How we say WHEN it happens (Linguistische Arbeiten) (German Edition) 3484304553, 9783484304550

Citation preview

Linguistische Arbeiten

455

Herausgegeben von Hans Altmann, Peter Blumenthal, Hans Jürgen Heringer, Ingo Plag, Heinz Vater und Richard Wiese

How we say WHEN it happens Contributions to the theory of temporal reference in natural language Edited by Hans Kamp and Uwe Reyle

Max Niemeyer Verlag Tübingen 2002

Die Deutsche Bibliothek - CIP-Einheitsaufnahme How we say when it happens : contributions to the theory of temporal reference in natural language / ed. by Hans K a m p and Uwe Reyle. - Tübingen : Niemeyer, 2002 (Linguistische Arbeiten ; 455) ISBN 3-484-30455-3

ISSN 0344-6727

© Max Niemeyer Verlag G m b H , Tübingen 2002 Das Werk einschließlich aller seiner Teile ist urheberrechtlich geschützt. Jede Verwertung außerhalb der engen Grenzen des Urheberrechtsgesetzes ist ohne Zustimmung des Verlages unzulässig und strafbar. Das gilt insbesondere für Vervielfältigungen, Übersetzungen, Mikroverfilmungen und die Einspeicherung und Verarbeitung in elektronischen Systemen. Printed in Germany. Gedruckt auf alterungsbeständigem Papier. Druck: Weihert-Druck G m b H , Darmstadt Einband: Industriebuchbinderei Nadele, Nehren

Contents

Preface

Grzegorz Dogil Intonation of Aspectual Meaning: Remarks on noch in German

vii

1

Anette Frank and Annie Zaenen Tense in LFG : Syntax and Morphology

17

Hubert Haider Adverbials at the Syntax-Semantics Interface

53

Klaus Baumgärtner halten und die Einheit des verbalen Aspekts

71

Kurt Eberle Tense and Aspect Information in a FUDR-based German French Machine Translation System

97

Peter Blumenthal Zur Konstruktion französischer Wahrnehmungsverben: Aspekt und Abstraktheit

149

Uwe Reyle A Note on Enumerations and the Semantics of puis and alors

171

Hans Kamp and Michael Schiehlen Temporal Location in Natural Languages

181

Dov M. Gabbay Temporal Logic and Quantifier Translations

233

List of Contributors

280

Preface

The original impetus for putting together this collection was the 60-th birthday of Christian Rohrer. Yet - we must say this right away - this is not a "Festschrift" of the usual variety - it is not a liber amicorum, a collection of contributions whose first purpose it is to say various kinds of nice things to or about the central amicus. Such collections make charming and sometimes interesting historical documents, but as contributions to their own field their value tends to be limited. So we have aimed, from the very start, at something different. Since we wanted this volume to have, of all the usual attributes of a birthday present, at least that of surprise, we could not consult Christian Rohrer to ask him whether the result for which we were hoping was in his spirit. But we thought we knew him well enough to be confident that this was so. Our goal has been a collection of papers which, both individually and together, make a significant contribution to their subject matter and thus will be of interest to the scientific community at large, and not just to a select group of amici. Essential to this, it seemed to us, was a common topic, which would be broad enough to be of interest to all those who we hoped would be willing to contribute; which would be of interest to a significant number of scholars in the field; and which at the same time would be specific enough to give the enterprise a clear sense of focus. Finally, it was natural to look for a theme that has played a role of importance in Christian Rohrer's own intellectual career. Given these constraints the topic of tense, aspect and temporal reference seemed an obvious choice. It is a topic which continues to be in the centre of general scientific interest (witness the substantial numbers of contributions to it that are invariably submitted to international conferences in natural language semantics and appear in leading journals); it is specific enough to allow for significant connections between different contributions, even if these represent orthogonal perspectives or employ distinct methodologies; and it is a subject that has had Christian Rohrer's attention from the earliest beginnings of his career to the present day. It is in this last clause that lies hidden what we see as the essence of this collection, that which sets it apart from the many publications on tense and aspect that have appeared over the past two or three decades. For while Chistian Rohrer's interest in the topic of tense and aspect has continued, his general outlook on the study of language evolved dramatically, and in this his own development has been a reflection of a change that the past thirty years have seen in general. It has been within this period that computational linguistics became a field in its own right, rather than a tenuous symbiosis between linguistics, computer science and AI; and at least over the past decade, it has made an increasing impact on the views of many theoretical linguists about what the aims and methods of their field ought to be. This general development is reflected in particular by much recent work on questions of tense and aspect; some of that work addresses problems that have arisen within a computational setting; some of it deals with older problems from a computational perspective; at the same time we still see many studies which deal with the old problems that are

viii

still with us and which do so within a "classical" setting, in which the computational point of view does not play a dominant role. It it is not surprising that these various points of view, and the kind of work to which they give rise, are represented within the circle of those who we thought would be natural contributors to the present collection. For that circle overlaps with each of the different scientific milieus through which Christian Rohrer has moved in the course of his career: from more traditional to formal linguistics, then from formal to computational linguistics, and then, within computational linguistics, gradually from a predominantly theoretical stance to a position in which the demands of actual applications plays an increasingly important part. So it was possible to bring together a group of papers which reflect this plurality of perspective and methodology within a comparatively well-defined thematic domain. The result is a collection which bears witness both to how, in the course of the past three decades, linguistics and computational linguistics have developed, and to the diversity of perspectives and methodologies which are alive today. The plurality we see in the study of language today is nothing if not bewildering. Perhaps the thing that current work on language needs more than anything else is an integration of the methodologically and perspectivally varying approaches which, it appears are often being pursued without much awarenes of what is being done next door. This task, of mutual comparison and integration, is always difficult; but it is always needed and always worth trying. It is our hope that this collection, apart from containing papers that are worth reading on their own account, will make some contribution towards that ultimate goal. And we would like to think that if it does, it will have proved a worthwhile birthday present, too.

Hans Kamp Uwe Reyle

Grzegorz Dogil Intonation of Aspectual Meaning: Remarks on noch in German

1

Noch, noch, noch

Noch is among the most frequently used and the most flexible expressions in German. It has a wide variety of meanings, and it interacts with many grammatical subsystems like syntax, semantics and intonation. Noch (and its pragmatic partner schon) has received more attention in the linguistic literature in the last quarter century than any other German particle 1 (cf. Shetter 1966; Traugott / Waterhouse, 1969; Doherty 1973; König 1977, 1991; Abraham 1980; Hoepelman /Rohrer 1981; Löbner 1989). Focus particles, like other focusing expressions, establish a relation between the semantic value of the focused expression and a set of alternatives (Rooth 1985, 1992). Noch belongs to the 'additive' (inclusive) subclass, the subclass whose members imply the inclusion of some alternatives as values for the variable they bind. 2 In case of focal noch a scale (usually a temporal scale) of alternatives is 'evaluated' in the sense that the particle fixes the value 'relatively late' on this scale (cf. Hoepelman / Rohrer 1981:106; König 1991:141). These focal, additive and scalar uses of noch may be illustrated in the example in (1) from König 1991:141. (1)

Noch 1950 waren solche Probleme unbekannt. 'As late as 1950 such problems were unknown.'

The set of alternatives to the basic proposition of this sentence (the focus value induced by noch) is evaluated at earlier times than the one given in the case of noch. The more basic use of noch, however, is that of an 'aspectual operator'. As such noch is used to localise an event in time by specifying an interval within which it takes place. Consider the illustrative cases in (2) and (3). (2)

Hans schläft noch. 'Hans is still sleeping.'

(3)

Hans ist noch ledig. 'Hans is still unmarried.'

1

2

Noch, belongs to the class of focus particles together with expressions like: auch, gerade, insbesondere, schon, zumal, selbst, geschweige denn, sogar, ausgerechnet, bloss, eben, erst, genau, lediglich, nur, auschließlich, etc. (cf. König 1991:15, 33). Further members of this subclass are: auch, gerade, insbesondere, schon, zumal, selbst, geschweige denn, sogar. Other focus particles of German axe 'restrictive' (exclusive) in the sense that they imply that none of the alternatives satisfies the relevant proposition. The examples axe: ausgerechnet, bloss, eben, erst, genau, lediglich, nur, auschliesslich.

2

Grzegorz Dogil

In this basic use noch has been classified by traditional grammarians as a sentential adverb in an imperfective sentence 3 . Elegant formal semantic analyses of this use of noch have been provided in model-theoretic frameworks (cf. Hoepelman / Rohrer 1981; König 1991; Damova 1995). The aspectual properties of noch are far more intricate than the traditional term 'imperfective operator' leads one to believe. Abraham (1980) and Hoepelman /Rohrer (1981) identify the 'perfective' uses of noch in contexts illustrated by the following sentences: (4)

Wir gewinnen noch. 'We will win yet. / / We are still the winning side.'

(5)

Hans redet noch. 'Hans will speak yet/eventually. / / Hans is still speaking.'

The 'perfective' reading in cases like (4) and (5) is preferred, although the traditional 'imperfective' reading is still an option. However, in combination with the particle doch, or the adverb schließlich 'eventually', the perfective reading is forced. (6)

Wir gewinnen doch noch. 'We will win yet.'

(7)

Hans redet doch noch. 'Hans will speak yet/eventually.'

The perfective interpretation is also forced in cases where noch used without doch induces ungrammaticality (cf. Hoepelman / R o h r e r 1981, and note 3). (8)

Die Bombe explodierte doch noch. 'The bomb finally did explode.'

(9)

Hans kam doch noch an. 'Hans finally did arrive.'

König / Traugott (1982) identified an additional shade to the perfective meaning of noch. They describe this meaning shade as 'an additive function' and illustrate it with examples like the one in (10). (10) Ich trinke (doch) noch ein Bier. 'I will still have a beer.' The contribution of noch in these cases is that of 'adding up to a large whole' - the additive aspectual function. This function is enforced and modified by putting noch under sentence stress. Consider the cases in (11) and (12). 3

Hoepelman / Rohrer (1981:105) show that noch, used in a sentence with perfective ('achievement') verbs like explodieren 'explode', ankommen 'arrive', etc., render these sentence ungrammatical: *Die Bombe explodierte noch. *Hans kam noch an.

Intonation of Aspectual

Meaning

3

(11) Ich trinke (doch) NOCH ein Bier. 'I will have another beer.' (12) Hans besitzt NOCH ein Haus. 'Hans owns another house.' A very specific reading emerges in these cases. The additive function of noch implies an alternative value or quality within its scope, however, noch with a pitch-accent on it induces a value or quality of the same kind - i.e. it induces the additive aspect. (11) clearly means that I had (at least one) beer already, and (12) implies that Hans owns (at least) two houses. In this paper I will investigate in some detail how these various functions and readings of noch interact with the intonational properties of phrases that are shaped with its meanings. First I will introduce the theoretical framework of intonational grammar. Next I will suggest a set of theoretical hypothesis concerning the set of possible correlations between the temporal and the non-temporal (focal) uses of noch and the corresponding intonations. In the empirical section I will confront these stipulated correlations with the real-life data which has been grammatically and prosodically annotated at IMS during the recent years. The implications of these results for computational linguistics and for automatic speech processing will form the final part of this study.

