVE juxtaposed to a host verb in such circumstances, while typical extensional languages, like for instance West-Atlantic and Bantu languages, use benefactive or applicative affixes to modify the verbal phrase. Most of what can be expressed by verbal extensions in some languages can be directly transferred to serializing languages by means of serial GIVE, and vice versa. Focus is directed to the fact that a malefactive event may be linguistically marked by an item corresponding to the verb GIVE, or by an extension which is typically used to mark benefactive events. The apparent paradox is accounted for on the basis of the semantics of the host verb (and its complement's), if any), together with semantic bleaching of the original benefactive marker. Malefactive GIVE is sometimes merely an affection marker, the malefactive content being carried by the host verb.

1. Introduction The present study is concerned with two common ways of expressing benefactive and malefactive in some selected languages. By the term malefactive we understand the linguistic coding of an event describing that something is done to the detriment of somebody or some situation is bad or hurtful for somebody. Adversative is another term applied to cover the same meaning. Malefactive coding is often linked to benefactive in a formal way, by the fact that the two often share morphophonological patterns. Serializing languages, which

204

Öle T. Fagerli

include among others the Kwa and Gur groups of West-Africa; languages of South-East Asia; and some pidgin and creole languages, very often use a word identical to the verb GIVE1 juxtaposed to a host verb in such circumstances, while typical extensional languages, like for instance Atlantic and Bantu languages, use benefactive or applicative affixes to extend the verbal phrase. Most of what can be expressed by benefactive extensions in some languages can be directly transferred to serializing languages by means of serial GIVE, and vice versa. Focus in this paper is directed to the fact that a malefactive event may be linguistically marked by an item corresponding to the verb GIVE, or by a verbal extension which is otherwise used to mark benefactive events. The apparent paradox is accounted for on the basis of the semantics of the host verb (and its complement, if any), together with extension2 of sense of the original benefactive marker. Malefactive GIVE is sometimes merely an affectedness marker, the malefactive content being carried by the host verb or its complement. The presentation and analysis of data is accompanied by an attempt at approaching the phenomenon within the framework of Cognitive Grammar. Recent works by Cognitive Linguists, most importantly Newman (1996, 1997), confirm the validity of the results of the present study. At the same time, it is believed that this presentation (slightly) enriches the already wellelaborated typological description of the GIVE verb.

2. The prototypical GIVE scene The meaning of GIVE is complex. I hold the literal, concrete sense of GIVE as basic and as a point of departure for related meanings. This basic predicate can be viewed schematically as in Figure 1. There are three participants, a GIVER, a THING, and a RECIPIENT. The GIVER hands the THING over to the RECIPIENT. In more detail, the transfer typically involves physical contact with the THING, first in the hands of the GIVER, then in the hands of the RECIPIENT. Consequently change of control or possession is often associated with the act. In the terms of Langacker (1991: 292), giving involves

Malefactive by means of GIVE

205

an energy flow, initiated by the benevolent GIVER (the "energy source") and ending up with the RECIPIENT (the "energy sink") who normally accepts the THING. This is referred to as the force-dynamics of the verb. The energy consists of mental activity on the part of both the GIVER and the RECIPIENT, and physical movement of all three participants.

TIME Figure 1. A basic representation of the GIVE scene

The scenario sketched above may be viewed from different angles, some parts of it may be focused while other parts may be omitted in appropriate contexts. Some parts, the THING participant for instance, may be extended metaphorically to denote abstract entities. Or the movement to a recipient may be focused/extracted and used to describe movement along a path to a GOAL, leaving other subparts of the meaning of GIVE aside. A rather complex verbal meaning like GIVE capturing such a basic human experience has the potential of being extended to other semantic fields than just the literal GIVE scene. There are indeed various ways of both extracting subparts of the literal meaning and of extending the meaning figuratively in order to apply the verb to "new" contexts. Cognitive Grammar emphasizes the importance of discovering relatedness of meanings and of establishing and explaining (wherever possible) polysemous patterns. In the sections to follow, GIVE morphemes in complex predicates of different kinds are used to express among other things benefactive and malefactive.

206

le T. Fagerli

3. Benefactive by means of verbal extensions In Fula, an Atlantic language spoken from Senegal to Sudan, we find a range of verbal suffixes, extending or modifying the meaning of the verb in various ways, and being linearly ordered between the verbal radical and its obligatory inflectional marker. I will concentrate on Fula data in the following two sections, in particular the Aadamaawa Fula dialect (Cameroon and Nigeria). For an overview and details about verbal extensions in Fula, see Arnott (1970), Fagerli (1994), and Breedveld (1996). Benefactive, malefactive, and allative is marked by the same verbal extension -AN-.3 Benefactive here covers both the prototypical case in which we find a GIVER and a RECIPIENT of a concrete THING, and cases in which there is a benefactor and a beneficiary of an action, often referred to as a dative event. The difference between the two types is mainly one of metaphorical extension of the prototypical act of giving into "giving" of an action. These two main types of benefactive, the recipient-of-thing type and the beneficiary-of-action type, are found in the sentences below. In (1) and (2), it is possible to interpret the RECIPIENT as either recipient of thing or beneficiary of action. In (3) and (4) we find examples of the latter: (1)

Mi winnd-an-i Yaaya pataakeewol. I write-BEN-PERF Yaaya letter (i) Ί wrote a letter to Yaaya.' (ii) Ί wrote a letter for Yaaya.'

(Fula)

(2)

Abbo te6-an-i Didi maygoro. Abbo pick-BEN-PERF Didi mango (i) 'Abbo picked a mango to Didi.' (ii) 'Abbo picked a mango for Didi.'

(Fula)

F.W. Taylor (1932) reports the same for e.g. sorr-an- 'to sell-BEN', which may mean 'to sell to' or 'to sell for'.

Malefactive by means of GIVE

207

(3)

Mi huuw-an-i mo. I work-BEN-PERF him Ί worked for him.'

(Fula)

(4)

Yaya wacf-an-i yam jahaarjgal rjgal. Yaya make-BEN-PERF me journey this 'Yaya traveled on behalf of me.'

(Fula)

A more subtle classification of benefactive events in Fula is available in Fagerli (1995).

4. Malefactive by means of verbal extensions Constructions with the -AN- extension in Fula may have malefactive interpretations in suitable circumstances. Three characteristic types of malefactive events are easily observable. First, in an act of giving the transferred element may be regarded as negative for the RECIPIENT. While picking a ripe mango or a flower to somebody is generally seen as benefactive to the RECIPIENT, picking up a rotten mango or garbage to give to somebody is generally malefactive to the RECIPIENT, cf (5) below. Second, the host verb of the extension may denote an action or state that is generally regarded as bad or negative, like LIE for instance, cf. (6). Third, the host verb of -ANmay have ablative characteristics, i.e indicate motion away from the patient, cf (7) and (8). In the second and third cases, the patient participant is affected by a negatively loaded action. Thus we may speak of two maleficiary-of-action types of malefactive events, along with the noted maleficiary RECIPIENT-of-thing type. (5)

Abbo ted-an-i Didi maygoro nyolnde. Abbo pick-MAL-PERF Didi mango rotten 'Abbo picked a rotten mango to Didi.'

(Fula)

Note however that this sentence also may be interpreted in the beneficiary-of-action way, meaning that Abbo picked a rotten mango

208

Öle T. Fagerli

on Didi's behalf. (6) below is ambiguous. There are two possible maleficiary-of-action readings of the sentence: (6)

Buuba few-an-i Jekariaw haala mbeewa. (Fula) Buuba lie-MAL-PERF Jekariaw affair goat (i) 'Buuba lied to Jekariaw concerning the goat.' (ii) 'Buuba lied about Jekariaw concerning the goat.'

If the situational background is that Buuba has stolen a goat, the first interpretation (6i) is presumably that Buuba lies when he tells Jekariaw that he, Buuba, is not the thief. In the second interpretation (6ii) Buuba blames Jekariaw for the theft. In both cases Jekariaw is a maleficiary of the action, though in the second case he may not be aware of being a RECIPIENT, if at all he is. The semantics of the sentence is a bit tricky. The RECIPIENT of the LYING message might be somebody else, while the maleficiary no doubt must be the indirect object referent (Jekariaw). A potential third reading of the sentence, meaning that Buuba lied for Jekariaw, is rejected by language informants, and is at best a highly marked interpretation. In (7) and (8) the movement described by the verbs hoo'- 'TAKE' andja6-t- 'SNATCH/TAKE' explain the malefactive reading. (7)

Aaman hoo'-an-i deerdiiko mbeewa. Aaman TAKE-MAL-PERF brother-his goat 'Aaman took a goat from his brother.'

(Fula)

(8)

O ja6-t-an-ilaw-i yam defte. he get-REV-MAL-CEL-PERF me books 'He quickly snatched my books from me.' (REV = REVERSIVE; CEL = CELERATIVE)

(Fula)

The THING (transferred element) moves from the TAKEE to the TAKER, in other words in the opposite direction of the THING in an act of GIVING. The verbal extension only indicates to whom the action is directed, in other words it functions as a kind of affectedness marker in these sentences. Senses of concrete recipiency

Malefactive by means of GIVE

209

and benevolence, typically associated with GIVE, are suppressed. The effect of the adversative act is transferred to the RECIPIENT (the affected maleficiary). Thus there is movement along two lines, one THING (the goat) moves from TAKEE to TAKER, while the other THING (the adversative act) metaphorically moves from TAKER to TAKEE. The subject of the sentence is both TAKER and GIVER. The force-dynamics of complex predicates like these is far more complicated than in "simple" GIVE predicates. But it seems clear that the benefactive morpheme functions as a carrier of energy flow all the way to the energy sink, the negatively affected RECIPIENT. We get an adversative reading with other extended verbs as well, where the host verb follows the TAKE pattern, e.g. wujj-an- 'steal from' (rather than 'steal for'). It would be reasonable however, in the case of polysemous structures in the individual languages, to bear in mind the domain of human interest associated with the ablative verbs in question. BUY has got the same characteristics as TAKE with respect to direction of movement, but there is normally no adversative effect on the person with whom one trades. This would be the case in Aadamaawa Fula, where the extended verb stem sood-anpreferrably is understood as 'buy for', hence a benefactive reading. For a special case of BUY in serializing languages, see example (24).

5. Allative In addition to benefactive and malefactive, the -AN- extension also indicates movement towards a GOAL. The allative function is sometimes referred to as goal. (9)

Pallaandi dogg-an-i legged. lizard run-ALLATIVE-PERF tree 'The lizard ran towards the tree.'

(10) Nagge umm-an-ake cow rise-ALLATIVE-PERF The cow started up at him.'

mo. him

(Fula)

(Fula; Arnott 1970)

210

Öle T. Fagerli

Noye (1989) uses the term Destinative as a cover term for benefactive, malefactive, and allative. Though it is not commonly named so in the linguistic literature, it seems like a suitable notation on semantic grounds: There is movement of some element, concrete or abstract, to a certain destination in all three cases. The polysemy of the extension is better accounted for by the term destinative than for instance applicative, which is a rather meaningless term. Applicative is commonly used to denote a common verbal extension in Bantu languages. These languages are typologically close to Fula with respect to verbal extensions. Swahili may express benefactive, malefactive and allative in a similar fashion to the one described above for Fula. However, I will not undertake a thorough comparative analysis here, the Swahili applicative extension being extremely polysemous. In addition to the possibilities mentioned, it serves to express among other things manner, purpose, and instrument. In Fula, these latter three form parts of an independent verbal extension.

6. Benefactive by means of serial GIVE Typologically very different from extensional languages, serializing languages commonly apply the verb for GIVE or alternatively SHOW to express not only recipiency-of-thing but also beneficiary-of-action. Such serial verb constructions consist of two verbs most commonly sharing the same subject and direct object, the GIVE verb, which is the second one in the string, introducing the RECIPIENT or beneficiary role into the complex predicate. The verb is rendered in English as the prepositions for or to. These languages normally have no productive prepositions corresponding directly to the English ones. The serial constructions below are complex predicates comparable to the Fula verbal extensions, being able to express both the recipiency-of-thing (11 and 12i) and the beneficiary-of-action (12ii15) types of benefactive. Introducing several serial languages at this point shows that the same benefactive construction type is common throughout Africa and further. Whether all of these languages posit

Malefactive by means of GIVE

211

equal polysemous structures for the the GIVE verb remains an object of investigation. It is not necessarily the case. (11) Kofi to-o boll no ma-α Ama. Kofi throw-Pst ball Def GIVE-Pst Ama 'Kofi threw the ball to Ama.'

(Akan; Saether0 1997: 19)

(12) Megbe sefofo ac[e na wo. I-pluck flower indef. GIVE you (i) Ί picked a flower to you.' (ii) Ί picked a flower for you.'

(Ewe; Essegbey: p.c.)

(13) Kofi og na Paramaribo gi mi. Kofi go to? Paramaribo GIVE me 'Kofi went to Paramaribo for me.'

(Sranan; Jansen et al. 1978 in Holm 1988)

(14) Mi wroko gi en. I worked GIVE him Ί worked for him.'

(Sranan; Voorhoeve 1975 in Lord 1993)

(15) Ali tono koro la a naa. Ali work+imp GlVE+imp a.m. a.def. chief 'Ali works for the chief.'

(Dagaare; Bodomo: p.c.)

The item expressing the benefactive meaning in Fula is a productive verbal extension, while the Akan, Ewe, Sranan, and Dagaare items expressing benefactive are words which are identical to independent GIVE verbs in the respective languages. The benefactive marker in Ewe, see (12), although glossed as 'for' and reported to be morphologically reduced in some serialization contexts (Eli Saether0 Andenes: p.c.), corresponds phonologically to the GIVE verb. All the mentioned GIVE verbs are elsewhere in the syntax used as full-fledged independent verbs. Examples (11) and (12i) are instances of complex predicates where the GIVER and the THING are identical to the subject and direct object referents of the host verb. GIVE introduces a new participant

212

Öle T. Fagerli

on the scene, namely the RECIPIENT, the indirect object referent to whom the THING is being transferred concretely. The only detail differing from prototypical literal GIVE in the two sentences is the lack of physical contact between the GIVER and the RECIPIENT in (11), due to the nature of the host verb THROW. The second reading of (12) would normally involve concrete recipiency, but it need not. It could be that the RECIPIENT is simply a beneficiary of the action, and that the THING is being transferred to somebody else. A more accurate translation into English would then be picked a flower on your behalf. Examples (13) to (15) are instances where the RECIPIENT is clearly a beneficiary of the action, no other THING than the action of the GIVER affecting the RECIPIENT. It is a controversial issue whether one should call the instances of GIVE above prepositions or case markers rather than verbs because their use has been extended to include the beneficiary-of-action type of benefactive not typically associated with the verb. From the point of view of many European languages, they also occupy positions normally associated with prepositions or case markers. It is ascertained that GIVE has developed into uninflected benefactive markers in some languages, for instance the Senufo language Karaboro (Lord 1993: 42). I believe serial GIVE as in the quoted examples are good candidates for being grammaticalized, but they still retain too much of their verbal morphophonological properties in these languages to deserve other category labels than verbs. For a discussion of category status of serial verbs, cf. Lord (1993) and Osam (1994). The linguistic material available presenting GIVE in serializing contexts is relatively rich, I have presented a selection of data only. Languages such as Mandarin and Hokkien Chinese, Thai, Vietnamese, Yao Samsao, Chamorro, Siya, and Yoruba also posit the same potential: GIVE can be used as a benefactive marker, cf. Lord (1993: 43) and Newman (1996: 212-223).

Malefactive by means of GIVE

213

7. Benefactive by means of SHOW GIVE is not the only possible verbal source for benefactive marking; among others SHOW is used in serializing and compounding languages to denote benefactive events. In Akan (Twi) it is possible to use a verb corresponding to SHOW in such circumstances: (16) wo-tow tuo kyere borohene. (Twi; Christaller 1881 they-fire gun SHOW governor in Lord 1993) They fire guns in honor of the governor.' (17) ama kd-d asem kyere-ε kofi. Ama say-PST matter SHOW-PST Kofi 'Ama told Kofi something.'

(Akan; Saether0: p.c.)

According to Eli Sasther0 (p.c.) (17) represents a standard way of expressing TELL in Akan. Schematically, SHOW exhibits the same type of participants as well as a similar flow of action as GIVE. However, I will not pursue the SHOW case further in this study.

8. Malefactive by means of serial GIVE Using GIVE to express malefactive seems like a contradiction. It is certainly a marked interpretation of sentences like (18) below, where the beneficiary-of-action type is preferred. However, just like in the case of verbal extensions in Fula, serial GIVE can serve as a malefactive marker, as in (19): (18) meh gbecfuc[o na wo (Ewe; Essegbey: I-collect rubbish GIVE you p.c.) (i) Ί collected rubbish for you.' (='on your behalf) (ii) '?*! collected rubbish to you.' (marked interpretation)

214

OleT.Fagerli

(19) Ο ngma la zirii ko Amai oi yideme yele. he cut a.m. lies GIVE Ama she housepeople matter 'He lied to Ama about her family.' (Dagaare; Bodomo: p.c.) (19) shows that the patient is negatively affected by the lies sent out by the malefactor. It is structurally and semantically similar to the examples in (11) and (12), and the use of GIVE is perhaps not surprising, although unusual. (20) is different and more surprising in that the primary verb indicates motion of the THING away from the male-ficiary, seemingly contradictory to the direction expressed by GIVE: (20) Kofi kibri wan sani gi Gado. Kofi hide one something GIVE God 'Kofi is hiding something from God.'

(Sranan; in Boretzky 1983)

Hiding something in this case means to keep it out of reach or out of sight of someone who normally has access to it. I believe it is right to apply the general term ablative for this relation, but more specifically the situation of the patient (God) is malefactive. This is so in spite of the use of GIVE as a second verb. The construction is very similar to the ablative examples in (22-23). As previously seen, GIVE implies movement of some entity from the benefactor to the beneficiary/recipient, in the case of (21) from Kofi to God. This is the opposite direction of that one signaled by HIDE. This apparent paradox can be dealt with if we assume that there is movement at two levels. HIDE implies that the entity here referred to as 'something' moves (or is kept away) from someone, while GIVE has the function of showing that the action of hiding affects God. It is merely an affectedness marker in this case. Affectedness may be extracted as a subpart of the meaning of GIVE, involving a flow of action from an energy source to the affectee. Any idea of benevolence as an attribute to GIVE is suppressed, and this is exactly what happened in (19), too. Thus this is a maleficiary-of-action sort of construction, a negative counterpart of the beneficiary-of-action type discussed in previous examples.

Malefactive by means of GIVE

215

Pike (1970: 10) cites an example from Bariba, conforming to the HIDE type above: (21) ba kuku - na (tonal morpheme for "benefactive")(Pike 1970: 10) They hid for each other (from each other).' Examples of this kind of malefactive or adversative are not easy found, languages using GIVE in serial verb constructions do not always go as far as to allow for semantic extension of GIVE to the extent necessary in malefactive constructions. In Akan one has not encountered serial ma 'GIVE' in malefactive constructions of the TELL LIES or HIDE type (Saether0 Andenes: p.c.). They would rather be interpreted in the beneficiary-of-action sense. In Yoruba, however, one can find a clear structural and semantic parallell in the ablative examples below. Ablative is normally used to denote an in-flective case relation which indicates the source of a movement. Fun (fu) is the word for GIVE in the language. By the grammarians it is glossed as 'from' below, evidently because it suits the reading of the sentence better than GIVE or FOR: (22) nw5 she

sa fü run-away from

u him

(23) mo yera fü I myself withdraw from

(Yoruba; Rowlands 1976 in Boretzky 1983) u him

(Yoruba; Boretzky 1983)

First of all, the complex predicate here lacks the usual kind of THING participant associated with the GIVE scene. Syntactically there is no direct object, if we assume that the postverbal part functions as an oblique object. Furthermore, it is again striking to see the contrast in energy flow comparing the host verb and GIVE in this context. In an utterance with GIVE we expect movement of something starting with the subject referent of the sentence, i.e the energy source, and having the direct object referent as the endpoint. Fun can be used that way too in Yoruba; as an independent verb and as a marker of recipiency or benefactive in complex predicates. But in (22) and (23) the subject

216

OleT.Fagerli

referent moves away from the object referent. It is clear that the change in direction is due to the verbs sa and yera which imply motion away from the direct object. This does not necessarily mean that the ordinary attribute of direction (to beneficiary/recipient) associated with GIVE is suppressed, but that there are two levels of direction; the agent's motion away (from someone), indicated by the first verb, and the connection this motion has to a specified patient of the motion, indicated by fü. Fü here has a character more like an ablative case marker than a full-fledged verb. The semantics of these sentences is similar to that of the ones marking malefactive affectedness above. Whether the object referent is negatively affected by the act however, such that we could apply the term malefactive, is uncertain from the scarce data I possess. My knowledge of Yoruba does not go far beyond the examples quoted here, and that keeps me from drawing any conclusion. (24) below may also be read in the ablative sense because something is being moved away from the patient. (24) abai di wosu da mi he-buy the house GIVE me 'He bought the house from me.'