2

Intonational meaning

The intonational grammar that I base my analysis on is the, by now standard, Tone Sequence Model [TSM]. The basic categories of the model are the phonological categories H and L. H and L are the abstract tone levels which acquire their fundamental frequency intepretations in the process of transforming the phonological into the phonetic representation (cf. Möhler 1998:80ff). The TSM grammar uses its abstract H and L categories to designate the heads and edges of intonational constituents. The heads, called pitch-accents, are conventionally marked with an asterisk (*), and the edges, phrase and boundary tones are usually marked with a minus (-) [intermediate phrase] and percent (%) [intonational phrase boundaries], signs. Edge categories are monotonal (H or L), the head marking pitch-accents, however, may consist of one or more tones. In cases of bitonal pitch-accent, the first tone is called the leading tone and the second one the trailing tone (e.g. H* + L; L + H*, etc.). The tone of a pitch-accent which is marked with an asterisk associates invariably with the stressed syllable. The leading tones (T-f) and the trailing tones (+T) of a complex pitch-accent are realised immediately before or after the peak (*) stressed syllables' tone. The phrase accents and boundary tones are realised on the syllables immediately preceding the designated boundaries. In Figure 1 some prototypical text-tune associations are illustrated (cf. Mayer 1997:27-32, for details).

4

Grzegorz Dogil

H*+L

L*+H

[[...:

a : cr* :]ip],p

IA

L+H* L-H%

Figure 1: Pitch-accent, phrase- and boundary tone associations. The intonational grammar defines an intonational phrase [IP] as sequence of tonal events such as in the formula EQ 1. (EQ 1) [T*(+T) ((T-) T*(+T)) T- T%]IP A well formed intonational phrase consists of at least one pitch-accent, at least one phrasal accent and one boundary tone. The individual categories of this grammar have been argued to contribute compositionally to the meaning of the propositions that they are associated with (cf. Pierrehumbert / Hirschberg 1990: 286-288; Mayer 1997:44-65):

Intonation of Aspectual

5

Meaning

• Pitch-accents convey information about the status of the individual discourse referents, modifiers, predicates, and relationships specified by the lexical items with which the accents are associated. • Phrase-accents convey the degree of relatedness of a designated intermediate phrase to preceding and succeeding intermediate phrases. • Boundary tones convey information about whether the designated intonational phrase is to be interpreted with respect to the succeeding phrase. The semantic information that the three types of accents convey is interpreted in terms of model-theoretic dynamic discourse semantics4 . The role of intonation is to update the mutual beliefs of the participants in discourse (Joshi 1982). The prepositional content of discourse referent(s) may be marked as salient (i.e. important for the correct interpretation of the proposition) by placing it under a pitch-accent. Similarly, the phrase-accents and boundary tones may mark the interpretive salience (forward/backward relatedness) of the proposition. Since noch is a typical discourse referent modifier we will be concerned with the contribution to meaning of pitchaccents associated with noch, and will disregard the contributions of phrase accents and boundary tones. The meanings of accents are composed of the meanings of the tones that constitute them. According to Mayer (1997:53) the following correlations obtain. Consider Table 1.

Peak (*) and leading (prefix) tones Trail (suffix), phrase, and boundary tones

H

L

salient + new

salient + not new

incomplete, open-ended

underspecified for completeness

Table 1: The meaning of H and L tones To illustrate the contribution of a pitch-accent to the propositional meaning of a discourse referent, consider the following example (cf. Mayer 1997:54). (13) A: Alan's such a klutz. B: He's a good badminton Player. L*+H L-H% The discourse referent 'badminton' is marked as salient (important for the correct interpretation of the discourse) by the fact that it is pitch-accented. L* tone signals 4

See FraCaS 1994 for a comprehensive summary of approaches to dynamic (discourse) semantics.

Grzegorz Dogil

6

that the salient information is not new. According to the interpretation suggested in Mayer (1997), information that is not new may be interpreted either as given or as false. In the particular case of (13) we assume that the information is given, i.e. t h a t both discourse participants know t h a t Alan is a good badminton player. The (+H) trail tone signals that the proposition is 'incomplete and open-ended', i.e. that its relevance for the discourse beliefs expressed so far still has to be clarified. The example provided above should elucidate the principles by which the intonational contribution to meaning is viewed in this study. Since we are dealing with the discourse referent expressed by a single phonological word (noch itself or discourse referents within its scope), we will be concerned with the intonational meaning of pitch-accents. Pitch-accents are the only categories of the intonational grammar whose scope is the phonological word. German intonational grammar within the TSM framework has five bi-tonal pitchaccent types (cf. Fery 1993, Mayer 1995). • H*+L • L*+H • H H * + L 'early peak' • L*+HL 'late peak' • H*M 'call contour' The 'call contour' is restricted to a single, vocative context. The 'early peak' and the 'late peak' accents are also restricted to specific contexts and meanings 5 which coincide with the very specific contexts and meanings of noch. Hence, we are left with the two pitch-accents [H*+L; L*+H] to express the full range of its meanings. In the following sections I will show that the usage of these accents correlates with the usage of noch. The leading hypothesis will establish how the basic meanings of noch which we identified in the first section, coincide with the meanings of accents which we described in the second section.

3

noch meanings and their prosody

In the first section I have argued (following König 1991) that noch may be used basically in the following four functions: • as an imperfective aspectual operator (see example (2)-(3)) 5

Cf. Kohler (1991), who first described the phonetic distribution and semantic contribution of these specifically German accents.

Intonation

of Aspectual

Meaning

7

• as a perfective aspectual operator (see example (4)-(7)) • as an inclusive focal particle (see example (1), (10)) • as a focal particle with an 'additive function' (see example (11)-(12)) From the point of view of intonational grammar noch is a phonological word which may carry a pitch-accent. As a focus particle it may also induce an accent on the head of the focused phrase within its scope. In order to test the initial plausibility of any systematic dependence between the basic meanings of noch and the pitch-accents I devised a simple production task. Three male native speakers of German, all of whom have substantial linguistic a n d / o r phonetic training, were asked to produce a set of sentences [2, 3, 4, 5, 1, 10, 11, 12]. The speakers were instructed to intone the sentences carefully, so that the expected implication of the proposition clearly came through. They were given five judgements for each sentence, three of which were then chosen for the analysis. All sentences were recorded in an anechoic chamber of the IMS using professional equipment. Speech signals were stored on an SGI workstation and processed by means of esps/waves software. The intonation of each of the 72 token sentences was labelled according to the TSM principles (cf. Mayer 1995). Let me first consider some of the expected results. In the imperfective use of NOCH we expect an intonation which makes the proposition open for further completion or interpretation. Such an intonational meaning is signalled by an H trail, a phrase and a boundary tone (cf. Table 1). The actual realisation of the representative examples looks as follows: (14) Peter schlaft noch. L*+H % L*.. H * + L % (15) Hans ist noch ledig. L*+H % H*+L L*+H %

(6) (3) (8) (1)

The high trail tone [+H], which we stipulated to correlate well with the imperfective reading is present in all realisations. 6 The prevailing [L*+H] pitch-accent together with a progredient terminal contour {[L*+H] %}, that is the contour present in most of the realisations, correlates very well with the semantic analysis of an imperfective meaning. Interestingly, in 4 realisations (all of them from the same speaker), in 6

The tune {L*.. H*+L}, which apparently does not contain an [+H] trailing tone is actually a product of a tonological linking rule (cf. Fery 1993:113ff). The rule converts underlying {[L*+H] [H*+L]} to a surface{L*.. H*+L}. From the point of view of intonational meaning the linked tone is present for interpretation as it can always be phonologically recovered.

Grzegorz Dogil

8

addition to the [L*+H] accent on the semantic head, noch has been assigned the [H*+L] pitch-accent. T h a t is, from the point of view of intonational meaning it has been marked as new and salient in the discourse (H* tone) and underspecified as far as completeness is concerned ( + L tone). I interpret this use as an a t t e m p t to mark the event of sleeping and being single as lying somewhere at the extreme margin of the time scale. The accenting noch with a [H*+L] tune signals the expectation t h a t Peter will soon wake up and t h a t Hans will get married pretty soon. This meaning is consistent with what semantics described as the scale/phase properties of noch. For the perfective realisations of noch we expect a realisation in which the salient elements are considered new and complete. The intonational grammar provides the [H*] peak-accent as a marker of newness and the [ L] trail, phrase and boundary tone as means of marking the sentence as underspecified for completeness. Hence the expected German pitch-accent to mark a perfective meaning is the [H*+L]. The realisation provided by the production experiment is as follows: (16) Hans redet noch. H*+L % (17) Wir gewinnen noch. H*+L % H H*+L %

(9) (5) (4)

The prevailing realisation chosen by the speakers is the expected {[H*+L] %} tune, the tune which correlates well with the perfective reading of this sentence. For (17) one speaker consistently used the 'early peak' [HH*+L] pitch-accent. The same production was also present in one of the realizations from another speaker. This 'early peak' accent, which was first discribed by Klaus Kohler (cf. Kohler 1991), may only be used in words whose pretonic syllable is weak (typically the German {ge-} or {be-} prefix, e.g. gelogen 'to lie', bekommen 'to get', gewinnen 'to win'). The prefix [H ] tone is associated with this pretonic syllable, and the [H*+L] pitch-accent is associated with the syllable carrying the main stress. Kohler interpreted the meaning of the 'early peak' as "..[setting] the frame for an established fact and the summing up of an argument, which is brought to a close" (1991:130). This proclamation very clearly renders what in formal semantics is meant by 'perfective reading'. In our compositional approach to intonational meaning, the leading [H ] tone expresses the newness and salience of the referring expression. This evaluation of the basic meaning is repeated by the [H*] peak accent, and the perfectivity is expressed by the trailing [+L] tone. The leading [H ] tone is in a sense redundant and restricted to a specific metrical context. If this metrical context (i.e. a metrically weak syllable) is missing (cf. (16)) the same information has to be expressed with a 'simple' [H*+L] pitch-accent. T h a t this is an option is indeed illustrated by the renditions of this example chosen by other speakers in the experiment. It has to be stressed, however, t h a t the production of a [H H*+L] pitch-accent does have a dramatic 'perfectivity' and 'finality' effect, as described by Kohler. In Figure 2 we illustrate the fO tracing of one of the productions of the [H H*+L] accent on sentence (17). Sentences (16) and (17) have also an imperfective reading (cf. section 1 of this study and Konig 1991:142). On this reading it is implied by (16) t h a t Hans is still

Intonation of Aspectual Meaning

9

200.1 i i i i , i i i

TT

150-

. V V A 101 .