(Saramaccan; Grimes/Glock 1970 in Boretzky 1983)

We would expect GIVE in this case to have a benefactive/recipient reading, like in (11-12), thus meaning 'He bought a house for/to me'. Allegedly, Saramaccan uses GIVE differently in this case from the other serializing languages we have looked at. GIVE here has lost its idea of benevolence, and probably functions more like a marker of affectedness. Despite the inherent ablative characteristics of BUY, there is no real negative affectedness associated with the act. The patient is nevertheless affected by the buying act, in the sense that he loses control over a thing. Thus, the forward (allative) direction normally indicated by GIVE needs not stand in conflict with the ablative direction indicated by the verbs HIDE, WITHDRAW, and BUY. The malefactive and ablative interpretations of the sentences are due to these verbs.

Malefactive by means of GIVE

217

9. The polysemy of GIVE morphemes We may consider both the serial GIVE and the benefactive extension approach at coding the related types of events described in the previous sections as special instances of a cognitively and semantically basic GIVE scene. Note that the uses of GIVE in the serializing languages encountered here do not complete the picture of polysemous structures for GIVE in the respective languages. We have only seen a few examples of occurences of GIVE in a limited syntactic environment, that of serial verb constructions. And this fact would naturally limit the range of potential semantic extensions of the verb as a whole. On the other hand, it has been shown that in the serial verb construction, where two or more verbs share semantic and syntactic structures, GIVE tends to be viewed as an affectedness marker of the complex predicate. With the help of the meaning of the host verb or its complement, GIVE can be viewed as a malefactive or ablative marker. As far as I know, and plausibly on semantic grounds, GIVE as a simple predicate can never mark malefactive or ablative. In Newman (1996: 233, 249) an overview of possible figurative extensions of GIVE typologically is presented within the framework of Cognitive Grammar. Newman in general considers both GIVE as an independent full-fledged verb and instances of dative case, serial verbs and directional prepositions. Benefactive and allative verbal extensions are not included, however. The main categories of semantic extensions are as follows: (i) (ii) (iii) (iv) (v) (vi) (vii) (viii)

interpersonal communication (give somebody a message) emergence / manifestation (give good crops) causation / purpose (focus on the GIVER'S role as volitioner or causer) enablement (enable somebody to do or be in control of something) schematic interaction (agent-patient interaction, give the car a wash) recipient / benefactive movement (into or away) completedness (relatable to inherent perfectivity of GIVE)

218

OleT.Fagerli

It would take us too far from the scope of this study to comment on all of these semantic extensions, suffice here to refer to Newman (1996) for a detailed analysis. However, the absence of the malefactive in the list deserves some comments. Malefactive would certainly not count as a main category of semantic extension of GIVE, I have argued that it belongs to an overall recipient category, thus falling into group (vi) above. Newman counts harmful things given to recipients as atypical of the GIVE scenario, so no specific status is given to that kind of malefactive. But he notes that the dative case in for instance German, Czech, and Polish may encode both benefactive and malefactive effects on the recipient (Newman 1996: 115-117). The dative is also an issue in Janda's (1997) article on GIVE, HAVE, and TAKE in Slavic. Experiencer is a cover term used by Newman and Janda to denote the beneficiary and the maleficiary. Whichever the label, what is shown to be the case is that an apparent opposition in meaning can be given a natural semantic explanation. The GIVE morpheme by itself never seems to mark malefactive, it is the context in which it appears which triggers the overall malefactive reading. Schematic meaning of entity to which an act is directed

Schematic meaning of person affected by an act

Affectee not involved in act

Schematic meaning of goal to which movement is directed

Affectee involved in act as recipient of concrete thing

Goal at end of path

The lizard ran towards the tree

Beneficiary of act

Maleficiary of act

Beneficiary of recipiency

Abbo works for me Abbo lied against me 1 picked a mango He took it from me to her

Figure 2. Schematic network of Fula -AN-

I picked a rotten mango to her

Malefactive by means of GIVE

219

What about allative (goal)? In most cases, this function is not considered a figurative extension of GIVE. It is rather a semantic subpart of literal GIVE, both participants, the THING moving and the GOAL to which it moves, are usually concrete (whether animate or inanimate). The verbal extension in Fula has been given a fairly complete treatment of polysemous structures. The meanings related to the -ANsuffix are summarized in the schematic network in Figure 2. Serial GIVE as we have looked at it, would follow approximately the same network of meanings. I have attempted to present a rather fine gradation of how different malefactive and benefactive events are distributed in the relatedness network. The maleficiary-of-action type could be further divided by two different kinds of verb meanings involved in the complex predicate, one like TAKE, which is of the ablative kind, and another like LIE, which is inherently an adversative act. Somehow, literally or figuratively, all these subcategories of -AN- are instantiations of the recipiency part of GIVE events.

10. Additional remarks We have seen that what is in the unmarked cases a benefactive marker, whether it is a verbal extension or serial use of GIVE, can extend its semantic space in appropriate circumstances to denote malefactive situations. To a large extent the two differing structural types, serial verb constructions and verbal extensions, match each other semantically. So far I have not come across serial GIVE in expressions of allative the way it may be used in Fula. It is perhaps not plausible to think of the verb corresponding to GIVE in such languages as the most natural source for denoting an event of motion towards a goal. Serializing languages rather apply directional verbs like GO or BE DIRECTED TO for this purpose. Actually, Fula dialects may apply the verb GO in serial constructions. Although not used productively as a benefactive marker,

220

le T. Fagerli

various expressions of allative as well as concrete recipiency can readily be found: (25) Mi neldi derewol yah-ugo Nigeria. I send-VAP sheet GO-INF Nigeria Ί sent a letter to Nigeria.' (26) Mi winnd-ii leeter faa-de I write-TNS letter GO-INF Ί wrote a letter to Aminata.' (Pulaar = Fula of Senegal)

(Fula)

e Aminata. (Pulaar; PREP Aminata Sylla 1979)

The beneficiary-of-action type and malefactive is not expressed through GO. We would not single out the grammaticalization potential of the verb, however. In a historical perspective, it is reasonable to hypothesize that the source for the destinative extension in Fula could be some directional verb like GO rather than GIVE. This may explain the difference in semantic function between the extension in Fula and serial GIVE in other languages. The allative function would probably not be traced back to GIVE, while it is easy to imagine that it could originate in a former verb for GO.4 GO is presumably more basic conceptually than GIVE. The polysemy of the destinative extension and of serial GIVE does not end there, though. A matter of further investigation is the relationship between benefactive and causative markers, both in serializing and extensional languages. It is clear that the word for GIVE in several languages is used to express causative relations. In Newman's list of main categories of semantic extensions of GIVE in the above section, causation is one of them. Examples (27)-(28) from Akan show the word ma 'GIVE' used with a causative function. In (28) it is even used twice in one utterance, first with a causative (manipulative) effect, then with a clear beneficiary-of-action meaning. Additional examples are available from Akan in Saether0 (1997) and from diverse languages cited in Newman (1996: 17Iff.). Examples of a combined causative and benefactive verbal affix are

Malefactive by means of GIVE

221

discussed for Ngiti and Lendu by Kutsch Lojenga (1997), and for Hausa by Mark Lobben (conference paper, June 1998). (27) Kofi ma-α abofra no bo-o famu. Kofi GIVE-PST child def fall (=hit)-PST ground(in) 'Kofi made the child fall down.' (Akan; Saether0 1997) (28) o-ma-a me ye-ε adwuma ma-α no. s/he-GIVE-PST me do-past work GIVE-PST him 'He made me work for him.' (Akan; Saether0 1997) No direct parallels are found with benefactive extensions in Fula of today. However, Endresen (1994) has looked at the etymological relationship between the Fula benefactive and causative suffixes, -ANand -(I)N-, respectively. Phonologically, they could have the same origin, the vowel in the -Vn- suffix being subject to vowel assimilation. Endresen's grammaticalization hypothesis implies that an allative element has developed in stages to include first the recipient function, then benefactive, and lastly the causative function. I assume that the malefactive function as described in this paper accompanies the recipient and benefactive stages.

Notes 1. Capitalized letters are used to represent basic-type categories, such as the verb GIVE as well as participant categories within the verb. 2. The term extension in this paper is used both to denote verbal extensions as a morphological phenomenon in some languages and semantic extension. The context will hopefully clarify any potential confusion caused by the use of this (polysemous) term. 3. Broken Fula in Aadamaawa (or pidginized Fula, sometimes referred to as Bilkiire) tends to use the form -in- instead of -an-, a linguistic variant not noted earlier by grammarians as a systematic one. The -in- suffix coincides with the causative suffix -(i)n-, a matter of confusion to learners of Fula. Goal, however, is never rendered by any other form than -an-.

222

Öle T. Fagerli

4. According to Eli Saether0 Andenes (p.c.), Akan uses serial ma 'GIVE' in cases where the recipient is a person (recipient of thing), e.g. 'throw to Ama', while the verb Jfo 'go' is used for inanimate goals.

The Norwegian and Russian reflexive-middlepassive systems and Cognitive Grammar Hans-Olav Enger and Tore Nesset

The paper presents an extensive analysis and comparison of the Norwegian and Russian reflexive-middle-passive systems, drawing especially on Kemmer (1993), but presenting some improvements. Further, the paper shows why indirect reflexives with beneficiary and recipient participants are best analysed as two different situation types. With reference to the Norwegian distinction between seg and seg selv, the term "superheavy marker" is introduced. It is shown how the categorisation networks of Cognitive Grammar (Langacker 1987α, 1990, 1991) are valuable for the typological study of voice systems.

1. Introduction: Aims and organisation This paper has both an empirical and a theoretical aim. The former is to present an extensive analysis of the Norwegian and Russian reflexive-middle-passive systems. In this respect we draw on Kemmer's (1993) analysis of Norwegian and Janda's (1993) of Russian. However, our account goes beyond these in that we consider a wider range of data. In particular, we consider the important distinction between Norwegian seg and seg selv, which has been amply studied in generative work on Scandinavian syntax (cf. e.g. Dalrymple 1993; Hellan 1988; Hestvik 1992), but which Kemmer has chosen not to discuss. The middle voice has been a topical research problem in cognitive linguistics (cf. e.g. Janda 1993; Maldonado 1988, 1993; Manney 1990, 1992, 1995, 1998). Whereas the focus in most of these works has been on particular properties of the middle voice in particular

224

Hans-Olav Enger and Tore Messet

languages, our theoretical aim is to show how categorisation networks of the type employed in Cognitive Grammar, cf. Langacker (1987a, 1990, 1991) provide a valuable tool for typological research. Our analysis yields two implications for the typological study of voice systems: • We improve Kemmer's terminology by introducing the term "superheavy marker". • We demonstrate why indirect reflexives with beneficiary and recipient participants are best analysed as different situation types. The paper is organised as follows. After presenting the relevant data in section 2, we establish categorisation networks for the two languages in section 3 and section 4. The former focuses on extension relations, the latter on instantiation relations. Section 4 concludes and summarises the contributions of the present paper.

1. The Norwegian and Russian systems: Situation types and markers In what follows we shall explore 18 situation types and see how they are marked in Norwegian and Russian. This provides us with the data to be accommodated in terms of Cognitive Grammar categorisation networks in section 2 and section 3, but it also allows us to make two points of more general interest for typology. First, we shall improve Kemmer's (1993) terminology by advancing the term "superheavy marker". Secondly, we shall argue that reflexives with beneficiary and recipient markers represent different situation types, a question that is left open by Kemmer. We shall discuss the functions of the Norwegian and Russian pronouns seg (selv) and sebja and the related affixes -s and -sja.1 We shall refer to them jointly as "reflexive-middle-passive markers". In keeping with Kemmer's (1993) terminology, we shall label the pronouns "heavy markers" (H) and the affixes "light markers" (L)

Reflexive-middle-passive

225

since the pronouns display more phonological body (cf. Kemmer 1993: 25). The inventory of forms is set out in Table 1. As indicated by the table, Norwegian seg is in some sentences (e.g. (1) below) accompanied by selv, whereas in other cases it occurs in isolation (cf. e.g. (7) below). In what follows we shall refer to both seg and seg selv as "heavy markers", but when it is necessary to distinguish between them, the former will be referred to as a "bare heavy marker" and the latter as a "superheavy marker". We believe this to be a significant improvement to existing terminology. Kemmer (1993), who does not treat the distinction between Norwegian seg and seg selv, lacks appropriate terminology. Admittedly, she discusses e.g. Surselvan sesez under the heading "emphatic reflexive", but this term is unfortunate from a methodological point of view. In trying to clarify the various meanings or functions of certain forms, it is essential to apply a terminology that itself refers to form only. Otherwise, one runs the risk of circularity in taking the relation for granted before it is established on independent grounds. The term superheavy has the advantage of being consistent with Kemmer's heavy and light marker. Table L Markers in Norwegian and Russian Norwegian Russian

Heavy marker (H)

Light marker (L)

seg (selv) sebja

-s -sja

Our analysis relies on the notion of "situation type", which Kemmer (1993: 7) defines as "sets of situational or semantic/pragmatic contexts that are systematically associated with a particular form of expression". Kemmer (1993: 7) emphasises that "semantic/pragmatic contexts" are not to be equated with "simple 'real-world contexts' existing independently of the language-user; situational contexts include real-world information, but that information is necessarily filtered through the conceptual apparatus of the speaker. This conception of situational contexts thus allows for the obvious role of the language-user in construing particular 'real-world' situations in different ways (cf. ... Langacker 1987a) [emphasis original]."

226

Hans-Olav Enger and Tore Nesset

Table 2 summarises the situation types and markers to be discussed in the present study. The list of situation types is adapted from Geniusiene (1987), Pederson (1991), and Kemmer (1993), but we exclude types not attested in either language. Table 2. The uses of Norwegian -s and seg and Russian -sja and sebja 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18.

Prototypical reflexive Indirect reflexive (recipient) Indirect reflexive (beneficiary) Natural collective Natural reciprocal Autocausative Grooming Reflexive causative Converse reflexive Mental events Spontaneous events Modal-deagentive Subject demotion Object demotion Absolutive Quasi-passive Reflexive passive Impersonal

HH: Superheavy marker (seg selv) L: Light marker (-s, -sja)

Norwegian

Russian

HH HH H H — H H H H H H — — — — L L L

H H H L L L L L L L L L L L L L L —

H: Heavy marker (seg, sebja) —: No marker attested

We are not going to discuss the characteristics of each situation type here, as they are already covered in the works cited above (see especially Kemmer 1993, chapters 3 and 4). However, representative examples from Norwegian and Russian are presented below for each situation type. A familiar problem is how to distinguish one situation type from another. We endorse Kemmer's (1993: 5ff.) approach which is to let cross-linguistic data decide. Thus, if one wonders whether what Kemmer refers to as "grooming" and "reflexive" represent one or two situation types, one should examine whether they are marked differently in some languages. Given that they are in

Reflexive-middle-passive

227

e.g. Russian (Ivan moetsja 'Ivan washes' vs. Ivan nenavidit sebja 'Ivan hates himself), one may conclude that one is dealing with different situation types. There may be practical problems involved in Kemmer's approach, but it nevertheless provides a well-motivated criterion. Examples for the attested situation types are given in (1) through (18).2 Notice that the numbering of the examples corresponds to the numbering of situation types in Table 2. Consider first situation types 1 to 3, prototypical and indirect reflexives, the only situation types where Russian has the heavy marker sebja. Norwegian evinces the superheavy marker in 1 and 2, but the bare heavy marker in 3. In the typological literature, indirect reflexives with recipient and beneficiary participants are normally regarded as subtypes of one and the same situation type. However, Kemmer (1993: 74-75) notes that recipients and beneficiary participants normally are typologically distinct, and hints at the possibility that "[c]loser typological investigation may result in the necessity of treating these two situation types as distinct". For the purposes of her study, however, Kemmer treats the two as one type. Given that recipient and beneficiary NPs are marked differently in indirect reflexives in Norwegian, we submit that they must be considered different situation types according to the criterion for delimiting situation types discussed above. Thus our analysis allows us to answer a question left open by Kemmer (1993).

(1) PROTOTYPICAL REFLEXIVE a. Jeg elsker meg selv og har ingen rivaler. I love segselv.l.SG and have no rival:PL Ί love myself and have no rivals.' (Fredrik Stabel) b. On nenavidit sebja. he:NOM hate:3.SG sebja:ACC 'He hates himself.' (Haiman 1983: 804)

228

Hans-Olav Enger and Tore Nesset

(2)

INDIRECT REFLEXIVE (RECIPIENT) a. Han gav seg selv en bok. he gave seg selv a book 'He gave himself a book.' b. On podaril sehe knigu. he:NOM gave:MASC.SG sebja-.ΌΑΎ book:ACC.SG 'He gave himself a book.'

(3)

INDIRECT REFLEXIVE

(BENEFICIARY)

a. Jeg vil bygge meg en g rd. I will build seg:l.SG a farm Ί shall build myself a farm.' (children's song) b. On kupil sebe dom. he:NOM bought:MASC.SG sebja-.ΌΑΎ houserACC.SG 'He bought himself a house.' We now turn to situation types 4 through 11, which differ from the previous in that Russian evinces the light marker -sja instead of the heavy sebja. Norwegian uses the bare heavy marker.3

(4)

NATURAL COLLECTIVE a. Slektningene samlet seg rundt bruden. relative.the:PL gathered seg around bride.the The relatives gathered around the bride.' b. Sobralos' mnogo znakomych. gathered: 3. SG.sja many acquaintance :GEN.PL Ά lot of acquaintances gathered.' (Usakov 1935)

(5)

NATURAL RECIPROCAL a. (not attested) b. Korabli vstrecajutsja u vchoda ship:NOM.PL meet:3.PL.5/a at entrance:GEN.SG v gavan'. in harbour:ACC.SG 'The ships meet at the entrance of the harbour.' (Usakov 1935)

Reflexive-middle-passive (6)

229

AUTOCAUSATIVE

a. Jon reiste seg motstrebende, Jon arose seg reluctantly 'Jon arose reluctantly.' (Hellan 1988: 114) b. On lozilsja v poster. he:NOM lay:MASC.SG.5/a in bed:ACC.SG 'He lay down on the bed.' (Usakov 1935)

(7)

GROOMING a. Jeg vasker meg. I wash seg:l.SG Ί wash.' b. Ja kazdyj den' brejus'. I:NOM every:ACC.SG day.ACC.SG shave:l.SG.sja Ί shave every day.'

(8)

REFLEXIVE CAUSATIVE a. Har du kuppet de g hos Nilsen Fris0r? have you:SG cut seg:2.SG at Nilsen Fris0r 'Have you had your hair cut at Nilsen Fris0r?' b. Ona operirovalas' v novoj klinike. she:NOM operated:3.SG.sja in new:LOC.SG clinic:LOC.SG 'She had an operation at the new clinic.' (Janda 1993: 313)

(9)

CONVERSE REFLEXIVE a. Trtzrne speiler seg i vannet. tree:PL mirror seg in water.the The trees are mirrored in the lake.' b. Nebo otrazaetsja v ozere. sky:NOM.SG mirror:3.SG.sja in lake:LOC.SG The sky is reflected in the lake.' (Geniusiene 1987: 228)

(10) MENTAL EVENTS a. Alle barn gleder seg til jul. all children look-forward seg to Christmas 'All children look forward to Christmas.'

230

Hans-Olav Enger and Tore Nesset

b. / vot teper' my vse raduemsja and here now we:NOM all:NOM.PL are-happyil.PL.sja i veselimsja, and enjoy 1.PL.sja 'And now we are happy and enjoy ourselves.' (Usakov 1935)

(11) SPONTANEOUS EVENTS a. D0ren pnet seg langsomt. door.the opened seg slowly The door opened slowly.' (Hellan 1988: 113) b. Dveri zakryvajutsja. door:NOM.PL close:3.PL.5/a The doors are closing.' (Janda 1993: 313) Russian displays the light marker in examples (12)-(15). No marker is attested in Norwegian; the Russian system is somewhat more elaborate. Notice the dative case in the experiencer in (12) and (13), and the use of genitive in (13) and instrumental in (14). (12)

MODAL-DEAGENTIVE

a. (not attested) b. Mne ne rabotaetsja. I:DAT not work:3.SG..s/a Ί cannot work.' (Geniusiene 1987: 289)

(13) SUBJECT DEMOTION a. (not attested) b. Mne chocetsja I:DAT want:3.SG.s/ absolutive events deagentive «--

1

F

t subjectdemotion

t quasipassive

I Superheavy marker

*

subject demotion

Heavy marker

Light marker

t quasi passive

1

No marker

Figure 1. Networks for Norwegian and Russian

An alternative might be to alter the network proposed in Figure 1, e.g. by assuming a direct connection between collective and reflexive

234

Hans-Olav Enger and Tore Messet

situations. We shall not pursue this issue; our point is to illustrate the heuristic merits of Cognitive Grammar networks; they force us to reconsider linguistic data and improve our analyses. Taken together, these three points suggest that extensive analyses of the reflexive-middle-passive system of each language are possible in terms of Cognitive Grammar networks. In addition, however, Figure 1 illustrates the value of such networks in typological research. The juxtaposition of the networks for Norwegian and Russian brings out the similarities and differences between the two languages in a simple and straightforward way, and thus facilitates cross-linguistic comparison.5

4. Instantiation relations: Reflexive, middle and passive The networks presented in the previous section are very similar to the semantic maps employed by typologists like Pederson (1991) and Kemmer (1993). In this section we shall see that Cognitive Grammar goes beyond the typological tradition, in that the instantiation relation allows us to integrate the cover terms reflexive, middle and passive in the networks. This enables us to explicate the similarities and differences in the distribution of the markers in the two languages with regard to the cover terms. Once again, then, the Norwegian and Russian evidence supports our general claim about the value of Cognitive Grammar networks in typological research. Sentences (19)-(21) (also listed as (Ib), (7b) and (17b) above) deserve further discussion; they provide good examples of reflexive, middle and passive voice. Following Kemmer (1993: 201ff.) we may characterise these three cover terms by means of the five features listed in Table 3. The three situation types are united in that the chief participant is directly affected by the verbal action. In the case of reflexives and middles like (19) and (20), the affectedness is due to the agent subject being identical to the patient of the action, whereas in the case of the passive in (21), the patient is the subject of the sentence.6

Reflexive-middle-passive

235

(19) PROTOTYPICAL REFLEXIVE On nenavidit sebja. 'He hates himself.' (Haiman 1983: 804)

(20) MIDDLE (GROOMING) Ja kazdyj den' brejus'. Ί shave every day.'

(21) PASSIVE Takie knigi citajutsja sirokimi massami. 'Such books are read by the broad masses.' (Janda 1993: 313) The second feature, "degree of elaboration of events", is introduced by Kemmer (1993) in order to distinguish between middle and reflexive voice. The intuition behind this feature is that reflexive situations like hating are situations that are normally directed towards other persons or entities, whereas for middle situations like shaving direction towards oneself is the expected case. Since in middle situations the action is normally directed towards oneself, the two participants involved are distinguished to a lesser degree. The less the participants are distinguished, the less the situation is elaborated. Kemmer arranges reflexive and middle situations on a cline ranging from transitive two-participant events to intransitive one-participant events. Reflexive situations are placed in the middle of the cline and are accordingly assigned an intermediate degree of elaboration, while middle situations are placed between reflexive and intransitive situations and hence show low degree of elaboration of events. (Kemmer (1993: 205) considers the passive to be "midway between a two- and a one-participant event", and we have therefore assigned the value "intermediate" as far as the degree of elaboration of events is concerned.) Features c) and d) set the passive apart from the middle and reflexive. As opposed to these voices, the passive has an external agent insofar as the subject and the adjunct in the instrumental case are not coreferential in (21). Carrying a patient role the subject of passive sentences like (21) is obviously not volitional. The last

236

Hans-Olav Enger and Tore Nesset

feature, which concerns the salience of the participant that "might be conceived of as the originator of the event" (Kemmer 1993: 205), unites the passive and middle as opposed to the reflexive. Passive sentences like (21) involve demotion of the agent from subject to adjunct, and at least certain middle situations, e.g. those of emotion and spontaneous events, defocus the originator of the event in a similar way (cf. Kemmer 1993: 205). Table 3. Characterisation of reflexive, middle and passive a) b) c) d) e)

Dir. affectedness of chief part.: Deg. of elaboration of events: External agent: Volitionality of chief part.: Salience of originator of event:

Reflexive

Middle

Passive

+ intermediate

+ low

+ high

+ low

+ intermediate + low

If we apply the features to the data in Table 2, situation types 1-3 represent reflexive voice, 4-15 middle, and 16-18 passive.7 Given this, we are in a position to relate the use of the heavy and light marker in the two languages to the functions of reflexive, middle and passive. This is done in Table 4. Table 4. The marking of reflexive, middle and passive Reflexive: Middle: Passive:

Norwegian

Russian

HH H L

H L L

On the basis of this table, generalisations about the relation between meaning and form can be made. If we take meaning as the starting point, we can state that both languages mark reflexive by a heavy (superheavy or bare heavy) marker and passive by a light marker, but that the languages differ in the marking of the middle voice, insofar as Norwegian displays a heavy and Russian a light marker. If we take form as the starting point, we can say that in Norwegian the superheavy and bare heavy markers are reflexive and middle markers and

Reflexive-middle-passive

237

the light marker a passive marker, whereas in Russian the heavy marker is a reflexive marker and the light a middle and passive marker. Having considered the relationship between the situation types and the cover terms, we must ask how this relationship can be modelled. The typological approaches of Pederson (1991), Kemmer (1993) and others do not address this question, insofar as their semantic maps do not include explicit reference to the cover terms. Cognitive Grammar, on the other hand, offers a simple and straightforward answer to the question; the cover terms can be related to the situation types by means of instantiation and thus be incorporated in the networks. This is shown in Figure 2, where the instantiation relations are represented as solid arrows. In this network "reflexive", "middle" and "passive" represent Schemas that each subsume a number of the functions from the networks given in Figure 2. The figure is simplistic in that the schema "middle" subsumes more than the two Schemas "autocausative" and "grooming" and that the extension relations are excluded in order to avoid crossing lines. The latter problem would not arise in a three-dimensional representation where situation types and cover terms would constitute different planes connected by instantiation relations. However, the simplified two-dimensional Figure 2 suffices to illustrate that Cognitive Grammar offers a straightforward modelling of the relationship between situation types and the terms "reflexive", "middle" and "passive". This way Cognitive Grammar provides a welcome addition to the notional apparatus employed in the typological tradition. REFLEXIVE

PROTOTYPICAL REFLEXIVE

IND. REFL. (RECIP.)

IND. REFL. (BENEF.)

QUASIPASSIVE

REFLEXIVE PASSIVE

IMPERSONAL

Figure 2. Instantiations of reflexive, middle and passive

The instantiation relation also allows us to explicate the similarities and differences in the distribution of the markers in Norwegian and Russian with regard to the Schemas of "reflexive", "middle" and

238

Hans-Olav Enger and Tore N esset

"passive". If we let these notions be represented by the properties given in Table 3, we are in a position to characterise the two systems in terms of the networks in Figures 3 and 4.

dir. affec. non-high elab. no extern, ag. volit. chief part.

... seg...

Figure 3. Reflexive, middle and passive in Norwegian

The mid level in the networks brings out the differences between the Norwegian and the Russian reflexive-middle-passive systems; in Norwegian, the middle goes with the reflexive, whereas in Russian it goes with the passive.8 We see that not only does the instantiation relation provide a means for modelling the relationship between "reflexive", "middle" and "passive" and situation types; it also facilitates cross-linguistic comparison. This again illustrates the value of Cognitive Grammar in typological studies.

5. Conclusion The contribution of the present paper is twofold. Firstly, we have presented extensive analyses of the reflexive-middle-passive systems in Norwegian and Russian. Our analyses have both enabled us to improve Kemmer's (1993) terminology by introducing the term "super-heavy marker" and to show why indirect reflexives with

Reflexive-middle-passive

239

recipient and beneficiary participants are best analysed as different situation types (a question that is left open by Kemmer). As a second main point, we have shown how Cognitive Grammar categorisation networks allow us not only to account for the reflexive, middle and passive voices in each language, but also to set out the similarities and differences between the two languages in a lucid and straightforward way. Therefore, Cognitive Grammar proves very valuable in typological research.

dir. affec. non-high elab. low sal. of orig.

sja

PASSIVE sja

REFLEXIVE ... sebja

Figure 4. Reflexive, middle and passive in Russian Abbreviations: dir. affec.: non-high elab.: low sal. of orig.: no extern ag.: volit. chief part.:

direct affectedness of chief participant non-high degree of elaboration of events low salience of originator of event no external agent volitional chief participant

Notes 1. Whether -sja is an affix or a clitic is disputed. We follow Schoorlemmer (1997) and Nesset (1998a, b) who conclude that -sja is best regarded as an affix. It is worth mentioning that -sja shows some allomorphy, but since allomorphy is irrelevant for present purposes, the affix will be referred to as -sja throughout

240

2.

3.

4. 5.

6.

7.

Hans-Olav Enger and Tore Nesset

the paper. Norwegian seg is inflected for number and person, but in Table 1 below we give only the 3. person singular/plural form. There is also a possessive form sin which will not be discussed in the present paper. Full paradigms are given in Hellan (1988: 60). According to Pederson (1991), Norwegian -s must be analysed as several homonymous suffixes. However, we follow Enger (this volume) who argues that it is best considered one polysemous suffix. When no reference is given, Russian examples are checked with a native consultant. Some of the Norwegian examples are taken from secondary sources, but in most cases we rely on our own native speaker competence. In some examples the Norwegian heavy marker is meg (1st person singular) and deg (2nd person singular). The Russian light marker in some examples occurs in its postvocalic form -s'. Throughout the paper, Russian examples are cited in orthography transliterated according to the Scando-Slavica system. In the case of a restricted and relatively small set of verbs, the light marker is possible in these situations, but we will not treat such exceptional marking in this paper. For thorough discussion, see Enger and Nesset (1999). For a smaller and somewhat different network for the Russian light marker, see Janda (1993: 312). One may speculate whether the similarity between the Russian and Scandinavian reflexive-middle-passive systems may be due to a Sprachbund around the Baltic Sea (Dahl and Koptjevskaja-Tamm 1992: 17ff.). Although this question is interesting in itself, it does not bear on any conclusion to be drawn in this paper; Cognitive Grammar networks are equally valuable for comparison of linguistic systems whether their similarities are due to contact or not. It is debated in the literature whether affectedness really is an appropriate label. Dimitrova-Vulchanova (1996: 162-163) is skeptical towards Kemmer's notion of affectedness in connection with the reflexives: "it is highly disputable whether being affected by the act in question can be assumed for the elements encoded as subject." However, Dimitrova-Vulchanova accepts Kemmer's point about "low elaboration". For a discussion of affectedness and low individuation (low elaboration) from the perspective of cognitive linguistics, see Manney (1995). Admittedly, this is simplistic in one respect; as can be seen from Table 2, situation type 3 has a heavy marker in both languages. Thus, this situation type does not behave like a reflexive in Norwegian. We submit that it represents a middle situation in Norwegian, but a reflexive in Russian. This move may appear completely ad hoc, but is in fact supported by empirical evidence. In Norwegian, these verbs do not occur with a non-coreferential indirect object:

Reflexive-middle-passive

241

(i) *Jens bygde meg en gärd. 'Jens built me a farm.' This suggests that indirect reflexives with a beneficiary marker denote actions normally directed towards the self, and thus qualify as middle, not reflexive in Kemmer's terms. In Russian, on the other hand, indirect reflexives do occur with non-coreferential indirect objects, behaving like true reflexives, cf. (ii): (ii) Ivan postril mnefermu. 'Ivan built me a farm.' For an extended discussion, see Enger and Nesset (1999: 45-48). 8. Our analysis in figure 5 of -sja as signalling low salience of the originator of the event is in harmony with an analysis presented by Adger Williams, who characterises -sja as "a device that enables the speaker to bring an entity into focus; i.e. to talk about it, without having to be distracted by the mention of other entities" (Williams 1993: 181).

Norwegian reflexive-middle-passive -s: Homonymy/polysemy and morphology Hans-Olav Enger

This paper discusses the Norwegian reflexive-middle-passive marker -s. This suffix is inflectional when it functions as a passive marker, and derivational in all other functions. In §2, it is discussed whether the various s-suffixes are polysemously related or mere homonyms. The answer is "polysemy". This has some interesting consequences. It creates problems for many morphological models; apparently, a "split morphology" must be rejected. Another consequence relates to the view on polysemy/homonymy propounded by Pederson (1991), who argues that whenever the direct connection between two nodes in a network is cut off, homonymy must be posited for those pans of the network that no longer cohere. On the basis of the Norwegian case, this principle is questioned, and theoretical arguments against it are adduced. In §3, the question is raised, why only the passive is inflectional. It is argued that the other "meanings" of -s (notably reflexive and reciprocal) are of higher relevance, in the sense of Bybee (1985). Therefore, it is no surprise that they are derivational. This explanation is indirectly an argument for Bybee's relevance concept.1 1. Facts and questions In Norwegian Bokmäl, the verb suffix -s serves to mark a number of functions or meanings; • passive, as in b0ker skrives 'books are written', • reflexive, as in herrene samles rundt sigarene 'the gentlemen gather around the cigars', and • reciprocal, as in vi tales i neste uke 'we'll talk next week'.2

244

Hans-Olav Enger

The suffix -s can be called a reflexive-middle-passive marker (Enger and Nesset this volume). It is sometimes derivational and sometimes inflectional. As a passive marker, -s is inflectional and productive. In all other functions, -s is derivational and fairly improductive. Two questions arise: 1. Should we posit one polysemous suffix -5, or several homonymous suffixes? 2. If we prefer polysemy: why is -s inflectional only when it is passive?

2. Homonymy or polysemy — and some implications 2.L Homonymy versus polysemy We begin with question 1 above: are the various -s'es polysemically connected or mere homonyms? The homonymy-polysemy distinction is notoriously tricky, but we shall consider some standard criteria. A first criterion of polysemy is semantic relatedness. Semantically, there is a relatively close relationship between the reflexive and the reciprocal (cf. e.g. Langacker 1991: 370-371). The passive is related to both of these, since in the reflexive and the reciprocal, the grammatical subject is "both affected and affector", whereas in the passive, the subject is "affected". So this criterion indicates polysemy. We now turn to the second criterion, cross-linguistic comparison: as Croft (1990: 166) says, "if many diverse languages independently have the same pattern of "homonymy", then the meanings are closely related." The idea is that homonymy, being accidental formal identity, should not recur in many languages. This idea is defended at length by Haiman (1985). For example, the English preposition to can be used both to indicate direction towards some location (We took the bus to Hadeland) and to indicate transfer of possession (He

Norwegian reflexive-middle-passive marker -s

245

gave a book to his father). Now, English also has a homophonous numeral two. However, it can be shown that the formal identity between direction towards location and transfer of possession is quite common, cross-linguistically. By contrast, the formal identity between direction towards location and the numeral '2' is uncommon. Therefore, the former case of formal identity qualifies as polysemy, the latter as phonological homonymy. Now, the passive and the middle can be expressed by the same suffix in Norwegian, as they can in Russian (cf. Enger and Nesset this volume), and the explanation probably has nothing to do with genetic affiliation. This indicates polysemy. A third criterion and argument for polysemy is the fact that within the Norwegian grammatical tradition, verbal -s is (more or less explicitly) treated as a polysemous suffix (see e.g Bei to 1986 or Faarlund, Lie, and Vannebo 1997). This may not be a strong argument in itself, but it is surely relevant. A fourth criterion and argument for polysemy is methodological. Other things being equal, polysemy is preferable to homonymy for two reasons. Firstly, semantic distinctions can be maximised indefinitely, so unless polysemy is given priority, homonymy can be posited almost everywhere (Lyons 1977: 550-569). Secondly, to posit homonymy is to say that a case of formal similarity is a mere coincidence. To posit polysemy, by contrast, is to say that there is some reason for the formal similarity. A priori, scientists look for reasons and do not postulate coincidence unless forced to. Thus, polysemy should normally be the more attractive hypothesis. This argument is especially important for those linguists that prefer to focus on the non-arbitrary aspects of language. Leiss (1997: 136) even talks of the imperative of a non-arbitrary grammar: Die Form ist unbedingt ernst zu nehmen! ("The form must absolutely be taken seriously!")3 The upshot is that the Norwegian verbal suffix -s is a case of polysemy rather than homonymy. To specialists in Norwegian, this conclusion is unsurprising, but it has some interesting consequences, to which we now turn.

246

Hans-Olav Enger

2.2. A morphological problem Within a Word-and-Paradigm (WP, or realisational) model of morphology (e.g. Anderson 1990, 1992; Matthews 1972, 1991; Aronoff 1994) it is difficult to show the similarity between -s as an inflectional suffix and as a derivational one. The reason is that the two suffixes are found in different "modules" in the grammar. Within a WP model, two forms have been created either by one and the same rule, or by two different rules, and there is no convincing way of generalising over rules. Thus, the derivational suffix -s is introduced by a derivational rule, the inflectional suffix -s by an inflectional rule, and the two implicitly emerge as unrelated, i.e., as homonyms.4 There might be various ways of trying to save a WP model here. One could be to use the "meta-rules" suggested by Matthews (1991: 201-204). This would amount to postulating a third rule, a rule on a meta-level, which generalises over the inflectional rule that introduces -s and the derivational rule that introduces -s. There are problems with meta-rules, however. They are introduced as an extra device in order to cope with some problematic examples, and do not have any independent empirical justification. Furthermore, if linguistic theory strives for psychological plausibility, the meta-rules may seem ad hoc,5 and a defense that relies on meta-rules is too weak. Another possible defense of the WP model might rely on morphomes, as we know them from Aronoff (1994) -s might be analysed as a morphome, essentially an element of form without meaning. This morphome would then be unspecified with respect to the inflection/derivation distinction (just like Aronoffs (1994: 34-36) Latin example of a morphome). The problem is that whereas morphomes are meaningless, -s is polysemous (cf. §2.1); it does have a meaning, and it should therefore not be analysed as a morphome. The root of this problem may be that WP theorists have exaggerated the distinction between inflection and derivation. This is understandable, as the distinction is both controversial and valuable. Yet it should not be considered categorical, i.e., even if we accept the inflection/derivation distinction, we should not accept a "split morphology", where inflection and derivation are seen as categori-

Norwegian reflexive-middle-passive marker -s

247

cally different (see e.g. Wurzel 1984: 40-51; Bybee 1985: 81-110; Dressier 1989; Booij 1993; van Marie 1996). It is the sharp and categorical split between the two kinds of morphology that ultimately makes it difficult to generalise over the different uses of -s in WP models.6 A sketchy and incomplete Cognitive Grammar description of Norwegian -s is given in Figure 1. direct affectedness of chief participant non-high elaboration of events

Figure L Norwegian -s

A more comprehensive study is presented by Enger and Nesset (this volume), so we need not go into all the details here. However, it may be worth pointing out that at least some problems can be avoided within Cognitive Grammar. Firstly, the inflection/derivation distinction need not be seen as totally discrete (cf. Langacker 1987a: 18). Secondly, in Figure 1, the labels "reflexive", "reciprocal" and "passive" denote the semantic content of the suffix. These semantic contents are not identical, but they are fairly similar, and they can be subsumed under a common label (viz. 'direct affectedness of chief participant' and 'non-high elaboration of events'). In Cognitive Grammar, this is described by positing a relation of instantiation. Thus, it is shown that the three meanings are related without being one and the same. As for phonological form, all the three look alike, as indicated by Figure 1. The passive is boldtyped in Figure 1, to show that it is more entrenched. This description seems promising, but, admittedly, the inflection/derivation distinction is not directly brought out in Figure 1.

248

Hans-Olav Enger Chaining They followed themselves out of the room.

Coref w/ sub-ord. subject He said himself to do it.