snLL

38.60

Wir

ge

. . , ! , .• 38.80

Winnen 1

•

1

• • •

39.00

noch

••••I

I••

39.20

39.40

Figure 2: A 'perfective' realization of noch with an 'early peak' on gewinnen. speaking, and by (17) that the game is still being played (and that the chance of winning is small). Test speakers realise these 'imperfective' readings consistently with the {[L*+H] %} tune, the tune which we established to correlate best with the imperfective usage of noch as an aspectual operator. As a focal 'inclusive-evaluative' particle noch opens a scale of alternatives and 'evaluates' it in the sense that the particle fixes the value 'relatively late' on this scale. Intonationally a pitch-accent of salience and newness (i.e. [H*]) is expected on the head of the phrase which is within the scope of focus. This, in fact, is predicted for all inclusive focal particles. Because German has only bitonal pitch-accents we expect a realisation of a [H*+L] pitch-accent on the heads focused by noch. The actual realisations of the speakers in our production experiment fully support this prediction. (18) Noch 1950 waren solche Probleme unbekannt. H*+L % H*+L H% (19) Ich trinke noch ein Bier. H*+L %

(6) (3)

(9)

The [H*+L] pitch-accent realised on 'Bier' in (19) identifies it as focus. Similarly, the main stressed syllable in the scope of focus in (18) (i.e. fiinfzig 'fifty') is realised with this same pitch-accent. The boundary tone (H%) in (18), which was realised by one of the speakers, suggests that the focus phrase is to be interpreted in relation to the phrase following it. This contribution of the boundary tone is consistent with the syntax of the proposition but it is independent of the interpretation and realisation of the focal noch. When focal noch itself is stressed, the proposition carries a special 'additive function' meaning (recall the discussion in section 1). This additive aspectual function is uniquely expressed by the [H*+L] pitch-accent realised on noch, the accent whose meaning correlates very well with the intended semantics of the proposition. (20) Ich trinke noch ein Bier. H*+L

(9)

10

Grzegorz Dogi1

(21) Hans besitzt noch ein Haus. H*+L

(9)

The results of the production experiment described above show a high consistency in the use of intonation with various semantic uses of noch. All three speakers vary slightly in the realisations, however, categorical phonological choices are almost invariant. The basic pattern of results is summarised in table 2. aspectual operator

focal particle

imperfective

perfective

inclusive

additive

pitch-accent

L*+H

H*+L

H*+L

H*+L

alignment

head of the modified proposition

head of the modified proposition

head of the focused phrase

noch

Table 2: Results of the production experiment As an imperfective aspectual operator noch selects a [L*+H] accent on the semantic head of the proposition that it modifies. As a perfective operator it selects the [H*+L] (or [H H*+Lj) accent. As a focal particle it assigns the [H*+L] accent to the head within its scope, and when it fulfils the 'additive function' it carries the [H*+L] pitchaccent itself. I have shown that the results of the phonetic production experiment are consistent with the hypotheses predicted by the theoretical work on prosody and its correlations with semantics. The results are also consistent with the research results on formal properties of aspectual operators and focal particles. Such clear theoretical results are of immense value for computational linguistics. Notice that the inventory of expressions of the type of noch (aspectual operators and focal particles) is finite in any language, and if such clear correlations between pitch-accent and usage could be established many areas of computational linguistics could benefit. For example, in machine translation the distinction between perfective and imperfective usage of a particle is a source of serious mistakes. In the area of speech recognition an appropriate classification of a recognised word greatly reduces the search space in a language model. In the area of language generation and speech synthesis the selection of noch (and other particles) with appropriate accentual properties would contribute not only to prosodic naturalness but would also boost the interpretability of the generated message. It has often been observed, however, that such neat theoretical results based on the analysis of highly selective constructed data and tested in a very limited laboratory set up, are of little use for 'real life' computational applications like speech recognition, speech synthesis and machine translation. Often this reservation against theoretical linguistics and constructed data comes from language engineering approaches which attempt to draw conclusions from the analyses

Intonation of Aspectual Meaning

11

of large data bases. These approaches show that theoretically obtained generalisations disappear, or at least become statistically irrelevant, as soon as large data bases of real speech are investigated. At IMS we are in the lucky position of having access to very large data bases of 'real' language and speech as well as the tools to process these data bases. In the following I will show how our theoretical predictions on the intonational features of noch fare against such 'real life' data.

4

Exploratory data analysis

The basis for this investigation was created by the pioneering work of Stefan Rapp. Using the basis technology of automatic speech recognition, Rapp (1995, 1996, 1998) developed a series of tools which allow the automatic annotation of speech with orthographic, syllabic and phonemic transcription (alignment); automatic annotation with parts of speech (POS tagging); and, most importantly - automatic prosodic annotation of any speech sample with pitch-accents and boundary tones. All these tools are robust, dependable and highly functional. The results that they produce are not much worse than those delivered by human labellers (cf. Rapp 1998: 49). The tools are fully automatic, and hence allow unrestricted creation of highly annotated speech data bases. The investigation presented below is based on three such automatically created data resources. • Acted Speech Data Base. A word aligned and prosodically (automatically) labelled literature reading: "Die Leiden des jungen Werther" by Johann Wolfgang Goethe. The data base was created on the basis of a CD-ROM which contains the text and complete performance by a professional speaker (ISBN 3-15-100002-9). 4 hours of acted speech (Rapp 1996). • Radio News Data Base. Word aligned and prosodically (automatically and manually) labelled extracts of news stories from "Deutschlandfunk". 1,5 hours of rehearsed speech, various speakers. • Feature Series Data Base. German version of Stardate (called "Sternzeit") recorded daily since 1996 August 1st. Radio station: Deutschlandfunk - digitally transmitted over satellite radio and downsampled to 16khz. The text was taken from the internet (Deutschlandfunk's Sternzeit page). Speech and text were automatically word aligned and automatically (and partly manually) labelled. 2 hours of read speech, 1 speaker.

12

Grzegorz Dogi1

From each of these data bases all the occurrences of noch and their discourse environment were automatically selected.7 The Radio News Data Base is the most intricate one because some of the noch's were repeated in the same context at successive news broadcasts. As these repetitions were produced by various speakers we could compare the consistency of intonational use of noch by various speakers. Radio presenters proved to be very consistent in the use of their intonation in general (Cf. Rapp 1998:166ff), and in the case investigated here there appear to be no inter-speaker differences. The part of this data base that I concentrated on contains 30 occurrences of noch. In one case noch is stressed, the accent is the expected [H*+L]. The context in this case implies clearly that this noch is used in the 'additive function', as predicted by the theory.8 (22) Es gab danach jedoch noch eine wissenschaftliche Aufgabe L*+H H*+L für Apollo 11. H*+L % 'There was yet another scientific task for Apollo 11 after that.' There are 7 cases of [L*+H] accent in the vicinity of noch, and all of them may be classified as cases of 'imperfective aspectual operator' use - again as predicted by the theory set out in the earlier parts of this study. (23) Noch in diesem Jahr kauft man einen Jumbojet.... L*+H (L) L*+H 'Before the end of this year a jumbo jet will be bought...' The remaining 22 noch's are realised with a [H*+L] accent within their scope. This is a situation predicted for the use of noch as an inclusive focal particle or a perfective aspectual operator (see Table 2). Scrutiny of examples revealed not a single case of noch as a perfective aspectual operator. Actually, we found no cases of this use of noch in our 'real life' data. This obviously does not have to mean that such uses do not exist. The experiment described in a previous section shows that they are clearly present in the competence and performance of the native speakers. They are, however, clearly constructed as they induce a contrastive situation and tenuous ambiguity, which, possibly, is avoided in 'real life' speech. Whatever the reason might be, we have no data to make any generalisations about the perfective use of noch in 'real life' speech. Unfortunately, all the 22 realisations of noch with a [H*+L] accent within its scope can not be interpreted as cases of the focal particle use (as the theoretical results summarised in Table 2 would predict). Many of these uses (actually 10 out of 22) are examples of noch as an imperfective aspectual operator. Consider the example (24) below. 7 8

I would like to thank Karin Müller for helping me with the selection. Müller (1998) found in her extensive study of real speech data that almost all cases of pitch-accented noch carried a H*+L accent and had the 'additive' meaning.

Intonation of Aspectual

13

Meaning

(24) Über die Kapitulation der 1500 Mann starken Einheit, die sich in den Bergen versteckt halte, werde noch verhandelt. H*+L% 'The capitulation of a 1500 man unit, which is still hiding in the mountains, is still being negociated.' The discrepancy between the 'imperfective' function of the operator and the 'perfective' accent is not restricted to this particular data base. In the Features Series Data Base ("Stardate" read by one male speaker) noch as imperfective operator is realised with the 'incorrect' [H*+L] accent very frequently. In the Acted Speech Data Base the situation is less dramatic, i.e. the actor recites his text often with the 'correct' [L*+H] accent when imperfective use of noch is implied. There are, however, numerous examples when [H*+L] is used in contexts in which it is not predicted by the laboratory experiment. The other uses of noch are better harmonized with intonation. In its 'additive' function noch is always stressed and the accent it takes is the predicted [H*+L]. In the use as an inclusive focal particle noch assigns the pitch accent to the head of its focal scope. Here, however, the variation of the accent type is much greater than in the cases of the additive focal use. There are also other contexts and accents (mainly boundary tones) relevant to noch which will not be considered in this study because their relation to the semantic and prosodic theories discussed at the outset is not yet clear (cf. Miiller 1998:section 2.1, for a more detailed discussion of these cases). The main trends of the data base supported correlations between noch and pitch-accents are given in table 3. aspectual operator

focal particle

imperfective

perfective

inclusive

additive

L*+H and H*+L

no clear cases found

H*+L L*.. H* H*..

H*+L

pitch-accent

head of the modified proposition

head of the focused phrase and the particle itself

noch

alignment

Table 3: Correlations between noch and pitch-accents

14

5

Grzegorz Dogil

Implications for language and speech technology

The 'real life' data turn out to be very much more knotty than the theoretical hypotheses and the experiments would lead one to believe should be the case (compare Tables 2 and 3, for the main trends). This is, however, surprising only at first sight. The results considered in this section are based on behavioural data. Unlike in pure theoretical linguistics, where one or very few counterexamples are enough to refute a claim, claims based on behavioural data, which are the result of a complex cognitive process, are seldom absolute. If a given factor influences the examples found in behavioural data corpora more often than chance would predict, than we infer that it plays a role in processing. This does not mean that all data will be governed by that factor; in fact, a number of counterexamples may arise. In the case discussed in this study, the counterexamples to the intonational behaviour of noch as an aspectual operator may have several sources. First of all, the imperfective meaning of the propositions may be so strongly signified by the context that no additional accentual reinforcement appears necessary. Secondly, in numerous contexts noch may be considered semantically 'underspecified' in the sense of Discourse Representation Theory (cf. Reyle 1986, 1993). In those cases no accent at all or a default pitch-accent ([H*+L] in German) is customary in the vicinity of noch. Other, paralinguistic, factors may also introduce counterexamples. These are speech style, text type or simply speech errors. Fortunately, the redundancy of behavioural data is such that although the prosodic and the semantic features of a proposition apparently do not match, flawed inferences are rarely the case. A non-trivial question for real life applications of computational linguistics is what kind of data should be modelled in order to achieve optimal results. Should it be theoretically predicted and experimentally validated data as in Table 2, or data derived through statistical approximation of large data bases as in Table 3. There is no definitive answer to this question yet. However, I would like to point to some unexpected experimental results from speech generation which will probably guide our future research. I mentioned at the close of the theoretical section of this paper that a good correlation between abstract predictions and the results of the production experiment could be used as a basis for speech synthesis. We could very well imagine that various functions of noch could be used to control intonation in a full text-to-speech generation system. In the section on 'real life' data I argued that these rules clearly overgenerate, and that the accentual patterns that they predict are rarely, or even not at all found in the 'corpus' data. At IMS we asked ourselves, in the framework of research on text-to-speech synthesis (cf. Möhler/Dogil 1995, Möhler 1998), what would be the case if we actually simulated such an ideal corpus. That is, what would the reaction of listeners be to a set of data in which all accents behaved according to the linguistically validated rules of prosody. Möhler (1998:163) carried out a perceptual evaluation of speech samples 9 containing the original intonation curve and the intonation curve that has been automatically generated from a set of rules. The listeners not only 9

For better comparability the speech samples were replaced by diphone synthetic stimuli.