Grooming He combed himself.

Collective They gathered themselves in the square.

Prototype/ Semantic ref. He hit/saw himself

"Dative" reflexives -recipient -benefactive/affected -oossessor reflexive

Autocausatives - Body posture self-raise - Trans, motion self-run Non-trans, motion self-bow

Reflexive causative / sewed myself a shirt at the tailor's.

Partitive He buttoned himself up. Inchoative/ Deagentive/ Anticausative The door opened itself

Converse reflexives The trees reflect themselves in the water.

Modal-deagentive The book read itself quickly by me. Absolut) ve Watch out! That dog bites itself.

Quasi-Passives: Subject Demotion It doesn 't work itself for me.

Potential The books read themselves well.

I

Necessity The "gh" pronounces itself like an "/". ψ Resultative , The shoes wore themselves out.

ι

Reflexive passives: With external force The door opened itself from the wind. With external agent The door opened itself by me.

Impersonal Itself signs the document.

Figure 2. Network of potential functional extensions of the reflexive construction

Norwegian reflexive-middle-passive marker -s

249

2.3. Networks and homonymy/polysemy A clear tendency in Cognitive Grammar has been to maximise polysemy. It is therefore surprising that according to a principle suggested by Pederson (1991), who is inspired by Cognitive Grammar, Norwegian -s should be analysed as homonymous. Pederson (1991: 92) posits a semantic map, a network of functions for "reflexive markers"; see Figure 2. Furthermore, Pederson (1991: 348) claims that when the connection between two nodes is cut off, we must posit homonymy for those parts of the network that no longer cohere: "distal functions, relatable only via now lost central functions, are no longer conceptually linked. To wit, these severed distal functions are synchronically homonymous, despite the clear diachronic motivation for their expression by the same form." Norwegian -s would be a case in point, as the so-called semantic reflexive is expressed by seg. In Pederson's terms, the semantic reflexive is the "now lost central function". (For further details, see Enger and Nesset this volume). The implication of Pederson's words, then, is that -s should be analysed as homonymous, after all. However, a number of arguments for describing Norwegian -s as polysemous were given in §2.1 above; and some may be added here: generally, Pederson a priori denies the possibility of so-called doughnut categories ever being polysemous, and this may seem unfortunate, in the perspective of grammaticalisation theory. In the case at hand, Pederson's argument entails treating the passive -5 and other verb suffixes -s as homonymous suffixes. They are just as unrelated as, say, passive -s and the so-called possessive -s that is added enclitically to nouns in Norwegian (as in Pedersons avhandling 'Pederson's dissertation'). This seems somewhat counter-intuitive. More importantly, even if the connection between the passive and the middle may be weak, it simply does not follow that the connection must be non-existent: there is a difference between "weak semantic connection" and "no semantic connection". In other words, Pederson's argument presupposes that the homonymy-polysemy distinction is a simple binary dichotomy. This is unfortunate; the

250

Hans-Olav Enger

distinction is better seen as a gradient (cf. Lyons 1977: 552; Langacker 1990: 268). Enger and Nesset (this volume) argue that even though the semantic maps as employed by Pederson and others yield valuable insights, they are somewhat too simple, since they only model extension relations (i.e., relations between a prototype and its extensions). This leaves no room for "classical" criterial-attribute relations, which presumably also matter in categorisation; they are referred to as instantiation relations in Cognitive Grammar. Hence, categories may have something in common semantically, even if the direct connection between them on a semantic map has been cut off. For example, it is characteristic of all s-forms in Norwegian that the chief participant is directly affected by the verbal action, no matter whether the verb is "passive", "reflexive", "reciprocal" or something else. Even if the common meaning is fairly abstract, this is nevertheless an additional argument against characterising the different uses of -s as mere homonyms.7 To conclude, the Norwegian verbal suffix -s should be considered polysemous. However, if we regard the polysemy/homonymy distinction as gradient, we ought to ask about the degree of polysemy (Sandra and Rice 1995: 98). In fact, we could follow Pederson and use networks for this purpose: degrees of polysemy could be captured by counting "empty" nodes. If, for example, the connection is broken off at several points in a network, this may be a "weak" polysemy, i.e., closer to homonymy. This is a tentative suggestion, but it would be useful in the case at hand: the derivational categories expressed by -s are closer to each other than to the passive, and this is shown clearly both in Pederson's, Kemmer's and Enger and Nesset's networks. Thus, even if we claim polysemy for the verbal suffix -5, we have to accept the semantic distance being larger between the passive and the other categories than between for example the reflexive and the reciprocal.

Norwegian reflexive-middle-passive marker -s

251

3. Relevance S-forms with a passive meaning are inflectional; other s-forms are derivational (cf. §1). Why are only the passives inflectional? The answer may be that the passive is less "relevant", in Bybee's (1985) sense. In simplified terms, the passive makes a smaller semantic contribution than the other categories. This explanation takes semantics as its starting-point. Question 2 in §1 above can hardly be answered any other way, and it is not obvious (to me, at least) that any other current approach to morphology has any answer.

3.1. The concept of relevance According to Bybee, two parameters decide whether a category is expressed as inflection or as derivation. One of these parameters is relevance. A category is relevant for the verb to the extent that its meaning directly affects the lexical content of the verb stem (Bybee 1985: 15). The larger change of meaning a category induces, the more relevant it is.8 The other parameter is generality, i.e., semantic compatibility. A category that can be combined with many verbs is high in generality. Whether a category is expressed inflectionally or not is decided by these two semantic parameters. A category has to be so relevant that it can be inflectional; a category of very low relevance will not be grammaticalised. On the other hand, for a category to be inflectional, it must be relatively general as well, so it cannot be too relevant, either. Derivational categories may be of higher relevance. Relevance and generality are not completely independent; high relevance tends to detract from generality (Bybee 1985: 17). More relevant suffixes are expected to occur closer to the verb stem than less relevant ones.

252

Hans-Olav Enger

3.2. What are the meanings of -s ? We shall see that the passive is less relevant than the other categories expressed by the suffix -s. However, the immediate problem then is to decide what other "meanings" are expressed by -s. The literature gives no unanimous answer; Beito (1986) operates with four different meanings, Faarlund, Lie, and Vannebo (1997) with five. At this stage, however, we shall simply follow Faarlund, Lie, and Vannebo (1997): their grammar is reasonably new and authoritative, and has the advantage of being easily "translatable" into Bybee's terms. Faarlund, Lie, and Vannebo (1997: 511-512) say that the following five categories are expressed by -s: a) the reflexive, b) the reciprocal, c) the inchoative, e.g. dages 'become day', d) others, e) the passive.9 Nothing sensible can be said here about the category "others". What we need to show, is that the reflexive, the reciprocal and the inchoative are more relevant than the passive.

3.3. Bybeean relevance meets -s To begin with, the reflexive and the reciprocal are more relevant than the passive; they change the action denoted by the verb more than the passive does. This is to say that the semantic "distance" is shorter between the active and the passive than between the active and the reflexive and the reciprocal. The reason is that both in the active and in the passive, the agent is unaffected by the action. In the reflexive and the reciprocal, by contrast, the agent is both carrying out the action and is affected by it (Bybee 1985: 20). The higher relevance of the reciprocal and the reflexive can be illustrated by examples. Admittedly, even if a category is more relevant in one example, it need not follow that it always must be more relevant, but the examples are probably fairly representative. We begin with the reciprocal. Consider tales. It has two slightly different meanings. One example, viz. the reciprocal, was given in §1: Vi tales i neste uke 'we'll talk next week'. The other meaning is shown in e.g. Det tales sä mye om morfologi na om dagen lot is

Norwegian reflexive-middle-passive marker -s

253

said about morphology these days'. This second meaning is passive. The reciprocal is derivational here, the passive is inflectional. Compare paraphrases with bli 'become', another sort of passive in Norwegian: Del blir talt sä mye om morfologi na om dagen is acceptable, but *Vi blir talt i neste uke is not. The semantic distance between reciprocal tales and the active form tale is larger than the semantic distance between the passive tales and the active form. The reason is as follows: it is possible to tale without there being any conversation involved: Hitler talte til massene 'Hitler spoke to the masses'. Likewise, it is possible to tales —in the inflectional, passive sense—without there being any conversation involved: Det tales sä mye om morfologi, men del er bare enetale. lot is said about morphology, but it is just a monologue'. In the case of the derivational tales, by contrast, there must be a conversation. This indicates that the semantic distance between the active and the reciprocal is larger than that between the active and the passive; i.e., that the reciprocal is more relevant. At this stage, a possible objection should be considered: despite its intuitive appeal, the idea of 'semantic distance' may be difficult to measure. Even though the active form tale is similar to the passive form tales in that none of the forms obligatorily refer to a conversation, the passive form differs from the active form in that the agent is not only unexpressed, but also unspecified, and cannot easily be supplied ("Dei tales sä mye om morfologi av meg lot is said about morphology by me' is not perfectly grammatical). By contrast, reciprocal tales and active tale are similar as regards the specificity of the agents. This objection is interesting, but it may be discarded if it can be argued that the active-passive similarity is more important than the active-reciprocal similarity. Fortunately, such an argument can be supplied: conversation is somehow a different activity from monologue. Thus, the reciprocal -s induces a change in meaning that has to do with the world as we perceive it (whatever that may be!); in Bybee's (1985: 5) terminology, we may talk of a "semantic change". By contrast, whether an agent is expressed or specified, does not really change the activity denoted by tale. Thus, the

254

Hans-Olav Enger

objection can be discarded; the passive -s is less relevant than the reciprocal. We now turn to the reflexive; consider sarnies. An example of the passive sense is papir sarnies inn pä mandager 'paper is collected on Mondays'. (Compare the near-synonymous papir blir samlet inn pä mandager.) This is different from the reflexive example given in §1 above, herrene sarnies rundt sigarene. The reflexive simply cannot be paraphrased by a construction with bli. In the reflexive sarnies, there is an element of intentionality that is absent in the passive. In this way, the passive is more similar to the active, because the active samle is usually combined with an inanimate object. Notice also that the reflexive can be paraphrased by a different verb, e.g. stimle sammen. The reflexive is thus more relevant than is the passive.10 The inchoative, finally, is clearly fairly relevant. "To begin to do something" is different from "to do something". (A methodological problem here, however, is that those verbs that have -s as an inchoative suffix, do not usually have counterparts without -s.) We have now seen that the inchoative, the reflexive and the reciprocal are more relevant than the passive. In consequence, it comes as no surprise that these categories are derivational while the passive is inflectional. However, on the basis of Bybee (1985), we cannot predict that these categories must be derivational in Norwegian. Thus, our explanation is not "nomothetic"; rather, it is an explanation in terms of motivation. The notion of motivation is explicated at great length by a number of authors (see e.g. Lakoff 1987; Nesset 1998b: 27-34), so it will not be discussed in depth here. Suffice it to say that what is motivated, is not completely predictable, but still far from arbitrary. An explanation in terms of motivation may not fit with widespread positivist ideas of what constitutes a 'scientific explanation', since motivation does not give us any basis for prediction. Lass (1980) in particular has emphasised prediction as the ultimate criterion for scientific explanation. However, strict insistence on prediction may perhaps, somewhat irreverently, be characterised as a case of 'physics envy': after all, meteorologists do not make 100% correct predictions either (at least not in the Scandinavian

Norwegian reflexive-middle-passive marker -s

255

countries!); yet it would be absurd to deny meteorology the status of a science. An explanation in terms of motivation also qualifies as an explanation. To the best of my knowledge, no other morphological theory than Bybee's comes anywhere near explaining why only the passive -s is inflectional in Norwegian.

3.4. Relevance and the order of suffixes The value of relevance becomes even clearer when we consider the fact that Norwegian children sometimes rearrange the order of -s and other suffixes. The standard past tense of synes 'think, feel' is syntes. The past tense suffix t thus comes "inside" the -s. Since past tense is usually seen as inflectional in Norwegian, whereas -s here is derivational, this is anomalous.11 It is therefore no surprise that Norwegian children sometimes produce synste, thereby setting the order of the suffixes right, as it were, by getting more relevant suffixes closer to the root (cf. §3.1). In the terminology of Haspelmath (1993), this is a case of "externalisation of inflection" (see Enger 2000). Similarly, Icelandic children sometimes produce the "incorrect" forustum instead of "standard" forumst 'we perish', cf. Anderson (1990: 270, fn. 5). Anderson comments that "[s]uch forms [as forustum, HOE] apparently represent the misanalysis of -st as a simple derivational suffix, to which inflectional suffixes are then added, as opposed to the structure-making constituent argued for in the subsequent text here." By Anderson's analysis, then, forustum is just a mistake, and not terribly interesting, either. However, the parallelism with Norwegian indicates that there is more to this example. In fact, Birkmann (1997) presents good arguments for analysing Icelandic -st as a derivational suffix, and Ottosson (1992) has made a similar case for Old Icelandic -st. Furthermore, there are parallels to Norwegian synste in the dialect (or perhaps language) of Dalarna, Sweden, as described by Levander (1928: 168). Levander notes that in that dialect, -s (the cognate of Norwegian -s) "is often inserted in front of the suffix proper" (my

256

Hans-Olav Enger

translation), such as in minnsom 'we remember', where the -s is placed closer to the root than the person/number agreement marker om. This development, like the Norwegian one, makes perfect sense if we take -s to be derivational, and assume Bybee's relevance to be at work here. In fact, Kemmer (1993: 189), who does invoke Bybeean relevance in the Icelandic example, points to a parallel to the Icelandic development in Surselvan (a Romance language). Since an analysis by means of relevance make cases like forustum something more than an isolated error, this analysis seems superior to one which cannot.12

4. Conclusion This paper set out to answer two questions (cf. §1: 1) Should the Norwegian verb suffix -s be analysed as several homonyms or one polysemous suffix? We have found that the answer should be "one polysemous suffix". 2) Why is -s inflectional only when it is passive? The answer is that the passive is of lower relevance, in Bybeean terms. These issues should be of interest not only for the study of Norwegian. More specifically, even though the distinction between inflection/derivation is accepted, the paper provides arguments against a "split morphology"; in §2.2, it turned out that the categorical split between the two kinds of morphology is the reason why it is so difficult to generalise over the different uses of -s in WP models. While the paper is based in part on networks (also known as semantic maps), it has shown why we should be cautious in using them as the only evidence for homonymy/polysemy. The paper has also presented support for Bybee's (1985) relevance concept. These conclusions have bearing on both morphological theory and cognitive semantics.

Norwegian reflexive-middle-passive marker -s

257

Notes 1. I am indebted to Rolf Theil Endresen, Staffan Hellberg, Ronald Langacker, Tore Nesset, Tomas Riad, an anonymous referee, and the participants at The verb in Cognitive Grammar for very helpful comments on different versions of this paper. Thanks also to Osten Dahl for making me aware of -s in Dalarna, and to Mark Aronoff for e-mail clarification of the nature of morphomes. The mistakes that no doubt remain are my own. 2. We concentrate on Norwegian Bokmäl, but Nynorsk, Swedish and Danish seem similar in most relevant respects. 3. Some linguists take common etymology as a criterion for polysemy (e.g. Beard 1995: 33), but this is highly disputable (cf. Lyons 1977: 550-551, and, from a different angle, Leiss 1997: 145-147). Anyway, the -s suffixes can all be traced back to one common etymon sik. 4. This problem is probably not restricted to WP models. For example, Halle and Mohanan (1985: 63) posit two distinct English suffixes -ing (one adjective ending, one verb ending). The two suffixes are located at different "strata", and not explicitly connected. They are thus treated as if they were homonyms. However, this paper focuses on WP only. 5. The arguments against meta-rules are taken from Nesset (1994: 17-18) and Enger (1998a: 44-45). 6. Anderson (1990) has argued that Icelandic st-verbs, which are the cognates of Norwegian s-verbs, create problems for a morpheme-based model of morphology; problems that do not arise in a WP-type model (such as Anderson's "a-morphous morphology"). Yet Norwegian -verbs create problems for a WP-type model, too. 7. An additional objection is that Pederson's argument relies crucially on the architecture of his network. If the network had been drawn differently, what now looks like homonymy would instead look like polysemy. It is therefore important that the network approach in cognitive linguistics is not entirely unproblematic, cf. Sandra and Rice (1995). Sandra and Rice (1995: 94) even say that "there seem to be as many network models as there are network modellers"; Pederson (1991: 90) himself says that his diagram is greatly simplified, and his diagram differs in some minor details from Kemmer's (1993). Therefore, it seems premature to base decisions about homonymy versus polysemy on network evidence only. 8. Admittedly, the concept of semantic change is not quite as well developed in Bybee's theory as one might have wished. This problem cannot be pursued here, however, and we shall see that Bybeean relevance is still interesting.

258

Hans-Olav Enger

9. The label "others" may seem strange, but it is not unfair. In the original wording, Faarlund, Lie, and Vannebo (1997: 512) say: "d) Andre s-verb som ikke faller inn i betydningsgruppene ovenfor", in my translation: "Other sverbs that do not fall into the semantic groups above". 10. An interesting complication here is that the semantic distance between the active and the passive form of tale may be larger than that between the active and the passive form of samle. However, this complication does not challenge the analysis presented here. 11. Faarlund, Lie, and Vannebo (1997: 508) argue that, since -s comes "outside" the inflectional suffix in such cases, it is problematic to consider -s derivational. However, there is independent evidence for taking -s in these cases to be derivational, as has traditionally been done (cf. Enger 2000 for a more detailed discussion). Moreover, derivational suffixes do sometimes occur inside inflectional suffixes (see e.g. Wurzel 1984: 41, Dressler 1989, Booij 1993, Rainer 1996). After all, if derivation and inflection were defined in terms of position only, then it would amount to a non-empirical truism to say that derivation is inside inflection. 12. Recently, Bybee's concept of relevance has been challenged by Beard (1995: 54-55). He argues that affixes are ordered either according to some universal principle such as Bybeean relevance or according to language-specific rules (as assumed by Anderson 1992). Beard concludes that "the actual ordering of features is, as Anderson maintains, a variable of the parameter set by individual languages. The lack of consistent relevance of ordering across languages casts doubts on Bybee's hypothesis". However, it is not obvious why affix ordering cannot be influenced by both universal and language-specific factors. Stump (1997: 239, fn. 3) explicitly points out this possibility. Hence, Beard's argument may be an example of what Langacker (1987a: 28) has dubbed "the exclusionary fallacy". Much of morphology may in fact be subject to such partly conflicting influences between universal and language-specific factors; cf. Wurzel (1984). Without Bybeean relevance, it is also difficult to account for synste.

Control and transitivity: A study of the Norwegian verb love 'promise' Kristian Emil Kristoffersen

The Norwegian verb love 'promise' can appear in several different complementation patterns, among them the following two: (i) subject and infinitival complement, (ii) subject, nominal object and infinitival complement. These two patterns exhibit a phenomenon known as argument control. In this paper I propose an account of these two patterns in terms of construction grammar. Furthermore, I present some data which demonstrate that pattern (i) above is far more frequent than pattern (ii). Finally, I suggest an explanation of this difference in terms of type frequency.'

I. Overview This paper is an investigation of the relationship between transitivity and argument control in one Norwegian verb; love 'promise'. The term TRANSITIVITY refers to the ability of a predicator to take one or more objects, whereas ARGUMENT CONTROL refers to a relation of coreference between a logical, (obligatorily) unexpressed subject of a complement clause and another argument in the linguistic context of this subject. Consider the following Norwegian examples: (1)

a. Politiet päbyr folk police-the orders people b. Kari nekta Per a reise Kari denied Per to go

a to til to

sykle pä fortauet. cycle on pavement-the Roma. Rome

260

Kristian Emil Kristoffersen

(2)

a. Jeg I b. Jeg I

0nsker ä reise til Tyskland. wish to go to Germany planlegger ä reise til Tyskland. plan to go to Germany

In the two sentences in (1) the logical subject of the infinitival complement is coreferent with the direct object of the matrix verb; thus, they are instances of object control. In contrast, the sentences in (2) display what is called subject control, where the coreference relation holds between the logical subject of the complement verb and the matrix subject. In both cases, however, control arises from lexical semantic properties of the matrix predicates in combination with the presence of an obligatorily subjectless complement. The examples in (l)-(2) are part of a more general pattern in Norwegian and other languages, showing a strong correlation between on the one hand monotransitive syntax and subject control, as in (1), and on the other ditransitive syntax and object control, as in (2). In §2 below I analyse the syntactic and semantic patterns shown in (l)-(2) within the theoretical framework of CONSTRUCTION GRAMMAR. In my analysis these two sets of sentences instantiate two distinct constructions, which I call the OBJECT CONTROL CONSTRUCTION and SUBJECT CONTROL CONSTRUCTION respectively. It will be argued that the different control relations expressed by the two sets of sentences follow from the semantic characteristics of these constructions. However, not all Norwegian control verbs appear to fit easily into either of the two constructions mentioned above. The Norwegian verb love is a case in point here. Consider the examples in (3): (3)

a.

V-INF De hadde they had b. V-NP-1NF De hadde they had

lovet ä komme. promised to come lovet meg a komme. promised me to come

The Norwegian verb love 'promise'

261

(3a) has monotransitive syntax, and can thus be analysed as an instance of the subject control construction. (3b), on the other hand, has the syntax of the object control construction, while at the same time expressing a subject control relation. In other words, there seems to be a conflict between syntax and semantics in (3b). In §3 I suggest a solution to this apparent conflict in terms of the notion of MULTIPLE INHERITANCE. Concerning the complementation of love, (3a and b) are not the only possible syntactic patterns associated with this verb. Compare these sentences with the pair in (4): (4)

a. V-S De hadde they had b. V-NP-S De hadde they had

lovet at de skulle komme. promised that they would come lovet meg at de skulle komme. promised me that they would come

If we consider the data in (3) and (4) against the correlation noted above between monotransitive syntax and subject control, and ditransitive syntax and object control, we might expect the pattern illustrated in (3 a) (V-INF) to be more frequent in actual usage than the pattern illustrated by (3b) (V-NP-INF). Furthermore, we expect that the pattern in (4b) (V-NP-S) is more frequent than the pattern in (3b) (V-NP-INF). In §4 I present some data showing that these expectations are to a large extent borne out. Finally, in §5 I suggest an explanation of the difference in frequency between the V-INF and the V-NP-INF patterns in terms of Bybee's (1985, 1995a) correlation between high type frequency and productivity. In sum,I have the following four aims with this paper: 1. To propose an account of argument control in terms of two distinct control constructions (§2), 2. to propose an analysis of the two different syntactic patterns the verb love can appear in as a "control verb", i.e., V-INF and VNP-INF (§3),

262

Kristian Emil Kristoffersen

3. to present a set of data that demonstrate that the V-INF pattern of love is more frequent than the V-NP-INF pattern of the same verb (§4), and 4. to explain this difference in frequency (§5).

2. A construction grammar analysis of argument control In this paragraph I first briefly present relevant aspects of the theoretical framework employed in this paper, construction grammar (§2.1).2 Next, I propose an analysis of argument control within this framework (§2.2).

2.1. Construction grammar Within construction grammar syntactic constructions are regarded as basic linguistic units. Consider the following four sentences: (5)

a. b. c. d.

Mary gave Jill some flowers. Peter sent Mary the book. Jill told the children a story. John promised Peter a book.

These four sentences are taken to be instances of an abstract construction, which Goldberg (1995) calls THE ENGLISH DITRANSITIVE CONSTRUCTION. This construction is related to, but different from, the individual verbs that appear in it. It carries a particular meaning, which is expressed by a particular form. Within this framework there is no substantial difference between constructions and words, as both types of unit can be characterised as "form-meaning-function complexes [differing] only with regard to internal complexity" (Michaelis andLambrecht 1996: 216). The English ditranstive construction, which is instantiated by the four sentences in (5), is an example of a LINKING CONSTRUCTION. A basic property of linking constructions is that they express systematic

The Norwegian verb love 'promise'

263

correspondences between the thematic arguments conventionally associated with these constructions, and the syntactic elements that express these roles. For example, on Goldberg's analysis of the English ditransitive construction, the construction is argued to be a mapping between the three arguments agent, recipient and patient, and the syntactic functions subject, object and secondary object. Labels like agent, recipient and patient refer to what is called ARGUMENT ROLES. Argument roles belong to constructions. However, also the verbs which enter into linking constructions are associated with roles; these latter roles are referred to as PARTICIPANT ROLES. For example, Goldberg (1995: 50-51) calls the three participant roles of the English ditransitive verb hand, which appears in sentence (6), 'hander' (Mary), 'handee' (Peter), and 'handed' (book), while the corresponding argument roles are agent, recipient and patient. (6)

Mary handed Peter a book.

When a verb occurs in a construction, its participant roles are said to be fused with the corresponding argument roles. Thus, in sentence (6) the participant roles 'hander', 'handee' and 'handed' of hand are fused with the argument roles agent, recipient and patient of the ditranstive construction. A general constraint on fusion is that for a participant role and an argument role to be fused, they must be semantically compatible, i.e., the participant role must be an instance of the argument role (The semantic coherence principle, Goldberg 1995: 50). In other words, the participant role 'hander' can be fused with the argument role agent since the former can be construed as an instance of the more general agent argument role, i.e., a 'hander' is a kind of agent. Likewise, a 'handee' is a kind of recipient, and 'handed' is a kind of patient. Another aspect of syntactic constructions is that they are said to LICENSE the elements that instantiate them. The ditransitive construction, for example, licences three syntactic arguments, subject, object and secondary object.

264

Kristian Emil Kristoffersen

As noted above constructions are pairings of form and meaning. Concerning the meaning of constructions, it has been argued that many constructions have polysemous senses. A case in point is the English ditransitive construction which, according to Goldberg (1995: 38), has (7a) as its central sense, and (7b-c) as two of its related senses. Verbs motivating the various senses are given in (8). (7)

a. Agent successfully causes recipient to receive patient. b. Agent intends to cause recipient to receive patient. c. Agent acts to cause recipient to receive patient at some future point in time.

(8)

a. Peter gave Mary a cake, b. Peter baked Mary a cake. c. Peter left Mary a letter.

The fact that constructions are associated with several distinct, but systematically related senses is referred to as CONSTRUCTIONAL POLYSEMY. This is connected to the claim made within construction grammar that there is no substantial difference between words and constructions, cf. the following statement by Goldberg (1995: 32): "[SJince constructions are treated as the same basic data type as morphemes, that they should have polysemous senses is expected."

2.2. A constructional account of argument control As noted in the introductory paragraph the term ARGUMENT CONTROL designates a relation of coreference between the unexpressed (logical) subject of an infinitival complement and a syntactic argument (subject, object or oblique) in the linguistic context of this complement. In what follows I propose an account of control in which the notion of linking construction plays a central role. Compare first the sentences in (9) with those in (10):

The Norwegian verb love 'promise'

(9)

a. Karl nekta Per Kari denied Per b. Jeg planlegger I plan

(10) a. Kari nekta Per Kari denied Per b. Jeg planlegger I plan

ä to ä to

reise go reise go

til to til to

Roma. Rome Tyskland. Germany

265

(= (lb)) (= (2b))

mat. food en reise. a journey

An important difference between the two pairs of sentences is that the pair in (9) contains controlled complements, whereas the pair in (10) does not. This contrast can be described from two different perspectives. On the one hand one could see the two instances of nekte 'deny' as two verbs with distinct but related senses. One sense, which involves control, is present in (9a); the other, which does not involve control, is found in (lOa). Thus, control is seen as part of the semantics of the verb. However, this description fails to account for the fact that control in (9a) appears to arise from the combination of the argument structure of nekte (three-place) and a particular syntactic form of the third argument (infinitival complement). An alternative analysis which presents a solution to this problem would be to analyse the difference between (9a) and (lOa) as involving one and the same verb which participates in two distinct syntactic contructions. This is the basis for the analysis proposed below (§2.2.1-2.2.2), and argued more extensively for in Kristoffersen (1996), where object and subject control are analysed in terms of two types of constructions, the object control construction and the subject control construction.

2.2.1. The object control construction Object control refers to a coreference relation between a grammatical object and the logical subject of a complement infinitival which is a

266

Kristian Emil Kristoffersen

coargument of this object. Consider again the two examples of object control in (1) together with a third instance, all shown in (11): (11) a. Politiet päbyr folk ä police-the orders people to b. Kan nekta Per ä reise til Kari denied Per to go to c. Politiet tillot folk police-the permitted people

sykle pä fortauet. (= (la)) cycle on pavement-the Roma. (= (lb)) Rome ä sykle pä fortauet. to cycle on pavement-the

Most Norwegian sentences involving object control refer to events in which one participant tries to get another participant to perform some subevent (expressed by the infinitival complement), as in (lla), or to prevent another participant from performing some subevent, as in (lib), or to enable another participant to perform some subevent, as in (lie). In accord with this I suggest to analyse the meaning of the object control construction as metaphorical extensions of Talmy's (1988) concept of a force dynamic opposition between an antagonist and an agonist. Together these extensions constitute the polysemous meaning of the object control construction, described in (12).

(12) THE SENSES OF THE OBJECT CONTROL CONSTRUCTION • Antagonist acts to get agonist to perform the subevent. • Antagonist acts to prevent agonist from performing subevent. • Antagonist acts to enable agonist to perform subevent. The object control construction is a linking construction specifying three semantic arguments, antagonist, expressed by the subject, agonist, expressed by the NP object, and a propositional argument expressed by the infinitival complement. Accordingly, this construction licences three arguments. That the object argument of this construction refers to the controller follows from the relationship between antagonist on the one hand and agonist and subevent on the other.

The Norwegian verb love 'promise'

267

2.2.2. The subject control construction Subject control refers to a coreference relation beteween a grammatical subject and the logical subject of a complement infinitival which is a coargument of this subject. Two examples which involve subject control are given in (13): (13) a. Je g I b. Jeg I

0nsker ä wish to planlegger plan

reise til Tyskland. go to Germany ä reise til Tyskland. to go to Germany

In Kristoffersen (1996) I analyse verbs like 0nske in terms of a semantic element 'desire' which is attributed to the subject referent, who has a desire to realise the subevent. A verb like planlegge, on the other hand, can be taken to express a semantic element 'intention', again attributed to the subject referent, who intends to realise the subevent. Several linguists have pointed to the fact that 'desire' and 'intention' are semantically related, often with further reference to expressions of a grammatical meaning future (Bybee 1985: 166; Bybee, Perkins, and Pagliuca 1994: 256; Heine, Claudi, and Hunnemeier 1991: 173-175; Wierzbicka 1988: 32). Accordingly, these two meaning elements can be analysed as constituting the polysemous meaning of the subject control construction. I follow Bybee, Perkins, and Pagliuca (1994: 256) in their view that the two elements are related by metonymy.

2.2.3. Summary The subject control construction is like the object control construction in being a linking construction, in this case licencing two arguments, one (experiencer with the 'desire' verbs and committor3 with the "intention" verbs) expressed by the subject and one by the infini-

268

Kristian Emil Kristoffersen

tival complement. The properties of the two constructions are summarised in (14) and (15).

(14) THE OBJECT CONTROL CONSTRUCTION • is a polysemous linking construction based on metaphorical extensions of a force dynamic opposition; and • licences three arguments, antagonist, agonist, and soa-arg,4 and • has the direct object as controller.

(15) THE (MONOTRANSITIVE) SUBJECT CONTROL CONSTRUCTION • is a polysemous linking construction comprising two meanings, "desire" and "intention", which are metonymically related; and • licences two arguments, experiencer/committor and soaarg, and • has the subject as the controller. The reason for using the term monotransitive subject control construction in (15) will become apparent in the next paragraph, where a distinction is made between monotransitive and ditransitive subject control constructions.

3. A constructional analysis of love 3.L The exceptional status of love as a control verb One Norwegian verb which does not fit easily into the description of argument control given in the preceding paragraph is love 'promise'. Compare sentences (3b) and (la) above, repeated here as (16a and b):

The Norwegian verb love 'promise'

(16) a.

269

V-NP-INF

De hadde lovet meg ä komme. they had promised me to come b. V-NP-INF Politiet pabyr folk ä sykle pa fortauet. police-the orders people to cycle on pavement-the The contrast between (16a and b) poses the following problem: These two sentences appear to have the same syntax. However, they differ with respect to the controller. Paby Order' has object control, while love has subject control. Thus, if we consider the contrast between (16a and b) in isolation, love is clearly exceptional, at least from a purely syntactic point of view (cf. Bresnan 1982: 355; Larson 1991). The view that this verb is exceptional gains strength by the fact that whereas päby is one of several ditransitive verbs in Norwegian which involve object control, love is an isolated instance of a ditransitive verb with subject control.5 Within the constructional approach to control presented in the previous paragraph the difference regarding controller choice between these two sentences cannot be straightforwardly accounted for, since they have the same syntactic form. If a particular syntactic construction is taken to express a particular meaning, and if controller choice is taken to follow from semantic properties of this construction, we should expect (16a) to have object control, and not subject control. However, this sentence has subject control, a fact which from a constructional point of view raises the following problem: According to the description in (15) the subject control construction licences two arguments, viz. the subject and the infinitival complement, whereas love in (16a) has three arguments. Thus, we may ask how the recipient object is licenced in (16a).

270

Kristian Emil

Kristoffersen

3.2. Multiple inheritance To answer that question, consider first (3a), repeated here as (17): (17)

V-INF De hadde lovet they had promised

ä komme. to come

This sentence is an instance of the subject control construction, expressing among other things an intention on the part of the subject referent to perform the action referred to by the complement, cf. (15). The two arguments are licenced by the construction. Next, consider (16a). As mentioned, the problem in this case is that the subject control construction cannot license the recipient object. However, this argument is clearly licenced when love appears in another construction, the DITRANSITIVE CONSTRUCTION, cf. (18): (18) Fetter Peter

lovet promised

meg en bok. me a book

Thus, the use of love observed in (16a)can be related to two different syntactic constructions. Stated in terms of argument licencing, we could say that two of the arguments of love in (16a) are licenced by one construction, the subject control construction, whereas the third is licenced by another construction, the ditransitive construction. This view of love as used in the pattern in (16a) (V-NP-INF) can be handled in terms of the notion of MULTIPLE INHERITANCE, which is characterised as follows by Goldberg (1995: 97):6 "Multiple inheritance allows us to capture the fact that instances of some construction types seem to resist being uniquely characterized in a natural way [...]." A relation of inheritance holds between two constructions A and B if the two have formal and semantic properties in common. The relationship between A and B is asymmetric, in the sense that one inherits from the other. Multiple inheritance, on the other hand, occurs when one construction inherits from more than one other

The Norwegian verb love 'promise'

271

construction, i.e., if it has formal and semantic properties in common with more than one other construction. Argument licensing in examples like (16a) can be accounted for as a case of multiple inheritance. I call the construction instantiated by this sentence the DiTRANSiTiVE SUBJECT CONTROL CONSTRUCTION, in contrast with the monotransitive subject control construction described in (15) above. This construction stands in a relation of multiple inheritance to two other constructions, which jointly licence the three syntactic arguments: the monotransitive subject control construction (15) licenses the subject argument and the prepositional argument, and the ditransitive construction licenses the recipient object. Furthermore, the subject control construction contributes with the subject control interpretation. This relationship of multiple inheritance is outlined in Figure 1; with argument licensing indicated by arrows. To recapitulate, (16a), in which love is used as a control verb in a ditransitive pattern can be taken to instantiate two distinct syntactic constructions; on the one hand the subject control construction (see (15)) and on the other the ditransitive construction. The three arguments of love in (16a) are licenced by both of these constructions. MONOTRANSITIVE SUBJECT CONTROL CONSTRUCTION

NP V

DITRANSITIVE CONSTRUCTION

VP

NP V NP

NP V NP VP DITRANSITIVE SUBJECT CONTROL CONSTRUCTION

Figure 1. Argument licencing in the V-NP-INF pattern of love

NP

272

Kristian Emil Kristoffersen

4. Love: patterns of use in written and spoken Norwegian In §3 two complementation patterns of love were discussed, V-INF and V-NP-INF. In (4) two patterns were exemplified, with the infinitival complements (cf. (16)) replaced by finite complement sentences. Examples of all patterns are repeated in (19): (19) a.

V-INF De hadde they had b. V-NP-INF De hadde they had c. V-S De hadde they had d. V-NP-S De hadde they had

lovet a komme. promised to come lovet meg ä komme. promised me to come lovet at de skulle komme. promised that they would come lovet meg at de skulle komme. promised me that they would come

In §1 I formulated the following two assumptions concerning the actual usage of the different patterns in (19) on the basis of the correlation between transitivity and control: • The V-INF pattern in (19a) is more frequent than the V-NPINF pattern in (19b). • The V-NP-S pattern in (19d) is more frequent than the V-NPINF pattern in (19b). The first assumption accords with the fact that in the V-INF pattern the correlation between monotransitive syntax and subject control holds, whereas this is not the case with the V-NP-INF pattern; because of this one would think that speakers and writers would prefer the former pattern. Similarly, the second assumption addresses the fact that of the two patterns V-NP-S and V-NP-INF only the

The Norwegian verb love 'promise'

273

latter displays the mentioned "conflict" between syntax and semantics. To see if the two assumptions are supported by actual usage, I have (a) examined a corpus of Norwegian texts, and (b) given two elicitation tasks to speakers of Norwegian concerning the difference between the patterns in (19). In what follows I present the results. 4.1. Love in a corpus of Norwegian texts7 All sentences in the corpus containing love with either finite or infinitival complement have been classified according to two variables, first, whether the verb takes a finite or infinitival complement, and, second, whether or not the sentence also contains a nominal object expressing the recipient. Sentences with finite complements have been included only if the complement subject is coreferent with the subject of love. Tables 1 and 2 present the number of occurrences for each of the four patterns illustrated in (1) and (2). Table 1. Choice of complement when nominal object is absent (corpus) Finite sentence Infinitival complement Total

5 34 39

Table 2. Choice of complement when nominal object Finite sentence 5 Infinitival complement 6 Total Π

12.8% 87.2% 100.0 %

(19c) (19a)

is present (corpus) 45.5% (19d) 54.5% (19b) 100.0 %

These tables show that in the majority of the examples without a nominal object (Table 1: 87.2 %) love takes an infinitival complement, not a finite sentence. However, when a nominal object is present, the corpus does not point in favour of any of the two alternative complement types, cf. Table 2. Note, however, that the total number of occurrences here is low, only 11 in all.

274

Kristian Emil Kristoffersen

4.2. Elicitation tasks The aim of the elicitation tasks was to find out what complement type subjects choose when asked to complete a sentence beginning with Jeg lover deg ... Ί promise you ....' or Jeg lover ... Ί promise ...'. The subjects were given the Norwegian equivalents of one of the following two sets of instructions:8 • Imagine the following situation: You want to make a promise to someone about something. What you promise has to be an action performed by yourself. You start by saying: "I promise you..." How would you finish this sentence? • Imagine the following situation: You want to make a promise about something. What you promise has to be an action performed by yourself. You start by saying: "I promise ..." How would you finish this sentence? Tables 3 and 4 summarise the results of these tests: Table 3. Choice of complement when nominal object Finite sentence 10 Infinitival complement 35 Total 45

is absent (elicitation) 22.2 % (19c) 77.8% (19a) 100%

Table 4. Choice of complement when nominal object Finite sentence 89 53 Infinitival complement 142 Total

is present (elicitation) 62.7 % (19d) 37.3 % (19b) 100%

These results give a somewhat clearer picture of the variation between finite and infinitival complementation with love than the results from the corpus study. Table 3, presenting the results for sentences without nominal object, supports the results from the corpus study (Table 1), whereas Table 4 shows that a majority of my subjects prefers finite complement in the presence of a nominal object (Table 2). Table 5 summarises the results from the corpus study and the elicitation tasks:

The Norwegian verb love 'promise'

275

Table 5. Choice of complements with love (corpus and elicitation) Elicitation Pattern Corpus V-INF (19a) 34 87.2 % 35 77.8 % 22.2 % V-S 5 12.8 % 10 (19c) V-NP-INF (19b) 6 54.5 % 53 37.3 % V-NP-S (19d) 62.7 % 45.5 % 89 5

To sum up: I pointed to two possible differences in frequency concerning the four complementation patterns illustrated in (19); (i) that the V-INF pattern is more frequent than the V-NP-INF pattern, and (ii) that the V-NP-S pattern is more frequent than the V-NPINF pattern. As Table 5 shows, (i) is supported by the results both from the corpus examination and the elicitation tasks. The second difference, however, is supported only by the results from the elicitation tasks. I have no answer to the question why this should be so. It may possibly be related to the fact that the diverging results correlate with the difference between written and spoken language. However, the number of instances of sentences with a nominal object present in the written corpus is too low to draw strong conclusions.