Intonation of Aspectual Meaning

15

judged the resulting synthesized intonational contours to be highly natural, but they even rated the automatically generated intonation as more natural than the original one. This 'better-than-real-life' result indicates that theoretical linguistic hypothesis about intonation modelling and their interdependence on other parts of the grammar are correct and that they have the potential to be successfully applied to speech technology. Meeting listeners' expectations appears to be as important as uncovering (often hidden, ambiguous, underspecified) speakers' intentions.

6

References

Abraham, W. 1980. The syncronic and diachronic semantics of German temporal noch and schon, with aspects of English still, yet and already. Studies in Language 4, 3-24. Damova, M. 1995. Adverbs in the transfer module of MDS. Verbmobil Bericht 100, IMS, Universität Stuttgart. Doherty, M. 1973. Noch and Schon and their presuppositions. In F. Kiefer / N. Ruwet (eds.) Generative Grammar in Europe. Dordrecht: Reidel. Fery, C. 1993. German intonational patterns. Tübingen: Niemeyer. FraCaS [A Framework for Computational Semantics LRE 62-051] Deliverable D8, 1994. Describing the Approaches. The FraCaS Project Administrator, University of Edinburgh, Centre for Cognitive Science. Hoepelman, J. / Rohrer, C. 1981. Remarks on noch and schon in German. In Tedeschi / Zaenen (eds.) Tense and Aspect: Syntax & Semantics 14: 103-126. New York: Academic Press. Joshi, A.K. 1982. The role of mutual beliefs in question-answer systems. In, Smith, N. (ed.) Mutual Knowledge. New York: Academic Press. Köhler, K.J. 1991. Studies in German Intonation. AIPDS [Arbeitsberichte des Instituts für Phonetik und digitale Sprachverarbeitung] No. 25. König, E. 1977. Temporal and non-temporal uses of noch and schon. Linguistics and Philosophy 1: 173-198. — 1991. The meaning of focus particles: A compaxative perspective. Routledge: London. Ladd, R. 1996. Intonational Grammar. Cambridge University Press, Cambridge. Löbner, S. 1989. Schon - erst - noch. An integrated analysis. Linguistics and Philosophy 12: 167-212. Mayer, J. 1997. Intonation und Bedeutung. Aspekte der Prosodie-Semantik-Schnitt-stelle im Deutschen. Phonetik AIMS (Arbeitspapiere des Instituts für Maschinelle Sprachverarbeitung, Lehrstuhl für Experimentelle Phonetik, Universität Stuttgart), Vol. 3, No. 4. Möhler, G. / G. Dogil. 1995. Test environment for the two level model of germanic prominence. Eurospeech '95, Madrid: 1019-1022 Möhler, G. 1998. Eine theoriebasierte Modellierung der deutschen Intonation für die Sprachsynthese. Phonetik AIMS (Arbeitspapiere des Instituts für Maschinelle Sprachverarbeitung, Lehrstuhl für Experimentelle Phonetik, Universität Stuttgart), Vol. 4, No. 1. Müller, K. 1998. German focus particles and their influence on intonation. Diplomarbeit. IMS, Universität Stuttgart. Pierrehumbert, J. / J. Hirschberg, 1990. The meaning of intonational contours in the interpretation of discourse. In: Cohen, P. et al. (eds.) Intentions in Communications. Cambridge: MIT Press, 271-311.

16

Grzegorz Dogi1

Rapp, S. 1995. Automatic phonemic transcription and linguiatic annotation from known text with Hidden Markov Models/ An aligner for German. In Boguslavky, I. ed. Integration of Language and Speech in Academia and Industry, Moscow: IMACS. - 1996. Goethe for prosody. Proceedings of the fourth international conference on spoken language processing, ICSLP 96, Philadelphia: 1636-1639. - 1998. Automatisierte Erstellung von Korpora für die Prosodieforschung. Phonetik AIMS (Arbeitspapiere des Instituts für Maschinelle Sprachverarbeitung, Lehrstuhl für Experimentelle Phonetik, Universität Stuttgart), Vol. 4, No. 2. Reyle, U. 1986. Zeit und Aspekt bei der Verarbeitung natürlicher Sprachen. PhD Thesis, University of Stuttgart. - 1993. Dealing with Umbiguities by Underspecification. Journal of Semantics, 10: 123179. Rooth, M. 1985. Association with Focus. PhD Thesis, University of Massachusetts, Amherst. - 1992. A theory of focus interpretation. Natural Language Semantics 1, 75-116. Shetter, W.Z. 1966. The meaning of German "noch". Language 42, 42-66. Traugott, E.C. / Waterhouse, J. 1969. Already and yet: A suppletive set of aspect markers? Journal of Linguistics 5: 287-304.

Anette Frank and Annie

Zaenen1

Tense in LFG : Syntax and Morphology

1

The morphology-syntax interface in LFG

An important tenet of LFG is the lexical integrity principle which says that the leaves of c-structure trees are complete surface words. Given this principle, the morphological component is seen as distinct from the syntax. It can be modelled by sublexical rules as we will illustrate below but the principles that apply to these rules are different from those applying in the syntax (see Bresnan and Mchombo (1995) for discussion). The way LFG is set up allows single words and phrases to contribute the same or similar information to an f-structure. For example a form like parla, passé simple of parler, contributes information similar to that contributed by a parlé, the passé composé of the same verb. The framework allows a similar treatment for the two forms as well as the maintenance of lexical integrity and makes it possible to avoid word formation rules in the syntax without losing paradigmatic transparency (see Vincent and Bôrjars (1996) for discussion). These possibilities, however, are not always exploited as well as they could be, and using them transparently is made less easy than it could be by another architectural feature of LFG. The distinction that the architecture of LFG makes between c-structure and f-structure was meant to embody the insight that word order and other constituent structure differences are not necessarily indicative of profound syntactic differences among languages. LFG follows here the distinction made e.g. in Keenan (1976) between coding properties and genuine syntactic characteristics. The f-structure allows us to abstract away from superficial word order differences to bring out the more fundamental syntactic similarities (or differences) among languages. This abstracting away from certain differences is theoretically important but also practically, e.g. in the context of translation. It makes the f-structure into a structure that comes close to the underspecified representation used in the Core Language Engine (see Genabith and Crouch (1996); (Alshawi (1992), Alshawi and Crouch (1992)), which can be argued to be, from a practical point of view, a good candidate for input and output of transfer rules (see Dorna et al. (1998)). But the traditional architecture gets us only half way: while it abstracts away from c-structure phenomena, it encodes all the

1

We would like to thank the following people for helpful comments on earlier versions of this article: John Maxwell, who proposed a similar architecture in conversations with the second author, Ron Kaplan, Miriam Butt, Frederique Segond and Veronika Kniippel. In particular we thank Joan Bresnan for extensive comments and suggestions. The issues she raised could not be discussed in sufficient detail in this short contribution. Needless to say that the commentators do not necessarily share the perspective we are taking here. Special thanks go also to Marc Dymetman for judgements on French data. We alone are responsible for remaining errors.

18

Anette

Frank / Annie

Zaenen

morphological information in t h e f-structure. This information, however, is to a large extent as much encoding information as word order is. T h e way things stand in t h e traditional architecture, it is possible t o get all t h e f-structure relevant information a b o u t parla and a parlé into t h e right place in t h e fstructure. But it is not possible for t h e f-structure t o ignore t h e additional information needed to impose t h e right verbal form on parlé (past participle). In w h a t follows we discuss some proposals m a d e t o remedy this and t r y t o improve technically on them. We t h e n use our tools t o model t h e French auxiliary system.

1.1

Analyses of auxiliaries

Early analyses in L F G (Falk (1984), Bresnan (1982)) analyzed auxiliaries as raising verbs, assigning t h e m a PRED value, e.g., ' p e r f ' in t h e case of a t e m p o r a l auxiliary. Later approaches (Bresnan (1995), King (1995), Schwarze (1996)) t r e a t auxiliaries as non-subcategorizing elements, which contribute tense and aspectual information to t h e f-structure of t h e clause. Under this t y p e of analysis, the main verb is the functional head of t h e clause. Among the a r g u m e n t s t h a t are p u t forward for an analysis of auxiliaries as non-subcategorizing elements we find t h e following. Auxiliaries contribute essentially temporal and aspectual information, which indicates t h a t they should belong t o a different syntactic category t h a n ordinary predicational verbs (see B u t t et al. (1996a)). Even within a single language system - there may be analytic and synthetic tense forms which do not differ substantially in meaning (besides, of course, certain aspectual or t e m p o r a l differences). If t h e analytic form is analyzed as contributing a PRED on t o p of t h e tense and aspect information whereas t h e synthetic form doesn't, we of course do not express t h e functional similarity between the two cases. T h e auxiliary in the analytic form should therefore not be analyzed as a PRED-bearing element, but only contribute its tense a n d aspectual information (see King (1995)). B u t t et al. (1996a) consider cross-linguistic variation in (analytic vs. synthetic) tense formation as providing s u p p o r t for an analysis of t e m p o r a l auxiliaries as non-subcategorizing elements. In recent work Bresnan (1995) t r e a t s auxiliaries as functional categories in an L F G framework and enforces an analysis in which tense auxiliaries as well as passive ones are non PRED-bearing elements.

1.2

T h e m - s t r u c t u r e analysis of auxiliaries

In line with this movement towards a flat f - s t r u c t u r e analysis of auxiliary constructions, B u t t et al. (1996a) and B u t t et al. (1996b) propose a unified analysis of auxiliaries in English, French, and G e r m a n , with a flat f - s t r u c t u r e for all three languages. Their emphasis is on problems of parallel g r a m m a r development and machine translation. In particular, one way of looking at their proposal is t h a t - although morphology and c-structure make different contributions t o t h e functional and semantic analysis of a sentence in different languages - t h e corresponding f - s t r u c t u r e representations should not be distinct because their contribution t o t h e meaning of t h e sentence is the same.