5. Why is V-INF more frequent than V-NP-INF? If we consider the difference between the V-INF pattern on the one hand and the V-NP-INF on the other, another question emerges: Why do so many speakers and writers of Norwegian prefer the former pattern to the latter? It seems to me that a plausible answer to this question can be given in terms of Bybee's (1985, 1995a) use of type frequency in explaining productivity. On the basis of morphological data from several languages Bybee argues that grammatical affixes which in a language L have high type frequency are more productive than affixes in L having low type frequency. What, then, would the type frequency of a construction be? In construction grammar grammatical constructions are considered the syntactic counterparts of grammatical affixes. Goldberg (1995: 22-

276

Kristian Emil Kristoffersen

23) says the following about the correspondence between these two units: "On the present account, the closed class grammatical morpheme is analogous to the English skeletal construction; the verb stem plays the role of the main verb. The semantic integration of verb and morpheme is analogous to the integration of construction and verb in English." From this point of view, the type frequency of a given construction in a language L is the same as the number of verbs appearing in that construction in L (cf. Goldberg 1995: 134). Let us now turn to the construction types which have been discussed in this paper. Consider first the monotransitive subject control construction (MSC), illustrated by the sentences in (2) and described in (15). (20) lists 16 other Norwegian verbs participating in this construction: (20) avslä 'refuse'; elske 'love'; foreslä 'propose'; foretrekke 'prefer';/0rs0A:e 'try'; hate 'hate'; kreve 'demand'; like 'like'; nekte 'refuse'; planlegge 'plan';pr0ve 'try'; sverge 'sverge'; tenke 'intend'; trenge 'need'; velge 'choose'; 0nske 'want' Against the 16 verbs in (20), love is one of three Norwegian verbs which can appear with ditransitive syntax and subject control. The two others are tilby Offer' and garantere 'guarantee', which under special circumstances can appear with subject control (Engh [1982]; Runde (1997), cf. note 5 above). This means that the construction instantiated by for example the sentences in (2) has much higher type frequency than the construction instantiated by for example (19a) (the ditransitive subject control construction, DSC), cf. Table 6 on the opposite page, based on the number of verbs in (20), and the three verbs love, tilby and garantere). In view of this difference in type frequency it is expected that love as a control verb primarily appears without recipient object. If, on the other hand, one wishes to express the nominal object as well, one can instead invoke the pattern instantiated by (19d), which in fact is what the majority of my subjects do (cf. Table 5), although this result is

The Norwegian verb love 'promise'

277

not corroborated by the results from the corpus examination (cf. Table 2). Table 6. Type frequency of the two subject control constructions Construction type MSC DSC Number of verbs 16 3 Type frequency 84.2 % 15.8 %

6. Conclusion In this paper I have analysed the Norwegian verb love 'promise' when used as a control verb, as in the sentences in (21): (21) a.

V-INF De hadde they had b. V-NP-INF De hadde they had

lovet promised

ä to

komme. come

lovet promised

meg ä me to

komme. come

Both sentences express a relation of subject control, i.e. a coreference relation between the subject of the matrix verb (V) and the logical subject of the infinitival complement (INF). Concerning the properties of this verb I formulated four aims for the paper in § 1: 1. To propose an account of argument control in terms of two distinct control constructions (§2), 2. to propose an analysis of the two different syntactic patterns the verb love can appear in as a "control verb", i.e. V-INF and VNP-INF (§3), 3. to present a set of data that demonstrate that the V-INF pattern of love is more frequent than the V-NP-INF pattern of the same verb (§4), and 4. to explain this difference in frequency (§5).

278

Kristian Emil Kristoffersen

In §2 two different patterns of argument control were discussed; subject control (22a), and object control (22b): (22) a. Kari nekta Per Kari denied Per b. Jeg planlegger ä I plan to

ä reise to go reise til go to

til Roma. to Rome Tyskland. Germany

In the proposed analysis of argument control (22a) was analysed as an instance of a linking construction, called the object control construction, with three argument roles, antagonist, agonist and soa-arg. (22b) was analysed as an instance of another linking construction, the subject control construction expressing one of two metonymically related argument roles, committor and experiences The analysis treats argument control not as a property of individual verbs but as a property of the two constructions. As far as I can see this is a natural way of accounting for the fact that many "control verbs" can be used in syntactic contexts where control is not involved, compare (22) with (23) and (24): (23) Per nekta Kari mat. Per denied Kari food (24) a. Jeg I b. Jeg I

planlegger plan planlegger plan

en reise. a journey at jeg skal reise. that I shall go

Within lexical approaches to argument control like, i.a., Bresnan (1982), Larson (1991), and Pollard and Sag (1994), the variation between (17) on the one hand and (23) - (19) on the other needs to be treated in terms of variable verb meanings. On the analysis proposed here a verb like planlegge is the same both in (22b) and (24); semantic differences between the use of this verb in the three sentences are to be attributed to the constructions it appears in.

The Norwegian verb love 'promise'

279

§3 was devoted to the second aim, in particular the question how the pattern instantiated by (21b) should be analysed with respect to the control constructions described in §2. The proposed analysis here builds on the notion of multiple inheritance: Sentence (21b) is taken to instantiate two different linking constructions, the subject control construction and the ditransitive construction. The three syntactic arguments of this sentence was taken to be licenced by both constructions simultaneously, the subject and the infinitival complement by the subject construction and the recipient object by the ditransitive construction. My analysis of (2la and b) implies that the V-INF pattern (2la) is the basic one with respect to control, whereas the V-NP-INF pattern (21b) is exceptional in the sense that it inherits one more construction than the basic pattern does. This is the opposite perspective of that taken by most lexical theories of the control properties of love, where V-NP-INF is seen as the basic pattern, and the V-INF as a variant pattern in which the recipient object is omitted.9 In §4, devoted to the third aim, I presented some data which showed that among speakers and writers of Norwegian the V-INF pattern is more frequent than the V-NP-INF pattern. This is to be expected if the former is the basic control pattern associated with love. Finally, in §5 I looked at differences between the two patterns in terms of type frequency, and concluded that it is expected that love as a control verb primarily appears without recipient object, since the type frequency of the monotransitive control construction is higher than that of the ditransitive subject control construction.

Notes 1. Thanks to Jan Engh, Anders N0klestad, Kjell Johan Saeb0 and two anonymous reviewers for useful comments. Errors are mine. 2. Here I build primarily on Goldberg (1995); other recent studies employing this framework are Zwicky (1994), Michaelis and Lambrecht (1996) and Kay and Fillmore(1999). 3. The term committor (Pollard and Sag 1994: 287) refers to a participant who commits to perform the action denoted by the infinitival complement.

280

Kristian Emil Kristoffersen

4. The term soa-arg (Pollard and Sag 1994: 287) refers to an argument expressing a state of affairs. 5. Two other ditransitive verbs in Norwegian, tilby Offer' and garantere 'guarantee' can under special circumstances appear with subject control, see Engh [1982] and Runde (1997), However, these two verbs differ from love in that the latter has subject control as the unmarked reading. 6. See also Michaelis and Lambrecht (1996: 237) and Kay and Fillmore (1999) for discussion of this concept. 7. The corpus contains approximately 1.5 million word forms, and consists of (A) three issues of the Norwegian newspaper Aftenposten: 940116; 940115; 930322, and (B) the following Norwegian (bokmal variant) novels: A. Briseid: Kjedereaksjoner, 1977; A. Eriksen: Promenade, 1937; A. Jensen: Ikaros, 1957; A. K. Elstad: Magret, 1977; B. Vestre: Sparet αν en sti, 1957; Ε. Griffiths: Lenket til kjcerligheten, 1957; G. Allbjart: Flukten til livet, 1937; G. Brantenberg: Egalias d0tre, 1977; G. Johannessen: H0st i mars, 1977; H. Geelmuyden: Periferi og sentrum, 1937; H. S. Dehlin: Men skyggen f0lger efter, 1937; I. Heiberg: Hjeml0s, 1957; J. Bj0rkelund: Ragnarokk! 1977; J. Brinchmann: Mannen som kom tilbake, 1937; J. Bj0rneboe: Under en h rdere himmel, 1957; J. H. Jensen: D0den er stamgjest, 1977; K. Alnaes: Gaia, 1977; K. Askildsen: Davids bror, 1977; K. Bj0rnstad: Vinterbyen, 1977; K. Fasting: Det gode kj0bmannskap, 1937; L. S. Christensen: Amat0ren, 1977; N. Grieg: Spansk sommer, 1937; R. Arntzen: N r alt er nytt, 1937; R. Mager0y: Gunhild, 1957; S. Christophersen: Demningen, 1957; T. Stigen: Frode Budbcereren, 1957. 8. The two sets of instructions were given to two different groups of subjects; thus, no informant has answered both sets. 9. Pollard and Sag (1994) represent a notable exception to this trend, when they state that some subject control verbs allow a third argument in addition to the subject argument and the infinitival complement.

References

Aasen, Ivor 1864 Abelin, Asa 1999

Norsk Grammatik [Norwegian Grammar]. [Reprinted 1965, by Universitetsforlaget, Oslo.]

Studies in sound symbolism. Gothenburg: Department of Linguistics, Göteborg University. [Gothenburg Monographs in Linguistics 17] Anderson, Stephen R. 1990 The grammar of Icelandic verbs in -st. In Joan Maling and Annie Zaenen (eds.), Syntax and Semantics 24, modern Icelandic syntax, 235-287. San Diego: Academic Press. 1992 A-morphous morphology. Cambridge: Cambridge University Press. Arnott, D. W. 1970 The nominal and verbal systems ofFula. Oxford: Clarendon Press. Aronoff, Mark 1976 Word formation in generative grammar. Cambridge, MA: MIT Press. 1994 Morphology by itself. Cambridge, MA: MIT Press. Baayen, R. H. 1991 Quantitative aspects of morphological productivity. In Geert Booij and Jaap van Marie (eds.), Yearbook of morphology 1991, 109-149. Dordrecht: Kluwer. 1992 On frequency, transparency and productivity. In Geert Booij and Jaap van Marie (eds.), Yearbook of morphology 7992, 181-208. Dordrecht: Kluwer. Bauer, Laurie 1983 English word-formation. Cambridge: Cambridge University Press. 1988 Introducing linguistic morphology. Edinburgh: Edinburgh University Press.

282

References

Beard, Robert 1995 Lexeme-morpheme-base morphology. Albany, NY: SUNY Press. Beito, Olav T. 1986 Nynorsk grammatikk. Lyd- og ordlcere [Nynorsk Grammar. Phonology and Morphology]. Oslo: DetNorske Samlaget. Bennett, David 1992 Stratificational grammar. In W. Bright (ed.), International encyclopedia of linguistics, vol 4, 77-79. Oxford: Oxford University Press. Berko, Jean 1958 The child's learning of English morphology. Word 14: 150-177. Berkov, Valerij Pavlovic 1989 Die Reflexiva in den skandinavischen Sprachen. In Ernst Walter and Hartmut Mittelstadt (eds.), Altnordistik: Vielfalt und Einheit. Erinnerungsband für Walter Baetke, 144-160. Weimar: Hermann Böhlaus Nachfolger. Birkmann, Thomas 1997 Das neuisländische Mediopassiv: Flexion oder Wortbildung? In T. Birkmann et. al. (eds.), Vergleichende germanische Philologie und Skandinavistik. Festschrift für Otmar Werner, 81-89. Tübingen: Max Niemeyer. Bjerkan, Kirsten Meyer 2000 Verbal morphology in specifically language impaired children: Evidence from Norwegian. (Acta Humaniora 73.) Oslo: Unipub. Bjerkan, Kirsten Meyer and Hanne Gram Simonsen 1995 Past tense processing in Norwegian: Experimental evidence from children and adults. Paper presented at NELAS 4 and Workshop on Dyslexia and Early Language Development, Jyväskylä, Finland, 79 September 1995. 1996 Prosessering av preteritumsformer i norsk: Eksperimentell evidens fra barn og voksne [Processing of past tense in Norwegian: Experimental evidence from children and adults]. Norsk Lingvistisk Tidsskrift 14: 189-207. Bloomfield, Leonard 1935 Language. London: George Allen & Unwin.

References

283

Bokmälsordboka [The Bokmäl Dictionary] 1986 Mark Ingebj0rg Landr0 and Boye Wangensteen (eds.). Oslo: Universitetsforlaget. Booij, Geert 1993 Against split morphology. In Geert Booij and Jaap van Marie (eds.), Yearbook of morphology 1993, 27-49. Dordrecht: Kluwer. Boretzky, Norbert 1983 Kreolsprachen, Substrate und Sprachwandel. Wiesbaden: Otto Harrassowitz. Breedveld J. O. 1996 Form and meaning in Fulfulde: A morphophonological study of Maasinakore. Leiden: CNWS. Bresnan, Joan 1982 Control and complementation. In Joan Bresnan (ed.), The mental representation of grammatical relations, 282-390. Cambridge, MA: MIT Press. Brown, Roger 1973 A first language: The early stages. Cambridge, MA: Harvard University Press. Brown, Roger and Ursula Bellugi 1964 Three processes in the child's acquisition of syntax. In E. Lenneberg (ed.), New directions in the study of language. Cambridge MA: MIT Press. Burani, Cristina and Alfonso Caramazza 1987 Representation and processing of derived words. Language and Cognitive Processes 2: 217-227'. Burani, Cristina and Alessandro Laudanna 1992 Units of representation for derived words in the lexicon. In R. Frost and L. Kats (eds.), Orthography, phonology, morphology, and meaning, 361-376. Amsterdam: Elsevier. Bybee, Joan L. 1985 Morphology: A study of the relation between meaning and form. Amsterdam: John Benjamins. 1988 Morphology as lexical organization. In Michael Hammond and Michael Noonan (eds.), Theoretical morphology. Approaches in modern linguistics, 119-141. New York: Academic Press.

284

References 1995a

Regular morphology and the lexicon. Language and Cognitive Processes 10: 425-455. 1995b Diachronie and typological properties of morphology and their implications for representation. In Laurie Beth Feldman (ed.), Morphological aspects of language processing, 225-246. Hillsdale, NJ: Lawrence Erlbaum. Bybee, Joan, Revere Perkins and William Pagliuca 1994 The evolution of grammar. Tense, aspect, and modality in the languages of the world. Chicago: University of Chicago Press. Bybee, Joan L. and Dan I. Slobin 1982 Rules and Schemas in the development and use of English past tense. Language 58: 265-289. Caramazza, Alfonso, Alessandro Laudanna and Cristina Romani 1992 Lexical access and inflectional morphology. Cognition 28: 297332. Carstairs-McCarthy, Andrew 1992 Current morphology. London: Routledge. Chafe, Wallace 1994 Discourse, consciousness, and time: The flow and displacement of conscious experience in speaking and writing. Chicago: University of Chicago Press. Cho, Euiyon 1991 Notes on some tests for subjecthood in Korean. Studies in the Linguistic Sciences 21.2: 19-29. Chomsky, Noam 1986 Knowledge of language. Its nature, origin, and use. New York: Praeger. Collins, A. and M. Quillian 1969 Retrieval time from semantic memory. Journal of Verbal Learning and Verbal Behavior 8: 240-247. Cook, Kenneth W. 1988 A cognitive analysis of grammatical relations, case, and transitivity in Samoan. San Diego: University of California doctoral dissertation. Creider, Chet and Richard Hudson 1999 Inflectional morphology in word grammar. Lingua 107: 163-187.

References

285

Croft, William 1990 Typology and universals. Cambridge: Cambridge University Press. Cutler, Anne 1980 Productivity in word formation. Papers from the Regional Meeting of the Chicago Linguistic Society 16: 45-51. Cutrer, Michelle 1994 Time and tense in narrative and in everyday language. San Diego: University of California doctoral dissertation. Dahl, Osten and Maria Koptjevskaja-Tamm 1992 Language typology around the Baltic Sea: A problem inventory. Stockholm: Stockholm University. Dalrymple, Mary 1993 The syntax of anaphoric binding. Stanford, CA: CSLI. Davis, Garry W. 1992 What's in a word? On integrating recent approaches to secondary associations, submorphemic units and morphological segmentation. Word 43: 197-216. De Mauro, Tullio, Federico Mancini, Massimo Vedovelli and Miriam Voghera 1993 Lessico difrequenza dell 'italiano parlato. Milano: Etas. Derwing, Bruce L. and WJ. Baker 1986 Assessing morphological development. In Paul Fletcher and Michael Garman (eds.), Language acquisition. Studies in first language development. [2nd ed.] Cambridge: Cambridge University Press. Derwing, Bruce L. and Royal Skousen 1989 Morphology in the mental lexicon: A new look at analogy. Yearbook of morphology 2: 55-71. 1994 Productivity and the English past tense. Testing Skousen's analogy model. In D. Lima, R.L. Corrigan and G.K. Iverson (eds.), The reality of linguistic rules, 193-218. Amsterdam: John Benjamins. Dimitrova-Vulchanova, Mila 1996 Verb semantics, diathesis and aspect. Trondheim: Norwegian University of Science and Technology doctoral dissertation.

286

References

Dressler, Wolfgang Ullrich 1989 Prototypical differences between inflection and derivation. Zeitschrift für Phonetik, Sprachwissenschaft und Kommunikationsforschung 42: 3-10. Dyvik, Helge 1980 Har gammelnorsk passiv? [Does Old-Norwegian have passive?] In Even Hovdhaugen (ed.), The Nordic languages and modern linguistics 4, 81-107. Oslo: Universitetsforlaget. Elman, Jeffrey L. 1990 Finding structure in time. Cognitive Science 14: 179-211. Emmory, Karen and Victoria Fromkin 1988 The mental lexicon. In F. Newmeyer (ed.), Linguistics. The Cambridge survey, vol 3, 124-149. Cambridge: Cambridge University Press. Endresen, Rolf Theil 1982 Morfologi [Morphology]. In Harald Bjorvand, Even Hovdhaugen and Hanne Gram Simonsen (eds.), Sprakvitenskap. En elementcer innf0ring, 87-126. [Linguistics. An elementary introduction]. [2nd ed.] Oslo: Universitetsforlaget. 1991 Fonetikk og fonologi. Ei elementar innf0ring [Phonetics and phonology. An elementary introduction]. Oslo: Universitetsforlaget. 1992 Norske verb og kognitiv morfologi [Norwegian verbs and cognitive morphology]. Oslo: Unpublished ms., Department of Linguistics, University of Oslo. 1994 The etymological relationship between the benefactive and causative suffixes in Fula. Trondheim: Trondheim Working Papers in Lingustics 22: 89-100. Endresen, Rolf Theil, Hanne Gram Simonsen and Andreas Sveen 1996 Innf0ring i lingvistikk [Introduction to linguistics]. Oslo: Universitetsforlaget. Enger, Hans-Olav 1998a The classification of strong verbs in Norwegian with special reference to the Oslo dialect. (Acta Humaniora 26). Oslo: Universitetsforlaget.

References 1998b

287

Forholdet mellom presens -form og Infinitiv s-form [The relationship between present -form and infinitival s-form]. Maal og Minne 2 (1998): 193-206. 2000 Verbendelsen -s i norsk bokmäl: B0ying eller avledning? [The verbal ending -s in Bokmäl: Inflectional or derivational?] Norsk Lingvistisk Tidsskrift 18: 9-36. Enger, Hans-Olav and Tore Nesset 1999 The value of cognitive grammar in typological studies: The case of Norwegian and Russian passive, middle and reflexive. Nordic Journal of Linguistics 22: 27-60. Engh, Jan [1982] 1996 On control. Norskrift 89: 19-48. Faarlund, Jan Terje, Svein Lie and Kjell Ivar Vannebo 1997 Norsk referansegrammatikk [Norwegian Reference Grammar]. Oslo: Universitetsforlaget. Fagerli, Ole Torfinn 1994 Verbal derivations in Fulfulde. Trondheim: University of Trondheim cand.philol thesis. [Trondheim Working Papers in Lingustics 21]. 1995 Benefactive constructions: A typological study. Trondheim Working Papers in Linguistics 25: 100-118. Fauconnier, Gilles and Eve Sweetser (eds.) 1996 Spaces, worlds, and grammar. Chicago: University of Chicago Press. Fodor, Jerry 1983 The modularity of mind. An essay on faculty psychology. Cambridge, MA: MIT Press. Forrest, Linda B. 1996 Discourse goals and attentional processes in sentence production: The dynamic construal of events. In Adele E. Goldberg (ed.), Conceptual structure, discourse and language, 149-161. Stanford: CSLI.