19

Tense in LFG

Their analysis differs from the previous ones in providing a clear separation between cross-linguistically invariant f-structure features of temporal constructions, and language specific differences in the way this information is encoded. Such differences are exemplified in (1): while in English and German future tense is formed analytically (will turn/ wird drehen), French has a synthetic future tense (tournera). (1) a. The driver will turn the lever. b. Der Fahrer wird den Hebel drehen. c. Le conducteur tournera le levier. In this analysis, a new projection is introduced, the morphological structure, or mstructure. The m-structure is, just like the f-structure, an attribute-value matrix, but - while f-structure is the level of syntax that encodes grammatical functions, like SUBJ, OBJ, etc., and the PREDicate with its subcategorization requirements - the m-structure is viewed as the level of representation that encodes information about idiosyncratic constraints on morphological forms. Thus, m-structure is the level of representation where the language specific differences in the morpho-syntax of tense formation are represented. In much the same way as the f-structure is defined as the ^-projection off the cstructure, Butt et al. (1996a,1996b) define the m-structure as a /i-projection off the c-structure (2).

Syntax-morphology interface

(Butt et al. (1996a,b))

In this projection architecture, the analyses assigned to the sentences in (3) will be isomorphic at the level of f-structure (4). The associated morphological structures will be distinct for English and German, where the future is an analytic tense form (5a.-b.), as opposed to French (5c), where future tense is formed synthetically by inflectional morphology. The phrase structure rule will of course also encode the differences in word order which are likewise ignored in the f-structure. 2 (3) a. The driver will have turned the lever. b. Der Fahrer wird den Hebel gedreht haben. c. Le conducteur aura tourné le levier. 2

We do not a t t e m p t to give semantically motivated features for tense and aspect here. However, it is possible to define minimal semantic temporal conditions triggered by tense forms (like imparfait, passé simple, passé composé, etc.) in a semantic projection a of if>. K a m p and Rohrer (1983) propose that such an encoding be based u p o n notions like Reichenbach's (1947) temporal reference point, etc. Further refinements for a Reichenbachian approach to the French tense s y s t e m have been proposed by K a m p and Rohrer (1988) and Gosselin (1996). A substantial fragment of English is treated in K a m p and Reyle (1993).

20

Anette 'turn/drehen/tourner ( (f

PRED TENSE

SUBJ)

(t

OBJ)

Frank / Annie

Zaenen

)'

FUTPERF

'driver/Fahrer/conducteur'

PRED CASE

NOM

GEND NUM

(4) a./b./c.

SPEC

MASC SG DEF

PRED

'lever/Hebel/levier'

CASE

ACC

GEND

MASC

NUM

SG

SPEC

DEF

Structurally identical f-structures for English, German, French (3)

FIN AUX

(5) a.-b.

+ +

FIN

AUX

ß

+

VFORM DEP

AUX DEP

C.

BASE

VFORM

ß

AUX DEP

—

+ + AUX VFORM

— PERFP

PERFP

Structurally distinct m-structures for English, German (a./b.) vs. French (c.)

Following the projection architecture displayed in (2), the morphological structures in (5) are defined in terms of c-structure annotations. We will use the notation to refer to the m-structure node projected from the mother node * of the actual c-structure node *. In the same way, the traditional f can be expressed by . We will call such equations morphological equations. 3 (6) illustrates how structurally divergent m- and f-structures are projected from the c-structure: here the VP node of the auxiliary construction defines a hierarchical structure at the level of the /i-projection DEP) = ), whereas the functional equation of this node is a trivial one, t=4-

3

Butt et al. (1996a) misleadingly introduced the notation "fi to refer to the m-structure of the actual node's mother node. This is in fact not the correct notation for the architecture displayed in (2): it defines m-structure to be projected off the f-structure. In what follows we take the freedom to rephrase their approach in terms of the notation introduced above, which follows Butt et al. (1997). Alternatively, the annotations could equally well be restated in terms of the notation introduced in Kaplan(1987), where f i M * refers to the m-structure node projected from mother node M * .

21

Tense in LFG

VP

NP ( t s u b j ) = 4-

le conducteur Vaux (6)

Î = 4

VP t = 4

V = V

DEP) =

aura

NP

V

(f OBj)= I

*fi —

le levier

tourné

The lexical entries of auxiliary verbs and main verbs come with both functional and morphological equations, which define the language particular properties of morphosyntactic tense formation, as well as the corresponding tense information, which is encoded in the f-structure. We briefly illustrate this with the two lexical entries used in (6), the future auxiliary aura (will have), and the participle verb tourné (turned). The auxiliary is morphologically marked for future tense in French, and is constrained to combine with a past participle verb form, in its m-structure's DEP feature. The T E N S E feature in f-structure can therefore be set to the value F U T P E R F . 4 aura: Vaux ( î subj num) = sg (t subj pers) =

(7)

1.3

tourné: V (f p r e d ) = 'tourner((f s u B j ) ( t OBJ))' aux)=

3

( t T E N S E ) = FUTPERF

(V

-

vform)=

perfp.

( v fin)= + (* M A U X ) = + DEP V F O R M ) = c P E R F P .

Some problems of the current morphology-syntax interface

The syntax-morphology interface described above works very well for verbal morphology and the distinct temporal and aspectual constructions in the diverse languages described in the paper (see Butt et al. (1999)). Yet, one may have noticed t h a t the m-structure (5c) for the sentence displayed in (6) does not contain features like NUM, PERS, G E N D , or CASE, which prima facie should also be considered as morphological features. This raises the general question as to the distinction between morphological, functional syntactic, and finally also semantic information, and the distribution of these respective types of features over the various levels of representation assumed in the overall projection architecture of grammar. If morphological features like number, gender and person are to be represented in m-structure, we cannot simply introduce them by means of trivial morphological 4

The main verb avoir is assumed to have a different entry and is not discussed in this paper.

22

Anette

Frank / Annie

Zaenen

equations = . Since the various arguments of the verb may instantiate conflicting values of number, person, and case, the m-structure must specify "blueprints" of the fstructural grammatical functions S U B J , OBJ, etc., to host the respective morphological features. In the actual LFG grammar implementations of the ParGram project (see Butt et al. (1999) for information) the m-structure is defined to contain an attribute E X T - A R G (external argument), corresponding to the morphological structure of the f-structure's S U B J , a set-valued feature I N T - A R G S (internal arguments) for non-subject arguments (OBJ, OBJ2, OBL, etc.), a feature DEP corresponding to sentential arguments 5 ( C O M P , XCOMP), and finally a set-valued feature N O N - D E P for adjuncts.

1.3.1

Defining subject verb agreement

This "blueprint" of the functional argument structure of a sentence would in fact allow us to move agreement and case features to the morphological representation level. The subject verb agreement constraints of, e.g., a finite third person singular verb can then be stated in terms of the following morphological equations: (*M

E X T - A R G N U M ) = SG a n d

EXT-ARG P E R S ) =

3.

With this extension, the m-structure for (6) would spell out as in (8):6 EXT-ARG FIN

(8)

AUX DEP

PERS

3

NUM

SG

+ + AUX

—

VFORM INT-ARGS

PERFP

GEND NUM

MASC SG

This, however, starts to look suspicious: the m-structure proposal was originally motivated by the wish to have a clean f-structure representation of what matters to semantic interpretation. But now we start to get an m-structure representation that repeats most of the information pertaining to f-structure. The morphological representation level now contains subcategorization information, which is functional in 5

The feature EXT-ARG, e.g., was introduced to account for the morphological selection of infinitival vs. finite sentential adjective suBJects in cases like (i) and (ii). The verbal inflectional features FIN and INF being stated in m-structure, the morphological form of the sentential adjective SUBJ cannot be constrained in terms of these features without resorting to some EXT-ARG feature in m-structure. The distinction between finite and infinite sentential subjects could, alternatively, only be captured in terms of the f-structure attribute TENSE, which should then not be assigned in structures like (i) and (ii). (i)Obtenir son accord n'est pas facile. Getting her/his agreement is not easy (ii)Avoir obtenu son accord est une victoire. Having gotten her/his agreement is a victory.

6

For ease of exposition INT-ARGS is not represented as a set-valued feature here.

Tense in LFG

23

nature. It is becoming a complete "blueprint" for the functional structure. Besides this conceptual issue, it turns out that this approach is confronted with rather severe problems in the analysis of long-distance phenomena that involve morphological constraints.

1.3.2

The problem of long-distance dependencies

A typical example for a long-distance phenomenon in the morphology-syntax interface is past participle agreement in French. In object relative clauses (9) the past participle must agree, in number and gender, with the embedding head noun, as opposed to cases where the OBJ is realized in VP position, and where past participle agreement is illicit. (9) a. Les enfants adorent les histoires qu'on leur a déjà raconté*(es) mille fois. 'Children are found of the stories that one has told them already a thousand times' b. Les enfants adorent les histoires qu'on sait bien qu'on leur a djà raconté*(e.s) mille fois. 'Children are fond of the stories that one knows perfectly (that) one has told them aready a thousand times' The grammar assigns the relative pronoun the OBJect function, but can do so only in terms of a functional uncertainty equation, since relative clause constructions are unbounded in French (see (9b)). In order to trigger past participle agreement of the verb that subcategorizes for the object introduced by the relative pronoun, the morphological features of number and gender must be stated in the position of the relative pronoun, 7 yet have to be "transmitted" to the level of the subcategorizing past participle. In the architecture depicted in (2), this will only be possible if these morphological features are defined in terms of a functional uncertainty equation over the "f-structure blueprint" that is encoded in the morphological structure, in particular functional uncertainty over the DEP attribute (see (10) below).8 For concreteness, we add morphological equations to the annotations of the category PRONrel (for relative pronouns) in (10) that enforce the choice of the inflected participle form racontées (see (11)), which agrees with the head noun histoires in number and gender. The lexicon entry ( l l ) 9 constrains this inflected form to syntactic contexts in which the OBJ head-precedes the verb: 10 (T OBJ) • N[_num]: NUM)= _num. b. NP[sg] N[sg]: (tM NUM)= sg. c. NP[pl] N[pl]: (tM NUM)= pi. (52) shows how the crucial distinctions between auxiliary and main verbs, as well as finite, infinite and participle verb forms can be encoded in a system of parameterized (sublexical) rules. The two parameters used are _type, with possible values aux or main, and _ f i n , which allows for the values f i n , i n f , or p a r t (for participle forms). The important thing to note here is the effect of parameterization. Without any constraining parameter, any of the three alternative sublexical verb rules could be used to derive a V-category in grammar. With parameterization, however, it is possible to constrain certain positions in a grammar rule to specific complex V-categories. E.g., specifying V [ a u x , f i n ] in a grammar rule will constrain the V-category to finite auxiliary verbs, as defined in the sublexical rule for finite verbs in (52). In this way, reference to complex V-categories will allow us to hard-wire the relative order of main and auxiliary verbs in the flat VP structure, as illustrated in (51). VP[fin]

(51) V[aux,fin]

(V[aux,part])

V[main,part]

t = ;

t = i

t = i

By trivial functional equations t = 4- in the sublexical rules (52) the equations that are stated in the morphological lexicon (53) contribute to define partial f- and m-structures for the various complex V-categories. 27 Moreover, the sublexical constituents are annotated with equations that define, or instantiate the respective parameters. Thus, the complex V-type V[main, f i n ] , e.g., is restricted to sublexical structures with sublexical constituent TNS. In virtue of the distinct morphological features A U X = — and AUX= +, projected from the lexicon, the t y p e parameter of auxiliaries is correctly set to aux and, correspondingly, the parameter t y p e is instantiated to main for main verbs. Finally, the distinction between main verb and auxiliary participles (cf. (48) above) is now captured in terms of the complex sublexical category PART [_type], which is defined in the morphological lexicon for the respective types main and aux (see (53)).