288

References

Frauenfelder, Uli H. and Robert Schreuder 1992 Constraining psycholinguistic models of morphological processing and representation: The role of productivity. In Geert Booij and Jaap van Marie (eds.), Yearbook of morphology 1991, 165-183. Dordrecht: Kluwer. Fromkin, Victoria 1988 Grammatical aspects of speech errors. In F. Newmeyer (ed.), Linguistics. The Cambridge survey, vol 2, 117-138. Cambridge: Cambridge University Press. Fromkin, Victoria and Robert Rodman 1988 An introduction to language. [4th ed.] New York: Holt, Rinehart and Winston. Geniusiene, Emma 1987 The typology of reflexives. Berlin: Mouton de Gruyter. Givon, Talmy 1984 Syntax: A functional-typological introduction, vol. 1. Amsterdam: John Benjamins. Goldberg, Adele E. 1995 Constructions. A construction grammar approach to argument structure. Chicago: University of Chicago Press. Goldberg, Adele E. (ed.) 1996 Conceptual structure, discourse and language. Stanford: CSLI. Gopnik, Myrna 1994 Impairment of tense in a familial language disorder. Journal of Neurolinguistics 8: 109-133. Greenberg, Joseph H. 1963 Some universale of grammar with particular reference to the order of meaningful elements. In Joseph H. Greenberg (ed.), Universals of language, 73-113. Cambridge, MA: MIT Press. Gundersen, Helge 1995 Linjedansere og pantomine pa sirkhus. Folkeetymologi som morfologisk omtolking [Line dancers and pantomine at the circhouse. Folk etymology as morphological reanalysis]. Oslo: Novus.

References 1997

Hagen, J. E. 1994

289

Objektivisme i lingvistikken [Objectivism in linguistics]. In Anita Leirfall and Thor Sandmel (eds.), Festskrift til Johan Arnt Myrstad i anledning 50-arsdagen [Festschrift for J. A. M. on the occasion of his 50th birthday], 147-181. Trondheim: Tapir. Svake verb pä -Vde og -Vie i bokmäl og deres paradigmetilh0righet [Weak verbs ending in -Vde and -Vte in Bokmäl and their paradigm membership]. In J. Myking, H. Sand0y and I. Utne (eds.), Helsing til Lars Vassenden pä 70-arsdagen [Greeting for L.V. on his 70th birthday], 57-67. Bergen: Nordisk institutt.

Haiman, John 1983 Iconic and economic motivation. Language 59: 781-819. 1985 Natural syntax. Cambridge: Cambridge University Press. Halle, Morris 1973 Prolegomena to a theory of word formation. Linguistic Inquiry 4: 316. Halle, Morris and K.P. Mohanan 1985 Segmental phonology of Modern English. Linguistic Inquiry 16: 57-116. Halliday, Michael 1985 An introduction to functional grammar. London: Arnold. Hare, M., J.L. Elman and K.G. Daugherty 1995 Default generalisation in connectionist networks. Language and Cognitive Processes 10: 601-630. Haspelmath, Martin 1993 The diachronic externalization of inflection. Linguistics 31: 279309. Heine, Bernd, Ulrike Claudi and Frederike Hünnemeier 1991 Grammaticalization. A conceptual framework. Chicago: University of Chicago Press. Hellan, Lars 1988 Anaphora in Norwegian and the theory of grammar. Dordrecht: Foris. Hestvik, Arild 1992 LF movement of pronouns and antisubject orientation. Linguistic Inquiry 23: 557-594.

290

References

Hockett, Charles F. 1954 Two models of grammatical description. Word 10: 210-234. 1958 A course in modern linguistics. New York: Macmillan. van Hoek, Karen 1995 Conceptual reference points: A cognitive grammar account of pronominal anaphora constraints. Language 71: 310-340. 1996 A cognitive grammar account of bound anaphora. In Eugene H. Casad (ed.), Cognitive linguistics in the Redwoods: The expansion of a new paradigm in linguistics, 753-791. Berlin: Mouton de Gruyter. 1997a Anaphora and conceptual structure. Chicago: University of Chicago Press. 1997b Backwards anaphora as a constructional category. Functions of Language 4: 47-82. to appear Quantifier scope: A cognitive grammar perspective. Holm, John A. 1988 Pidgins and Creoles, vol 1. Theory and Structure. Cambridge, UK: Cambridge University Press. Hudson, Richard 1971 English complex sentences. An introduction to systemic grammar. Amsterdam: North Holland. 1976 Arguments for a non-transformational grammar. Chicago: University of Chicago Press. 1984 Word grammar. Oxford: Blackwell. 1990 English word grammar. Oxford: Blackwell. 1995 Word meaning. London: Routledge. 1996 Sociolinguistics. [2nd ed.] Cambridge: Cambridge University Press. 1998 English grammar. London: Routledge. 2000 */ amn 't. Language: 297-323. Isacenko, Aleksandr V. 1982 Die russische Sprache der Gegenwart. Formenlehre. Munich: Max Hueber Verlag. Jackendoff, Ray 1975 Morphological and semantic regularities in the lexicon. Language 51: 639-671.

References 1997

291

The architecture of the language faculty. Cambridge, MA: MIT Press. Jaeger, Jeri T., A.H. Lockwood, D.L. Kemmerer, R.D. Van Valin, B.W. Murphy and H.G. Khalak 1996 A positron emission tomographic study of regular and irregular verb morphology in English. Language 72: 451-497. Janda, Laura A. 1993 Cognitive linguistics as a continuation of the Jakobsonian tradition. The semantics of Russian and Czech reflexives. In Robert A. Maguire and Alan Timberlake (eds.), American contributions to the eleventh international congress of slavists, 310-319. Columbus, Ohio: Slavica. 1997 GIVE, HAVE, and TAKE in Slavic. In John Newman (ed.), The linguistics of giving, 249-265. Amsterdam: John Benjamins. Jensen, John T. 1990 Morphology: Word structure in generative grammar. Amsterdam: John Benjamins. Johannessen, Ole-J0rgen 1993 Ordlaging. In Jarle R0nhovd (ed.), Norsk morfologi, 125-153. Oslo: Ad No tarn Gyldendal. Kay, Paul and Charles J. Fillmore 1999 Grammatical constructions and linguistic generalizations: The What'sXdoing Y? construction. Language 75: 1-33. Keenan, Edward L. and Bernard Comrie 1977 Noun phrase accessibility and universal grammar. Linguistic Inquiry 8: 63-99. Kemmer, Suzanne 1993 The middle voice. Amsterdam: John Benjamins. Kirchner, D. M. and R. L. Klatzky 1985 Verbal rehearsal and memory in language-disordered children. Journal of Speech and Hearing Research 28: 556-565. Kristoffersen, Kristian Emil 1996 Infinitival phrases in Old Norse: Aspects of their syntax and semantics. Oslo: University of Oslo dr.art thesis.

292

References

Kumashiro, Toshiyuki in prep. The conceptual basis of grammar: A cognitive approach to Japanese clause structure. San Diego: University of California doctoral dissertation. Kuroda, S.-Y. 1972 The categorical and the thetic judgment. Foundations of Language 9: 153-185. Kutsch Lojenga, Constance 1997 Verbal derivation in Ngiti and Lendu. Handout for paper presented at the 2nd World Congress of African Linguistics, Leipzig. Labov, William 1972a Language in the inner city. Philadelphia: University of Pennsylvania Press. 1972b Sociolinguistic patterns. Philadelphia: University of Pennsylvania Press. 1994 Principles of linguistic change, vol 1. Internal factors. Oxford: Blackwell. Lakoff, George 1987 Women, fire and dangerous things. What categories reveal about the mind. Chicago: University of Chicago Press. Lakoff, George and Mark Johnson 1980 Metaphors we live by. Chicago: University of Chicago Press. Lamb, Sidney 1966 Outline of stratificational grammar. Washington, DC: Georgetown University Press. Lambrecht, Knud 1987 Sentence focus, information structure, and the thetic-categorical distinction. Proceedings of the Annual Meeting of the Berkeley Linguistics Society 13, 366-382. 1994 Information structure and sentence form: Topic, focus, and the mental representations of discourse referents. Cambridge: Cambridge University Press. 1996 On the formal and functional relationship between topics and vocatives. Evidence from French. In Adele E. Goldberg (ed.), Conceptual structure, discourse and language, 267-288. Stanford: CSLI.

References

293

Langacker, Ronald W. 1986 Abstract motion. Proceedings of the annual meeting of the Berkeley Linguistics Society 12: 455-471. 1987a Foundations of cognitive grammar, vol. 1. Theoretical prerequisites. Stanford, CA: Stanford University Press. 1987b Nouns and verbs. Language 63: 53-94. 1990 Concept, image and symbol. The cognitive basis of grammar. Berlin: Mouton de Gruyter. 1991 Foundations of cognitive grammar, vol. 2. Descriptive application. Stanford, CA: Stanford University Press. 1993 Reference-point constructions. Cognitive Linguistics 4: 1-38. 1994 Cognitive Grammar. In R. Asher (ed.), Encyclopedia of language and linguistics, 590-593. Oxford: Pergamon. 1995a Conceptual grouping and constituency in cognitive grammar. In Linguistic Society of Korea (ed.), Linguistics in the morning calm 3, 149-172. Seoul: Hanshin. 1995b Possession and possessive constructions. In John R. Taylor and Robert E. MacLaury (eds.), Language and the cognitive construal of the world, 51-79. Berlin: Mouton de Gruyter. 1997a Constituency, dependency, and conceptual grouping. Cognitive Linguistics 8: 1-32. 1997b A dynamic account of grammatical function. In Joan L. Bybee, John Haiman and Sandra A. Thompson (eds.), Essays on language function and language type dedicated to T. Givon, 249-273. Amsterdam: John Benjamins. 1999a Double-subject constructions. In Sung-Yun Bak (ed.), Linguistics in the morning calm 4, 83-104. Seoul: Hanshin. 1999b Virtual reality. Studies in the Linguistic Sciences 29.2: 77-103 Larson, Richard K. 1991 Promise and the theory of control. Linguistic Inquiry 22: 103-139. Lass, Roger 1980 On explaining language change. Cambridge: Cambridge University Press. Leira, Vigleik 1992 Ordlaging og ordelement i norsk [Morphology and morphemes in Norwegian]. Oslo: Det Norske Samlaget.

294

References

Leiss, Elisabeth 1997 Synkretismen und Natürlichkeit. Folia Linguistica 31: 133-161. van der Lely, Heather K.J. and Michael T. Ullman in press Past tense morphology in specifically language impaired and normally developing children. Language and Cognitive Processes. Leonard, Laurence B. 1998 Children with specific language impairment. Cambridge, MA: MIT Press Levander, Lars 1928 Dalmalet. Beskrivning och historia II [The Dal dialect. Description and history]. Uppsala: Appelbergs. Li, Charles N. and Sandra A. Thompson 1976 Subject and topic: A new typology of language. In Charles N. Li (ed.), Subject and topic, 457-489. New York: Academic Press. Lieber, Rochelle 1989 On percolation. Yearbook of morphology 2: 95-138. Lobben, Marit 1998 Force as a metaphor in grammar, with special emphasis on the Hausa verbal marking of causative and benefactive. Paper presented at The Verb in Cognitive Grammar, Gran, Norway. Lord, Carol 1993 Historical change in serial verb constructions. Amsterdam: John Benjamins. Liibben, August 1993 Mittelniederdeutsches Handwörterbuch. Facsimile. Darmstadt: Wissenschaftliche Buchgesellschaft. Luger, George and William Stubblefield 1993 Artificial intelligence. Structures and strategies for complex problem solving. [2nd ed.] New York: Benjamin/Cummings. Lyons, John 1977 Semantics, vol. 2. Cambridge: Cambridge University Press. MacWhinney, Brian 1977 Starting points. Language 53: 152-168. MacWhinney, B. and J. Leinbach 1991 Implementations are not conceptualizations: Revising the verb learning model. Cognition 40: 121-157.

References

295

Maldonado, Ricardo 1988 Energetic reflexives in Spanish. Proceedings of the annual meeting of the Berkeley Linguistics Society 14: 153-165. 1993 Dynamic construals in Spanish. Studi haliani di Linguistica Teorica e Applicata 22: 531-566. Manney, Linda 1990 Mental experience verbs in Modern Greek: A cognitive explanation for active versus middle voice. Proceedings of the annual meeting of the Berkeley Linguistics Society 16: 229-240. 1992 States, results, and the active/passive/middle continuum in Modern Greek. Journal of Modern Greek Studies 10: 205-233. 1995 Pragmatic motivation for inflectional middle voice in Modern Greek. Functions of Language 2: 159-188. 1998 The reflexive archetype and its various realizations in Modern Greek. Studies in Language 22: 1-48. Marantz, Alec 1992 Review of Jensen 1990. Language 68: 413-416. Marchman, Virginia A. 1997 Children's productivity in the English past tense: The role of frequency, phonology and neighborhood structure. Cognitive Science 21: 283-304 Marcus, Gary F., Ursula Brinkmann, Harald Clahsen, Richard Wiese and Steven Pinker 1995 German inflection: The exception that proves the rule. Cognitive Psychology 29: 189-256. Marcus, Gary F., Steven Pinker, Michael Ullman, Michelle Hollander, T. John Rosen and Fei Xu 1992 Overregularisation in language acquisition. Monographs of the society for research in child development 57.4, serial No. 228. van Marie, Jaap 1996 The unity of morphology. In Geert Booij and Jaap van Marie (eds.), Yearbook of morphology 1995, 67-82. Dordrecht: Kluwer. Marslen-Wilson, William 1980 Speech understanding as a psychological process. In J.C. Simon (ed.), Spoken language generation and understanding. Boston: Reidel.

296

References

Marslen-Wilson, William and Lorraine Tyler 1980 The temporal structure of spoken language understanding. Cognition 18: 1-71. 1981 Central processes in speech understanding. Philosophical Transactions of the Royal Society of London B295: 317-332. Martin, James 1992 Systemic grammar. In W. Bright (ed.), International encyclopedia of linguistics, vol 4, 120-122. Oxford: Oxford University Press. Matcovich, Paola F. 1998 The acquisition and processing of past participle morphology in Italian. An experimental study. Oslo: University of Oslo cand. philol thesis. Matsumoto, Yo 1996a How abstract is subjective motion? A comparison of coverage path expressions and access path expressions. In Adele E. Goldberg (ed.), Conceptual structure, discourse and language, 359-373. Stanford: CSLI. 1996b Subjective motion and English and Japanese verbs. Cognitive Linguistics!: 183-226. 1996c Subjective-change expressions in Japanese and their cognitive and linguistic bases. In Gilles Fauconnier and Eve Sweetser (eds.), Spaces, worlds, and grammarl24-l56. Chicago: University of Chicago Press. Matthews, P. H. 1972 Inflectional morphology. Cambridge: Cambridge University Press. 1991 Morphology. Cambridge: Cambridge University Press. 1993 Grammatical theory in the United States from Bloomfield to Chomsky. Cambridge: Cambridge University Press. McCarthy, John J. 1982 Prosodic templates, morphemic templates, and morphemic tiers. In Harry van der Hülst and Norval Smith (eds.), The structure of phonological representations, part 1,191-223. Dordrecht: Foris. Michaelis, Laura A. and Knud Lambrecht 1996 Toward a construction-based theory of language function: The case of nominal extraposition. Language 72: 215-247.

References

297

Naes, Oddvar 1972 Norsk grammatikk. Elementare strukturer og syntaks [Norwegian Grammar. Basic structures and syntax]. Oslo: Fabritius. Nesset, Tore 1994 Russian stress: Stress as an inflectional formative in Russian noun paradigms and Bybee 's cognitive morphology. Oslo: Novus. 1998a Affiks eller klitikon? Om elementet -sja i moderne standard russisk [Affix or clitic? On the element -sja i modern Standard Russian]. Norsk Lingvistisk Tidsskrift 16: 185-206. 1998b Russian conjugation revisited: A cognitive approach to aspects of Russian verb inflection. Oslo: Novus. Newman, John (ed.) 1996 Give: A cognitive linguistic study. Berlin: Mouton de Gruyter. 1997 The linguistics of giving. Amsterdam: John Benjamins. Nichols, Johanna, Gilbert Rappaport and Alan Timberlake 1980 Subject, topic, and control in Russian. Proceedings of the Annual Meeting of the Berkeley Linguistics Society 6: 372-386. Nida, Eugene 1949 Morphology: The descriptive analysis of words. [2nd ed.] Ann Arbor: University of Michigan Press. N0klestad, Anders 1996 A connectionist modelling approach to past tense acquisition in Norwegian. Oslo: University of Oslo cand.philol thesis. Noreen, A. 1923 Altnordische Grammatik. I. 5., unveränderte Auflage. Tübingen: Max Niemeyer. [Reprinted 1970.] Noye, Dominique 1989 Dictionnaire Foulfoulde-Fra^ais. Dialecte peul du Diamare, Nord-Cameroun. Paris: Librairie Orientaliste Paul Geuthner. Nuyts, Jan and Eric Pederson (eds.) 1997 Language and conceptualization. Cambridge: Cambridge University Press. van Oosten, Jeanne 1977 Subjects and agenthood in English. Papers from the Regional Meeting of the Chicago Linguistic Society 13: 459-471.

298

References 1986

The nature of subjects, topics and agents: A cognitive explanation. Bloomington: Indiana University Linguistics Club. Orsolini, Margherita and William D. Marslen-Wilson 1997 Universals in morphological representation: Evidence from Italian. Language and Cognitive Processes 72: 1-47. The Oslo corpus of tagged Norwegian texts http://www.tekstlab.uio.no/norsk/bokmaal/english.html Osam, E.K. 1994 From serial verbs to prepositions and the road between. Sprachtypologie und Universalienforschung 47: 16-36. Osthoff, Hermann 1878 Formassociation bei Zahlwörtern. In Hermann Osthoff and Karl Brugmann (eds.), Morphologische Untersuchungen auf dem Gebiete der indogermanischen Sprachen, 92-132. Leipzig: Hirzel. Ottosson, Kjartan G. 1992 The Icelandic middle voice: The morphological and phonological development. Lund: Department of Scandinavian Languages. Pederson, Eric William 1991 Subtle semantics: Universals in the polysemy of reflexive and causative constructions. Berkeley: University of California Ph.D. dissertation. Perlmutter, David M. (ed.) 1983 Studies in relational grammar, vol. 1. Chicago: University of Chicago Press. PICA 1962 Proceedings of the Institute on Childhood Aphasia. San Francisco: National Society for Crippled Children and Adults. Pike, Kenneth L. 1970 Tagmemic and matrix grammar applied to selected African languages. Summer Institute of Linguistics: Publications in Linguistics and Related Fields, no. 23. Pinker, Steven 1994 The language instinct. Harmondsworth, Middlesex: Penguin. 1997 How the mind works. Harmondsworth, Middlesex: Penguin. 1999 Words and rules. London: Weidenfeld and Nicolson.

References

299

Pinker, Steven and Alan Prince 1988 On language and connectionism: Analysis of a parallel distributed processing model of language acquisition. Cognition 28: 73-193. Plank, Frans 1981 Morphologische (Ir-)Regularitäten. Aspekte der Wortstrukturtheorie. Tübingen: Narr. Plunkett, Kim and Virginia A. Marchman 1991 U-shaped learning and frequency effects in a multi-layered perceptron: Implications for child language acquisition. Cognition 38: 43-102. 1993 From rote learning to system building: Acquiring verb morphology in children and connectionist nets. Cognition 48: 21-69. 1996 Learning from a connectionist model of the English past tense. Cognition 61: 299-308. Pollard, Carl and Ivan A. Sag 1994 Head-driven phrase structure grammar. Chicago: University of Chicago Press. Postal, Paul M. 1969 Anaphoric islands. Papers from the Regional Meeting of the Chicago Linguistic Society 5: 205-239. Prasada, Sandeep and Steven Pinker 1993 Generalisation of regular and irregular morphological patterns. Language and Cognitive Processes 8: 1-56. Prideaux, Gary D. 1984 Psycholinguistics: The experimental study of language. London: Croom Helm. Ragnarsdottir, Hrafnhildur, Hanne Gram Simonsen and Kim Plunkett 1997 The acquisition of past tense inflection in Icelandic and Norwegian children. In Eve Clark (ed.), Proceedings of the 28"' Annual Child Language Research Forum, 259-270. Stanford: CSLI. 1999 The acquisition of past tense morphology in Icelandic and Norwegian children: An experimental study. Journal of Child Language 26: 577-618. Rainer, Franz 1996 Inflection inside derivation. In Geert Booij and Jaap van Marie (eds.), Yearbook of morphology 1995, 83-91. Dordrecht: Kluwer.