27

We are using the actual XLE syntax, where a trivial equation f = 4- is added by default if no J. arrow appears in the annotation, but often state the trivial equation explicitly, for better readability.

41

Tense in LFG V[_type, _ f i n l

{

{ _type = main

V_BASE:

(TV A U X ) =

-

I _type = aux (TM A U X ) = + } ;

TNS:

t =4

VS_NUM: VS_PERS: VTAG V BASE:

t = 4t = l

_fin = fin;

{ _type = main (IV A U X ) =

-

I _type = aux ( t v AUX)=

(52)

INF:

+};

t = i _fin = inf;

VTAG V BASE:

{ _type = main ( t V DEP* A U X ) = —

I _type = aux ( t v DEP A U X ) = + } ;

PART[_type]:

t = l

VPART_NUM: VPARTGEND: V TAG

t = 4; t = 4 }

_fin = part;

Parameterized sublexical rules for complex V-category (preliminary version)

(53)

+Pres

TNS

( t T E N S E ) = PRESENT

+PaPrt

PART[main]

{ (TV DEP* V F O R M ) = PERFP

+P1 +3rd +P1

PART[aux] VS_NUM VS_PERS VPART NUM

(tV FIN)- +• | (TV DEP* V F O R M ) = PASSP } ; ( t V DEP V F O R M ) = PERFP. ((T SUBJ) M N U M ) = PL. ( ( t SUBJ) M P E R S ) = 3. {

((t

SUBJ)

A

NUM)=C

PL

| ( ( t OBJ)^ NUM)= c PL } . . .

+Fem

VPART NUM

{

((t

SUBJ)^

GEND)=

C

FEM

| ( ( t OBJ) M G E N D ) = C FEM } . . .

Morpho-syntactic annotations for sublexical constituents Based on the parameterized V-category defined in (52) and the (partially stated) functional annotations for the morphological tags in (53) we can impose fine-grained distinctions on the order of verbal elements in complex verb phrases by explicitly stating the order of the respective instantiated complex verb categories in the cstructure rules. The following (parameterized) c-structure rule for temporal auxiliary constructions encodes two possible structures for finite (_f i n = f i n ) or infinite (_f i n = i n f ) phrases. 28 The first disjunct allows for a single main verb of the appropriate finiteness-

42

Anette

Frank / Annie

Zaenen

type. The second disjunct captures a verbal sequence consisting of an (obligatory) auxiliary of the instantiated finiteness type, followed by a participle main verb, and an optional intervening participle auxiliary. (54)

VP[_fin]

{ V[main,_fin] | V[aux,_fin] (V[aux,part]) V[main,part] } . . . other VP-complements . . .

This rule accounts for the following sentence types. With _f i n set to f in: (55) a. Il vient. b. Il est venu. c. Il a eu travaillé. with _f i n set to i n f : Il peut ... (56) a. ... venir. b. ... être venu. c. ... avoir eu travaillé.

2.3.4

Temporal auxiliary vs. passive verbal structures

Abeillé and Godard (1996) argue for the following distinction. VP î=4

(57) a.

Vaux

(Vaux)

V

t=4-

t=4-

t=4-

flat c-structure (temporal auxiliaries)

b.

hierarchical c-structure (passive, predicational constructions)

To account for both the flat VP structure for temporal auxiliaries, as well as the hierarchical VP structure for passive and predicational constructions, the rule system presented above has to be slightly refined because passive constructions are constrained to c-structure configurations with an embedded VP, as opposed to a flat VP structure for non-passive, temporal constructions. 29 VP[_type,_fin] (58)

28

{ V[main,_fin] I V[aux,_fin] (V[aux,part]) V[main,part]: ( t PASSIVE) = | V[aux,_fin] (V[aux,partj) VP[main,part]: ( t PASSIVE)=C + } . . . other complements

This rule does not take into account complex predicate formation, as e.g. with causative faire. 29 In the rule fragment given below, we restrict ourselves to temporal and passive auxiliaries. The predicative auxiliary être could either be defined as only contributing morphological and aspectual features, along the lines of the passive auxiliary, or else as a PRED-bearing element which subcategorizes for an XCOMP complement.

43

Tense in LFG

The disjuncts in rule (58) are devised for the following construction types, respectively: (59) a. il vient / de venir / venant / venu b. il est venu / il a travaillé / il a eu travaillé c. il est attendu / il a été attendu / d'avoir été attendu The order of the verbal elements in complex verbal structures is captured in cstructure by reference to complex (parameterized) verbal categories, which we defined in the sublexical rules (52) and (53). With these rules in place, the L F G verb lexicon will consist of entries for verb stems rather than fully inflected forms. The morphological features of inflected verb forms are projected from the annotations of the sublexical rules and sublexical constituents. For auxiliaries of type aux, however, the lexicon has to encode a distinction between participle auxiliaries and other morphological forms. In order to define the correct formation of complex tenses, the participles été and eu must specify morphological annotations that are distinct from those of their finite (or infinitival) forms (recall this fact by reconsulting (47) vs. (48) above). A t the same time, the annotations for these two auxiliary participles été and eu are also distinct, and must therefore be specified in the corresponding lexical (stem) entries. In the refined sublexical rule for verbs (61) we therefore define a complex stem category V_BASE [ p a r t ] for (temporal) auxiliary participles. 30 This category shows up as the stem category V[part] in (60), the auxiliary stem lexicon for the temporal and passive auxiliaries avoir and être. It distinguishes participle forms V [ p a r t ] from other morphological forms, which are covered by the category V. The annotations of the respective disjuncts correspond fully to the ones given for fully inflected forms in (47) and (48). être:

V

(fM AUX)=

+

( t M DEP A u x ) = — (T M DEP A U X - S E L ) = C ÊTRE { (T^ DEP V F O R M ) = C PERFP

est venu

( t PASSIVE)= — ( t T E N S E ) = PAST I (TN DEP V F O R M ) = c PASSP (T PASSIVE)=c

est vu

+

( T T E N S E ) = PRESENT } ;

(60)

VI

V[part]

(îp

AUX)=

;

j, +

(T M DEP A U X ) =

+

(F M DEP DEP A U X - S E L ) = C ÊTRE { (T M DEP DEP V F O R M ) = C PERFP

a été arrivé

( t PASSIVE)= — (T T E N S E ) = PAST-ANTERIOR I ( t M DEP DEP V F O R M ) = C PASSP ( t PASSIVE)=C

a été v u

+

(T T E N S E ) = PAST } .

30

W e assume that V.BASE a n d V.BASE [ p a r t ] are considered as distinct categories.

44

Anette V

V[part]

AUX)= + (t passive)= — { (t^ DEP AUX)= (t,. DEP AUX-SEL)= c AVOIR (t TENSE)= PAST | (f M DEP AUX)= + } (t^ DEP VFORM)= c PERFP; (t„ a u x ) = + ( t passive)= DEP AUx)= + (f^ DEP DEP AUX-SEL)=c AVOIR (t,i DEP DEP VFORM)=c PERFP ( t TENSE)= PAST-ANTERIOR.

Frank / Annie

Zaenen

a travaillé

a eu travaillé, a été arrivé a été vu

a eu travaillé

The sublexical rule for the complex category verb now reads as follows. Note that it differs from (52) only with respect to the parameterization of the stem category V_BASE for auxiliary participles. V[_type, _fin]

{

V_BASE:

{ _ t y p e = main (t„ a u x ) = I _ t y p e = aux (tv AUx) = + };

TNS:

t = 4-

VS_NUM: VS_PERS: V_TAG V BASE:

t = 4; t = 4;

_ f i n = fin;

(61)

INF: V_TAG { V_BASE: I V_BASE[part]

{ _ t y p e = main (ÎM AUX)= I _ t y p e = aux (1V a u x ) = + }; t = ; _ f i n = inf; _ t y p e = main (t^ DEP* a u x ) = — _ t y p e = aux (t M DEP a u x ) = + };

PART[_type]:

t =4

VPART NUM: VPART GEND: V TAG

t = 4; t = 4; }

_ f i n = part;

Parameterized sublexical rules for complex V-category (final version) With the above rules and lexicon entries, we can now illustrate the relevant aspects of the analysis for II a été vu. The finite VP rule instantiates the parameter _f i n to f i n . The finite auxiliary a can fill the first position in the second disjunct of rule (58). The categorial parameters and functional annotations of both été (V [aux, p a r t ] ) and vu (V[main,part]) are appropriate to expand the structure further as given in (62).

Tense in LFG

45 S

í =4

(62)

Pron

VP[aux,fin]

( t SUBJ) = 4-

t = 4-

il

V[aux,fin]

V[aux,part]

VP[main,part]

t = 4-

1 = 4-

a

été

t = 4I

V[main,part]

t = 4-

The reader may verify, on this basis, the morphological and functional annotations that are defined by the lexical entries and sublexical rules in (60) and (61), and how they resolve to the wellformed morphological and functional structures (63).

'FIN AUX

+ + VFORM AUX

DEP

(63)

PERFP +

[VFORM DEP

PASSP

AUX

—

_AUX-SEL PRED TENSE

'voir ( N U L L ,

SUBJ [PRED

)' '

CASE GEND

PAST

PASSIVE

SUBJ

+ 'pro'

]

ÉTRE

NOM MASC

NUM

SG

PERS

3

The corresponding German sentence Er wurde gesehen will be assigned an equivalent f-structure representation, but a distinct m-structure, which misses one level of embedding. In German the tense information is introduced by a single past passive auxiliary wurde. 2.4

Some consequences and some possible extensions

2.4.1 Where do clitics go? As the reader might already have observed, the proposed system allows us straightforwardly to do away with most cases of clitic climbing in French. Clitics are local arguments in the f-structure and their functional annotations reflect this. We will assume a phrase structure rule, introducing the clitics as independent words which are attached to the verb. We could also consider them to be part of the verb, in LFG nothing hinges on this.

46

Anette

VP[_type,_fin] ->

Frank / Annie

Zaenen

(CL: { ( | OBJ) = | C A S E ) = C ACC I (t

(64)

OBJ2)=|

( 4 ^ C A S E ) = C DAT}

(CL:

( t OBJ) = |

...)

(CL: ( t OBJ2) = I

...))

V[_type,_fin]: t = I • The annotations on the personal clitics will be as illustrated in (65). la : C L

(65) V '

( t PRED) = (tfl N U M ) =

'pro' SG

(TM G E N D ) = FEM ( Î M C A S E ) = ACC.

The above clitic rule can be integrated into the complex VP rule established above as in (67). It not only allows for the ordinary cases of "clitic climbing" with auxiliaries like (66a), but in fact prevents illicit clitic positions as in (66b). (66) a. Il les a vu. b. *I1 a les vu. VP[_type,_fin] -> (CL: { (t

OBJ)=4 ...

(CL: (t 0 B J ) = 4 (CL: ( t O B J 2 ) = |

(67)

|

(f

OBJ2)=J. . . . }

...) ...))

{ V[main,_fin] | V[aux,_fin] (V[aux,part]) V[main,part]: (t PASSIVE) = | V[aux,_fin] (V[aux,partj) VP[main,part]: (t P A S S I V E ) = C + . . . other complements

The equations of y and en will have to be more complicated as it is well known that they can represent material that is not a direct argument or even a direct adjunct of the main predicate of the sentence as for example in (68). (68) J'en ai vu la première partie. I saw the first part of it. We do not go into that aspect of French syntax in this paper. 2.4.2

Possible extensions

The copular construction is not limited to être but can also occur with verbs like sembler, rester, etc. For those verbs we could not claim that they have no contribution to make beyond a morphological feature bundle and tense or aspect information. An explicit account of them is beyond the scope of this paper but sentences like the one in (69) indicate that the complex predicate approach proposed for causative constructions will need to be extended to them.

}

Tense in LFG

47

(69) Il lui reste fidèle. He remains faithful to him/her. If we do not want to complicate the clitic rules, these verbs will combine with their adjectival complements in the way faire combines with its verbal complements. LFG proposals for the treatment of complex predicates have been made, among others, in Alsina (1996), in Butt (1995) and for French in Frank (1996) and Dalrymple and Zaenen (1997). They need to be adapted to the proposal made here. This should not create any problems. The extension to rester, etc. should be straigthforward as these cases are simpler than the causative ones given that the subcategorization of the adjective does not change under the various types of embedding.

2.4.3 Possible problems Problems for an approach that is uniform across languages arise, however, in several other cases. French has not only a synthetic future tense it also has synthetic modal e.g. je travaillerais, 'I would work'. Should the English conditional be analysed like the French one? We assume the answer is yes but a further type of problem is raised by the existence of tenses like the immediate future in French as illustrated in (70): (70) II va le faire. He will do it. The value of aller in this context is very similar to that of a verb like will/shall in English or a morpheme like -ai/-as/-a/etc. in French. Here, however, the clitic placement facts do not plead in favor of a flat f-structure solution, whereas the general consideration about the relation between syntax/morphology and semantics do. It would not be impossible to extend the proposal made above to this case but further investigations will show whether the advantages of bringing the f-structure closer to a semantic structure are compatible with a perspicuous description of the syntax of individual languages.

3

Open conceptual issues

Our proposal of a morphology-syntax interface that separates functional-syntactic from morphological information in a sequenced projection architecture raises a number of important conceptual issues.31 LFG has been very successful in abstracting away from order constraints and in that way bringing out the similarities between typologically distinct languages. As we stated in the introduction this is important, not only from a theoretical point of view but also from a practical one. The typological distinction between synthetic and analytic languages is also adressed within the theory. LFG accounts in a straightforward 31

We are grateful to Joan Bresnan, who raised several of the following issues.

48

Anette Frank / Annie Zaenen

way for the fact that the same functional information can be encoded in one word in one language and spread over several words in another without giving up lexical integrity and without loosing the distinction between sentential syntax and word internal morphology (see e.g. Bresnan and Mchombo (1995)). However, as Butt at al. (1996a) observed the level of abstraction is pushed less far here. To push it further we need a careful study of what the informational content of various elements is. Extending Butt et al. (1996a) we propose here an architecture that would make a distinction between coding projections, such as the c-structure and the m-structure and informational projections, such as the f-structure and the semantic structure. The c-structure manages the order constraints among syntactic elements whereas the m-structure manages the purely morphological dependencies between word forms. Of course the same element will typically play a role in various projections: the c-structure orders e.g. an NP before a VP but this does not prevent this NP from contributing the functional role of a subject, on the contrary, in a language like English it is by virtue of its position that it contributes this information. Similarly, the tense markers which can be independent words or affixes, play a role in the c-structure and/or the m-structure but of course also in the f-structure and the semantics. In this paper we have not discussed these issues but it is obvious that the main raison d'etre of auxiliaries is not to take participle complements. What our proposal embodies is the claim that the functional contribution of the various morphological elements is not a one-to-one correspondence with the form of their encoding. As shown e.g. in Bresnan and Mchombo (1987) (on pronoun incorporation in Chichewa), the same morphological form can have different functional roles in one and the same language and across languages. Case has distinct functions across and within languages. In the spirit of several authors (e.g. T. Mohanan, A. Wierzbicka) we distinguish between morphological content and morphological forms. A same morphological form, e.g. a specific case can have a different content in different contexts. Under our proposal morphological form information would go into the m-structure whereas morphological content would contribute to the f-structure. The separation of representation levels lets us encode various kinds of mismatches across levels for features like number, person and gender in a straightforward and explitic way. Such a multilayered representation could also allow us to be more explicit about the presence or absence of morphological marking in a particular language. It is not possible to give an a-priori answer to what belongs to the various projections without detailed analysis of morpho-syntactic phenomena across a variety of languages. A case in point is a puzzle in asymmetric agreement under coordination in Welsh, studied and discussed in Sadler (1999). One of the alternative approaches Sadler proposes to solve the problem is to postulate agreement features at both f-structure and m-structure, following the projection architecture proposed in the present paper. A related but more formal issue arises from the proposed- sequenced, as opposed to a parallel projection architecture. The sequenced architecture that we proposed implies a functional mapping from f-structure to m-structure. This architecture does not permit a single partial f-structure to map to distinct m-structures. Now, could we be confronted with languages where two expressions that bear distinct morphological markings unify at the level of f-structure? Such a one-to-many relation could not be

Tense in LFG

49

represented in the sequenced projection architecture. Possible examples could be case attraction phenomena, for example with relative pronouns that appear in the case of the head noun, as opposed to the case of the syntactic argument they represent within the relative clause. Again, it has to be studied whether such cases can be accommodated by distributing inconsistent assignments over distinct levels of representation, or whether alternative approaches, such as the set-based feature theory of Dalrymple and Kaplan (to appear) - which is successfully applied to similar agreement problems - can accommodate such facts. Complex predicate formation is another example where elements with distinct morphological (and functional) features are mapped to a single f-structure unit. This has been discussed in Frank(1996), where a restriction-based approach (including a parallel m-projection) was proposed to account for various problems in the standard LFG treatment of complex predicate formation. In general, constructions which involve expressions with distinct morphological markings that are unified to the same f-structure unit are difficult to handle in the standard LFG architecture. The sequenced m-structure architecture can only handle such configurations if the mismatching elements are appropriately distributed over the distinct levels of representation, or else by adopting additional formal devices, such as the restriction operator originally introduced by Kaplan and Wedekind (1993), or the set-based approach to feature resolution by Dalrymple and Kaplan (to appear). Finally, and on a more technical note, one might wonder whether the formal device of complex c-structure categories that we used above to capture order constraints in the sequenced architecture could be extended to an approach where all morphological constraints are encoded in terms of complex c-structure categories. A separate level of representation for morphological constraints would then be unneccesary. At first glance it seems, though, that not all morphological distinctions can be naturally encoded in terms of c-structure categories. In the case of the French auxiliary system, for example, one has to express certain restrictions on tense formation which preclude ungrammatical constructions like *est eu travaillé as opposed to the well-formed a eu travaillé,

a n d similarly for *est été arrivé as o p p o s e d t o a été arrivé.

To c a p t u r e

these restrictions, an analysis that relies on purely c-structure categorial distinctions will have to encode the lexical form of the auxiliary, être vs. avoir, as a c-structure parameter of auxiliary categories. Here we would have to decide whether this kind of lexicalization is still within the range of a natural complex c-structure category.

4

Conclusion

In this paper we have proposed a new architecture for the /¿-projection in LFG. Our proposal has the advantage that it does not need to reproduce essentially functional information about syntactic arguments in the m-struture. We illustrated our approach with the description of a substantial fragment of the French auxiliary system and in doing so we showed that the problem of stating ordering constraints in the sequenced /¿-projection architecture can be solved in a rather elegant way through the use of

50

Anette

Frank / Annie

Zaenen

parameterized rules. T h e discussion of the French data elaborates further on the advantages of the ¿¿-projection approach advocated by B u t t et al. (1996a) but it also points to further phenomena that need to be investigated to get to a crisper view of what the division between language specific and universal aspects of syntax should be. T h e approach raises numerous interesting and intricate theoretical questions about the partitioning of linguistic features across the various levels of representation.

5

References

Abeillé, A. & Godaxd, D. 1996. La Complémentation des auxiliaires français. Langages 122: 32-61. Alsina, A. 1996. The Role of Argument Structure in Grammar. Evidence from Romance. CLSI Lecture Notes 62, Stanford, California. Alshawi, H. (ed) 1992. The core language engine, The MIT Press. Alshawi, H. & Crouch, R. 1992. Monotonie Semantic Information, in: Proceedings 30th Annual Meeting of the Association for Computational Linguistics, 32-38. Asher, N. 1993. Reference to Abstract Objects in Discourse, Kluwer, Dordrecht. Bresnan, J. 1995. Lexical-Functional Syntax. Lecture Notes, ESSLI 7, Barcelona. - 2001. Lexical-Functional Syntax. Blackwell Publishers, Oxford. Bresnan, J. & Kaplan, R.M. 1982. The Mental Representation of Grammatical Relations. Cambridge. Bresnan, J. & Mchombo, S.A. 1987. Topic, pronoun, and agreement in Chichewa, in -.Language 63,4, 741-782. - 1995. The Lexical Integrity Principle: Evidence from Bantu, in: Natural Language and Linguistic Theory 13, 181-352. Brun, C. 1996. Using Priority Union for Non-Constituent Coordination in LFG, in: M. Butt and T.H. King (eds): Proceedings of the first LFG conference, Rank Xerox Research Centre, Grenoble, CSLI Online Publications, http://www-csli.stanford.edu/publications/. Butt, M. 1995. The Structure of Complex Predicates in Urdu. CSLI Publications, Stanford. Butt, M., Nino, M.E. & Segond, F. 1996. Multilingual Processing of Auxiliaries in LFG, in: D. Gibbon (ed.): Natural Language Processing and Speech Technology: Results of the 3rd KONVENS Conference, Bielefeld, October, 111-122. Butt, M., Fortmann, Ch. & Rohrer, Ch. 1996. Syntactic Analyses for Parallel Grammars: Auxiliaries and Genitive NPs, in: Proceedings of COLING 1996, 182-187. Butt, M., Dalrymple, M. & Frank, A. 1997. An Architecture for Linking Theory in LFG, in: M. Butt and T.H. King (eds): Proceedings of the LFG'97 conference, University of California, San Diego, CSLI Online Publications, http://www-csli.stanford.edu/publications/. Butt, M., King, T., Nino, M.E. & Segond, F. 1999. A Grammar Writer's Handbook, Stanford, CSLI Publications. Couquaux, D. 1979. Sur la syntaxe des phrases prédicatives en français, in: Linguisticae Investigationes, III-2: 245-284. Dalrymple, M. 2001. Lexical-Functional Grammar. Volume 34 of Syntax and Semantics, Academic Press.