300

References

Reisberg, Daniel 1997 Cognition. Exploring the science of the mind. New York: Norton. Rhodes, Richard 1994 Aural images. In Leanne Hinton, Johanna Nichols and John J. Ohala (eds.), Sound symbolism, 276-292. Cambridge: Cambridge University Press. Rhodes, Richard A. and John M. Lawler 1981 Athematic metaphors. Papers from the Regional Meeting of the Chicago Linguistic Society 17: 318-342. Rumelhart, D.E. and J.L. McClelland 1986 On learning the past tense of English verbs. In J. L. McClelland, D. E. Rumelhart and the PDF Research Group (eds.), Parallel distributed processing: Explorations in the micro structure of cognition, vol. 2. Psychological and biological models. Cambridge, MA: MIT Press. Runde, Älov 1997 En semantisk-pragmatisk beskrivelse av noen norske kontrollverb [A semantic-pragmatic description of some Norwegian control verbs]. Oslo: University of Oslo cand.philol thesis. Sandra, Dominiek and Sally Rice 1995 Network analyses of prepositional meaning: Mirroring whose mind - the linguist's or the language user's? Cognitive Linguistics 6: 89130. Saether0, Eli 1997 'Minimal' verbs in Akan. A study of the contribution of the verb and construction in meaning. Trondheim: University of Trondheim cand.philol thesis. Say, Tessa 1997 Regular and irregular inflection in the mental lexicon: Evidence from Italian. Essex Graduate Students Papers in Linguistics 2: 5570. Schoorlemmer, Maaike 1997 Russian sja and the affix-clitic distinction. In Martina Lindseth and Steven Franks (eds.), Annual workshop on formal approaches to Slavic linguistics. The Indiana meeting, 7996, 253-274. Ann Arbor: Michigan Slavic Publications.

References

301

Shibatani, Masayoshi 1985 Passives and related constructions: A prototype analysis. Language 61: 821-848. Simonsen, Hanne Gram 1993 Norwegian Bokmäl: Strong and weak verbs. Oslo: Unpublished ms., Department of Linguistics, University Oslo. Simonsen, Hanne Gram and Kirsten Meyer Bjerkan 1998 Testing past tense inflection in Norwegian: A diagnostic tool for identifying SLI children? International Journal of Applied Linguistics 8: 251-270. Skousen, Royal 1989 Analogical modeling of language. Dordrecht: Kluwer. 1992 Analogy and structure. Dordrecht: Kluwer. Söderbergh, Ragnhild 1967 Svensk ordbildning [Swedish morphology.]. Stockholm: Svenska B okför laget/Nor stedts. Steele, Susan M. 1977 On being possessed. Proceedings of the Annual Meeting of the Berkeley Linguistics Society 3: 114-131. van Stek, Rob 1983 Enkelte aspekter ved suffiks-avledning i norsk [Some aspects of suffixal derivation in Norwegian]. Studier i skandinavistik 1: 45-75. Stump, Gregory T. 1997 Template morphology and inflectional morphology. In Geert Booij and Jaap van Marie (eds.), Yearbook of morphology 1996, 217-241. Dordrecht: Kluwer. Sweetser, Eve 1997 Role and individual interpretations of change predicates. In Jan Nuyts and Eric Pederson (eds.), Language and conceptualization, 116-136. Cambridge: Cambridge University Press. Sweetser, Eve and Gilles Fauconnier 1996 Cognitive links and domains: Basic aspects of mental space theory. In Gilles Fauconnier and Eve Sweetser (eds.), Spaces, worlds, and grammar, 1-28. Chicago: University of Chicago Press.

302

References

Swinney, D. 1979

Sylla, Yero 1979

Lexical access during sentence comprehension: (Re-)consideration of context effects. Journal of Verbal Learning and Verbal Behavior 5: 219-227.

Grammatical relations and Fula syntax. Ph.D. dissertation, University Microfilms International, London: Ann Arbor. Talmy, Leonard 1988 Force dynamics in language and cognition. Cognitive Science 12: 49-100. 1996 Fictive motion in language and "ception". In Paul Bloom, et al. (eds.), Language and space, 211-276. Cambridge, MA: MIT Press. Taylor, F. W. 1932 A Fulani-English dictionary. Oxford: Clarendon Press. Taylor, John R. 1989 Linguistic categorisation: An essay in cognitive linguistics. Oxford: Oxford University Press. 1996 Possessives in English: An exploration in cognitive grammar. Oxford: Oxford University Press. Tesniere, Lucien 1959/66 Elements de syntaxe structurale. Paris: Klincksieck. Thorell, Olof 1981 Svensk ordbildningslära [A course in Swedish morphology.]. Stockholm: Esselte. Tlearn neural network simulator http://crl.ucsd.edu/innate/tlearn.html Tomblin, J. Bruce, Nancy L. Records, Paula Buckwalter, Xuyang Zhang, Elaine Smith and Marlea O'Brien 1997 Prevalence of specific language impairment in kindergarten children. Journal of Speech, Language, and Hearing Research 40: 1245-1260. Tomlin, Russell S. 1995 Focal attention, voice, and word order. In Pamela Downing and Michael Noonan (eds.), Word order in discourse, 517-554. Amsterdam: John Benjamins.

References 1997

303

Mapping conceptual representations into linguistic representations: The role of attention in grammar. In Jan Nuyts and Eric Pederson (eds.), Language and conceptualization, 162-189. Cambridge: Cambridge University Press. Usakov, Dmitrij Nikolaevic (ed.) 1935 Tolkovyj slovar' russkogo jazyka [Monolingual Dictionary of Russian]. Moscow: Sovetskaja enciklopedija. Vangsnes, 0ystein Alexander 1999 Substantivfraser: Om kapittel 3 i Norsk referansegrammatikk [Noun phrases: On chapter 3 of The Norwegian reference grammar]. Nordica Bergensia 21: 34-50. Verfaillie, Karl and Anja Daems 1996 The priority of the agent in visual event perception: On the cognitive basis of grammatical agent-patient asymmetries. Cognitive Linguistics 7: 131 -147. Western, August 1921 Norsk Riksmals-Grammatikk [Norwegian Riksmäl Grammar]. Oslo: Aschehoug. Wierzbicka, Anna 1988 The semantics of grammar. Amsterdam: John Benjamins. Williams, Adger 1993 The argument structure of / -predicates. Journal of Slavic Linguistics 1: 167-190. Williams, Edwin 1981 On the notions 'lexically related' and 'head of a word'. Linguistic Inquiry 12: 245-274. Winter, Werner 1969 Analogischer Sprachwandel und semantische Struktur. Folia Linguistica 3: 29-45. Wurzel, Wolfgang Ullrich 1984 Flexionsmorphologie und Natürlichkeit. Berlin: Akademie-Verlag. Yelland, G. W. 1994 Word recognition and lexical access. In R. Asher (ed.), Encyclopedia of language and linguistics, 5009-5014. Oxford: Pergamon.

304

References

Zwicky, Arnold 1994 Dealing out meaning: Fundamentals of syntactic constructions. Proceedings from the annual meeting of the Berkeley Linguistics Society 20:611-625.

Index Aasen, L, 86 Abelin, A., 110 absolutive, 231 access, 50 activity level, 171 adversative, 203 affectedness, 214 affectedness marker, 204 agent, 234 Akan, 211, 213,215,220-221 allative, 206,209 analogical reinterpretation, 101 Andenes.E. S.,211 applicative, 204,210 Arnott, D. W., 206 Aronoff, M., 102 assemblies of symbolic structures, 18 Atlantic, 204 Augmented Addressed Morphology Model, 199 Baker, W. J., 130 Bantu, 204,210 Bariba, 215 base, 19 Bauer, L., 102, 105, 106 Beito, Ο. Τ., 86 Bellugi.U., 130 benefactive, 203, 210, 213 benefactor, 206 beneficiary, 206, 227 Berko, J., 130, 134 Bjerkan, K. M., 92, 115, 130, 140-141, 167-168, 170,177-179, 181,183 blends, 105 Bloomfield, L., 97, 102, 109, 110, 112

Bodomo, A.B. ,211,214 Bokm l, 74,76,78, 83-85, 89, 90-91,93-94,132,168 Boretzky, N., 214 Breedveld, J. O., 206 Brown, R., 130 building-block metaphor, 96 building-block model, 95 Burani, C, 199 Bybee, J. L., 99-122, 129-131, 134, 155,159, 166, 171, 184,189, 192 Baayen, R. H., 168, 170 consonant class, 90 Caramazza, A., 199 Carstairs-McCarthy, A., 113 categorical judgement, 44 causation / purpose, 217 C-class, 86 Chafe, W., 41 Chamorro, 212 clause-internal discourse topic, 45 clipping, 105 clipping compounds, 105 cognitive grammar, 153 cognitive morphology, 99 collective, 228, 232 completedness, 217 composite symbolic structure, 22 compounding, 105 Comrie, B., 49 concatenator, 118 conceptual reification, 38 conceptualist semantics, 18 conceptualizer, 19 conceptually dependent, 41 conjugation classes, 190 Conjugation I, 190, 193, 195

306

Index

Conjugation II, 190,193,196 Conjugation III, 190 connectionist models, 166,170, 189 connectionist network simulation, 146 consonantal conjugation, 86 constituency, 23 construal, 18 constructional schema, 23 control phenomena, 13 Cook,K.W.,51 correspondence, 22 cranberry morphs, 101 Creole, 204 Cutler, A., 106 Cutrer, M., 52 Czech, 218 Daems, A., 34 Dagaare, 211,214 Dalrymple, M., 223 Danish, 115 dative, case 218 dative event, 206 De Mauro, T., 194 Derwing, B. L., 130, 131, 199 destinative, 210 dominion, 24 double-subject constructions, 15 dual mechanism account, 130 dual mechanism model, 147, 189, 199 Dutch, 115 dynamic, 24 elaboration, 18 Elman, J. L., 52 emergence / manifestation, 217 enablement, 217 Endresen, R. T., 73, 86, 109, 115, 127,132-136, 155, 161,173, 221

energy flow, 205, 209,215 energy sink, 205, 209 energy source, 205,215 Enger, H.-O., 83, 90, 107 English, 26, 33,43-46,105,109, 111-116,126,130-134,150,165 entrenchment, 149, 162 Essegbey, J., 213 Ewe, 211,213 experiencer, 218 extension of sense, 204 extension, 232 Fagerli, O. T., 206 Fauconnier, G., 52 fictive motion, 50 finding things in space, 40 focal prominence, 21, 31 folk etymology, 101 force-dynamics, 205, 209 Forrest, L. B., 33 frame, 25 Frauenfelder, U. H., 199 French, 13,14,116 frequency, 135 Fromkin, V., 97 Fula, 210, 213,219,221 Faarlund, J. T., 86,94, 109,116 generative analysis, 104 generative grammar, 106 generative linguistics, 96 GeniuSiene, E., 226 German, 111,113,115-117,218 Germanic, 133 giver, 204 Givon, T., 12 global productivity, 168 goal, 205 Gopnik, M., 162 grammatical class, 21 grammatical construction, 22

Index grammaticalisation, 81,212 Greenberg, J. H., 34 Gundersen, H., 91, 95,101, 117 Gur languages, 204

Japanese, 17 Jensen, J. T., 97, 99 Johannessen, O.J., 116 Johnson, M, 95, 96

Hagen, J. E., 92, 170 Halle, M., 97 Harris, Z. S., 108 Hausa, 221 head, 23 heavy markers 224 Hellan, L., 223 Hestvik, A., 223 Hockett, C. F., 102, 107 van Hoek, K., 26, 35, 37, 52 Hokkien Chinese, 212 Holm, J. A., 211

Karaboro, 212 Keenan, E. L., 49 Kemmer, S., 223, 225 Korean, 15, 17 Kumashiro, T., 51 Kwa languages, 204

Icelandic, 149 iconic, 45 identification, 40 Indonesian, 14 infinitive, 139 inflection, 190 initial and secondary foci, 30 initial focus, 24 inner subject, 17 input factors, 134, 142, 146, 167 input layer, 171 instantiation, 234, 237 integrated, 22 interpersonal communication, 217 irregular forms, 190 irregular inflection, 131 irregular verbs, 83, 193,198 irregular weak verbs, 84 Italian, 146, 149,190 Jackendoff, R., 113 Jaeger, J. T., 130 Janda,L. A., 218, 223

307

Lakoff, G., 95, 96 Lambrecht, K., 14, 43,44,50 landmark, 21,30, 124 Langacker, R. W., 96, 100-111, 123-124, 153, 160, 166, 171, 189-190,204,224 larger weak (WL) class), 86, 90, 136, 147, 154, 160 Laudanna, Α., 199 Lawler.J. M., 109, 110 Leira,V., 115, 116 van der Lely, H. K. J., 162 Lendu, 221 Leonard, L. B., 153 level of specificity, 18 levels of organization, 42 lexical connections, 99 Li,C. N., 12, 14 Lie, S., 86, 109 Lieber, R., 97 light markers, 224 Lobben.M., 221 Lojenga, K., 221 Lord, C., 211,212 Low German, 113, 115, 116, 117 Luiseno, 15, 17 Liibben, A., 116 MacWhinney, B., 41

308

Index

Maldonado, R., 223 malefactive, 203,213 Mandarin, 212 Manney, L., 223 Marantz, A., 97 Marchman, V. A., 130, 131, 150 Marcus, G. F., 130-131, 166,183, 186, 189 Marslen-Wilson, W. D., 150, 199 Matcovich, P. F., 130,146,150 Matsumoto, Y., 52 Matthews,?. H.,97, 108, 111 McCarthy, J. J., 97,108, 113 McClelland, J. L., 130,131 mental contact, 24 mental lexicon, 99 mental scanning, 50 metaphor, 205 metaphorical extension, 206 metonymy, 26,40 middle, 234,235,238 morpheme-based structuralism, 97 morphemes as emergent, 111 morphemes as epiphenomenal, 111 morphological ambiguity, 101 Morphological Race Model, 199 morphological redundancy rules, 113 movement, 217 Naes, O., 86 Nesset, T., 98, 239 network, 17Ί, 234 network in the mental lexicon, 113 network models, 189 network training, 173 Newman, J., 204, 212,217-218, 220 Ngiti, 221 Nichols,!., 13 Nida, E., 97, 98, 99 Nida's principles, 98 N0klestad, A. 92, 130, 146, 167 nonce items, 195 nonconcatenative morphology, 122

non-processual relations, 21 Noreen, A., 86 Norwegian, 13,73- 86, 90- 95, 100-104,109-117,126,129136,140-149,167,223 Noye,D.,210 Nynorsk, 74,78, 83, 85,93-94, 104 108 object, 23 objectivist philosophy, 95 Old Norse, 86 van Oosten, J., 47 Orsolini, M., 150,199 Osam.E. K.,212 Osthoff, H., 112 outer subject, 17 output layer, 171 overgeneralisation, 131,165, 180 overgeneralisation errors, 141, 156, 185 overlapping, 105 overlapping items, 101, 104 passive, 234-235,238 past participle, 190 past tense, 130, 153, 165, 193 patient, 234 Pederson, E. W., 226,232 Perlmutter, D. M., 49 perspective, 18 phonestheme, 109,112 phonological coherence, 136, 144, 155 phonological difference, 136 phonological factors, 135 phonological openness, 136, 144, 148 phonological properties, 143, 192 phonological residues, 101 phonological similarity, 136, 145, 195, 199

Index phonological structure, 18 pidgin, 204 Pike, K. L., 215 Pinker, S., 131,165-166, 186, 189 pivot, 12 Plank,F., I l l Plunkett, K., 129-131, 134,150, 154, 166- 168, 170, 177, 182-183, 186-187 point of access, 35 poles, 18 Polish, 218 polysemy, 205,217 portmanteau words, 105 position verbs, 81 possession, 40 possessive constructions, 26 Postal, P. M., 52 Prasada, S., 131,165, 166,186, 189 present tense, 139 Prideaux, G. D., 106 primary and secondary clausal topic, 12 primary and secondary focal participants, 30 primary focal participant, 21 Prince, A., 131,166,186 process, 21 processing time, 24 product-oriented definition, 90 profile, 19 profile determinant, 22 profiling, 19 prominence, 13, 18 pronominal anaphora, 26,40 pronouns, 224 prototypical act of giving, 206 prototypical GIVE scene, 204 Pulaar, 220 radial category, 232

309

Ragnarsdottir, H., 129-130, 134, 136,149-150, 154,167-168, 170, 177,182-183, 186-187 Rappaport, G., 13 recipient / benefactive, 217 recipient, 204, 227 reciprocal, 228 redundancy rule, 113 reference point, 24, 30 reference point chain, 31, 36 reference tracking, 40 reflexive, 227,234-235, 238 reflexive-middle-passive systems, 223 regular forms, 190,198 regular inflection, 131 regular verbs, 83, 193 relationship, 20 residual morphs, 101 Rhodes, R. A., 109,110 Rodman, R., 97 Romani, C, 199 root-forming morphs, 110, 112 rule, 162 Rumelhart, D. E., 130, 131 Russian, 13,223 Saether0,E.,211,221 salience, 143, 149, 182 Saramaccan, 216 Say, T., 199 Scandinavian, 116 schema, 112,148,160,162, 171, 232 schematic interaction, 217 schematicity, 18 Schoorlemmer, M., 239 Schreuder, R., 199 secondary focus, 25 secondary focal participant, 21 segmentability, 182 semantic structure, 18 Senufo, 212

310

Index

serial verb constructions, 210 serializing languages, 203, 210 Shibatani, M., 49 Simonsen, H. G., 73, 86,92, 115, 127,132-135, 154-155, 161, 167-170,177-187 single mechanism account, 130 single mechanism models, 148, 186, 189,198 situation type, 224, 225,226 Siya, 212 Skousen, R., 131, 199 Slobin, D. I., 129, 130, 134, 155, 159 smaller weak (WS) class, 86, 90, 136,154 Söderbergh, R. 116 source-oriented, 90 South East Norwegian phonology, 74 specific language impairment, 153 spotlights of focal prominence, 31 Sranan,211,214 starting point, 41 Steele, S. M., 15 vanStek, R. 118 strong inflection, 131 strong verbs, 83, 86,136 structuralist linguistics, 96 subject, 23 subject honorification, 15 subjective motion, 19 submorphemic morphology, 106 submorphemic relations, 101 superheavy marker, 225 Surselvan, 225 Sveen, A., 161 Swahili, 210 Swedish, 110, 115-116 Sweetser, E., 52 Sylla, Y., 220 symbol, 160 symbolic, 18 symbolization, 40

Talmy, L., 52 target, 24, 30 Taylor, F. W., 206 Taylor, J. R., 27 telescoped words, 105 Thai, 212 thematic vowel, 190 thetic judgement, 44 thing, 20, 21,27, 204 Thompson, S. A., 12, 14 Thorell.O., 116 Timberlake, A., 13 tlearn, 173 token frequency, 135, 140, 145-146, 148-149,155, 173, 187,193, 199 Tomblin, J. B., 153 Tomlin, R. S., 33 tönernes, 76 topic construction, 28 trajector, 124, 21, 30 trajector/landmark alignment, 46 Twi, 213 type frequency, 90,135,143, 148, 155,191,193

Ullman, M. T., 162 unique morphs, 101 usage-based, 149 vowel class, 90 Vangsnes, 0. A., 109 Vannebo, K. I., 86, 109 V-class, 86 verbal extensions, 206 Verfaillie, K., 34 Vietnamese, 212 viewpoint shift, 50 vocabulary size, 176 vocabulary spurt, 173 vocalic conjugation, 86

Index weak inflection, 131 weak verbs, 83, 90 Western, A., 86,94 Williams, E., 97 Winter, W., 112 WL class, 86 word accents, 76, 78, 132

word tone, 132, 145, 149 WS class, 86

Yao Samsao, 212 Yoruba, 212,215

311