Tense in LFG

51

Dalrymple, M. & Kaplan, R. 1997. A Set-based Approach to Feature Resolution, in: M. Butt and T.H. King (eds): Proceedings of the LFG'97 conference, University of California, San Diego, CSLI Online Publications, http://www-csli.Stanford.edu/publications/. — Feature indeterminacy and feature resolution, in: Language 76 (4), 759-798. Dalrymple, M. & Zaenen, A. 1997. Les verbes causatifs "polymorphiques": les prédicats complexes en français, in: Langages 122: 79-95. Dorna, M., Frank, A., van Genabith, J . & Emele, J . M. 1998. Syntactic and Semantic Transfer with F-Structures, to appear in: Proceedings of the ACL/COLING conference, Montréal. Falk, Y. 1984. The English Auxiliary System: a Lexical-Functional Analysis, in: Language 60 (3): 483-509. Frank, A. 1996. A Note on Complex Predicate Formation: Evidence from Auxiliary Selection, Refìexivization, Passivization, and Past Participle Agreement in French and Italian, in: M. Butt and T.H. King (eds): Proceedings of the first LFG conference, Rank Xerox Research Centre, Grenoble, CSLI Online Publications, http://www-csli.stanford.edu/publications/. van Genabith, J. & Crouch, R. 1996. Direct and underspecified interpretations of LFG f-structures, in: Proceedings of COLING 1996, Copenhagen, Denmark, 262-267. Gosselin, L. 1996. Semantique de la temporalité en Français, Duculot. Hankamer, J . & Sag, I.A. 1976. Deep and Surface Anaphora. Linguistic Inquiry 7: 391-426. Huot, H. 1981. Les constructions infinitives du Français, Librairie Droz, Genève. Kamp, H. & Reyle, U.1993. From Discourse to Logic, Dordrecht, Kluwer. Kamp, H. & Rohrer, Ch. 1983. Tense in Texts, in: R. Bäuerle, Ch. Schwarze, A. von Stechow (eds): Meaning, Use, and Interpretation of Languge, Berlin: de Gruyter, 250-269. - 1988. Temporal Reference in French, ms, University of Stuttgart. Kaplan, R. 1987: Three Seductions of Computational Psycholinguistics, in: P. Whitelock, H. Somers, P. Bennet, R. Johnson, M. McGee Wood (eds): Linguistic Theory and Computer Applications, 149-188, London, Academic Press. Kaplan, R. & Wedekind, J . 1993. Restriction and Correspondence-based Translation, in: Proceedings of the EACL93, Utrecht, the Netherlands, 193-202. Kaplan, R. & Zaenen, A. 1989. Long-Distance Dependencies, Constituent Structrue, and Functional Uncertainty, in: M. Baltin, A. Kroch (eds): Alternative Conceptions of Phrase Structure, University of Chicago Press, 17-42. Reprinted in: M. Dalrymple, R.M. Kaplan, J.T. Ill Maxwell, A. Zaenen (eds) 1995: Formal Issues in Lexical-Functional Grammar, CSLI Lecture Notes 47, CSLI Publications, Stanford, California, 215-239. Keenan, E. 1976. Towards a universal definition of subject, in: C. Li (ed) Subject and Topic, Academic Press,303-333. King, T.H. 1995. Configuring Topic and Focus in Russian. Stanford, California: CSLI Publications. Vincent, N. & Börjars, K. 1996. Suppletion and syntactic theory, in: M. Butt and T.H. King (eds): Proceedings of the first LFG conference, Rank Xerox Research Centre, Grenoble, CSLI Online Publications, http ://www-csli.Stanford.edu/publications/. Reichenbach, H. 1947. Elements of Symbolic Logic, London: Collier-MacMillan. Sadler, L. 1999. Non-Distributive Features and Coordination in Welsh, in: M. Butt and T.H. King (eds): Proceedings of the LFG'99 conference, University of Manchester, CSLI Online Publications, http://www-csli.stanford.edu/publications/. Sag, I., Gazdar, G., Wasow, Th. & Weisler, S. 1985. Coordination and How to distinguish Categories, Stanford University, CSLI Publications.

52

Anette

Frank / Annie

Zaenen

Schwarze, Ch. 1996. The syntax of Romance Auxiliaries, in: M. B u t t and T.H. King (eds): Proceedings of the first LFG conference, Rank Xerox Research Centre, Grenoble, CSLI Online Publications, h t t p : / / w w w - c s l i . s t a n f o r d . e d u / p u b l i c a t i o n s / . Zaenen, A. &c Kaplan, R. 1995. Formal Devices for Linguistic Generalizations: West Germanic Word Order in LFG, in: M. Dalrymple, R.M. Kaplan, J.T. I l l Maxwell, A. Zaenen (eds): Formal Issues in Lexical-Functional Grammar, CSLI Lecture Notes 47, CSLI Publications, Stanford, California, 215-239.

Hubert

Haider

Adverbials at the Syntax-Semantics Interface

1

Introduction

Recent theorizing (cf. Chomsky 1995) within Generative Grammar generally opts for a strict positional characterization of the interface relations between syntactic structure and semantic interpretation: There is a syntactic representation, namely LF, that provides unique structural domains to be mapped on unique domains in the semantic representations. To name two examples, Diesing's tree splitting algorithm (Diesing 1992) is intended to provide a positional characterization of the division between the nuclear scope (namely VP) and the restrictor portion of a clause, while, for instance, according to Cinque (1997), adverbials are hosted by appropriate spec-positions of (silent) functional heads which are tied to a unique semantic interpretation. Thus the syntax of adverbials is characterized as directly correlated with their semantic value. This paper argues for a structurally relative interface between an autonomous syntax component and a semantic component. The structure of the syntactic representation of a clause is a function of autonomous structure principles. Apparent semantic effects on syntactic structuring are characterized as interface effects: Semantic interpretation does not directly interfere with syntactic structuring, but there is an indirect effect based on a monotony property of the interface: Hierarchies of syntactic structures are mapped monotonically on semantic domains. Thus, elements in a syntactically higher domain cannot be interpreted as elements of a semantically lower domain, unless there is a chain relation to a syntactically lower domain that is mapped onto the semantically lower domain. Here is an example: (1)

a.

This has frequently completely confused some of them

b.

"This has completely frequently confused some of them

The unacceptability of (l.b) follows either from structure or from interpretation. A structural account would have to assume that there is a specific structural position for manner adverbials, which is different from the structural positions of frequency adverbials and that the latter position is higher than the former. Obviously, this positions must be functionally defined, not categorially. On the other hand, (l.b) can be ruled out on semantic grounds in combination with appropriate interface conditions: Manner adverbs are predicates of a verbal denotation (i.e. the characterization of a process denoted by the verb). Frequency adverbials are predicates of properties of events. Let us take it for granted for the time being that the domain of events includes the domain of the verbal head, denoting a process, both in structural terms as well as in terms of the semantic interpretation. In this case, an adverb preceding and thereby c-commanding a frequency adverb cannot get an interpretation in lower a domain than the domain of the frequency adverb. Hence,

54

Hubert

Haider

the manner adverb in (l.b) cannot get a manner-interpretation. But, since this is its only available interpretation, 1 the clause contains an ill-interpreted term, whence its unacceptability. If the manner adverb in (1) is replaced by an adverb t h a t is ambiguous between a manner interpretation and an interpretation as a sentence adverbial, any one of the two resulting orders is acceptable: (2)

a.

He has frequently cleverly confused some of them

b.

He has cleverly frequently confused some of them

In (2b), just as completely in (l.b), cleverly cannot be interpreted as a property of the manner he brought about the confusion, since its c-command domain contains a frequency adverb which opens an event-domain interpretation. So, cleverly must be assigned a denotation that fits the interpretation of a V P t h a t is in the domain of a frequency adverb, that is, the interpretation of a property of events: Since this is indeed possible for cleverly, a well-formed interpretation results, namely: It was clever that he frequently confused some of them. The paper is organized as follows: In the next section, a particular implementation of a positional interface, namely Cinque's (1997) treatment of adverbial syntax, is put to test. The empirical as well as theoretical difficulties encountered are taken to be indicative of, and representative for, the difficulties of a positional interface. Section 3 briefly presents the outlines of an alternative account in terms of an incremental and monotonic mapping of relative c-command domains on domains of the semantic representation of an expression. More details and implications are presented in section 4, which analyzes properties of temporal adverbials. The conclusions are listed in section 5.

2

A positional interface? - Minimalist maximalization

In the Principle & Parameter-system, adverbials are taken to be elements adjoined to non-argument constituents, like VP or IP. The proliferation of silent functional heads in recent developments opens another avenue. Adverbs could be assigned to spec-positions of projections of functional heads, which remain silent and therefore void of a morphological representation. This has been argued for by Alexiadou (1994) and subsequently in full generality by Cinque (1997): There is a cascade of functional projections with silent F-heads on top of VP, whose specs host the respective adverbials. The order of adverbials in a clause reflects the nesting of functional projections. Since there must be a separate functional projection for each adverbial in which the semantic value of the adverbial in the spec-position harmonizes with the semantic value of the silent functional head, each adverbial gives rise to a functional projection: 1

Of course, the c-command domain could be mapped on a domain suitable for a manner adverb, but then the frequency adverbial contained in this domain would be squeezed into a domain that does not fits its requirements.

Adverbials at the Syntax-Semantics (3)

Interface

55

The new law certainly may possibly have indeed been badly formulated (Quirk et als. 1986: § 8.20, p. 495)

According to this analysis, the adverbs in (3) are each followed by a silent functional head that projects a functional projection. The same considerations apply to a language like German. In the following clause, eight functional projections are needed in order to host the adverbs plus negation that occur between the subject and the object. (4)

Könnte man hier denn nicht jetzt etwa mutwillig noch schnell ein Adverb mühsam hinzufügen? 'couldn't one here PRT 2 not now PRT sportively just quickly an adverb laboriously add ?'

It is worth pointing out that the order of the adverbs in (4) is one of various possible alternative orders, as for instance, the order given in (5). In Cinques's framework, this situation entails either of two moves: The meaning of the adverbs changes if their position changes, or the cascade of functioal projections must be freely rearrangeable. Cinque adopts the first alternative. But the difference between (3) and (4) is scopal at best: It is not the meaning of the items that has changed but only their relative scope. In a positional view, however, change of order is a change of spec-position, and hence a change of the inherent semantic relation between the element in a spec-position and its functional head. (5)

Könnte man denn nicht etwa hier mutwillig jetzt schnell noch mühsam ein Adverb hinzufügen? 'could one PRT not PRT here sportively now just quickly just laboriously an adverb add?'

The positional approach is in conflict with non-trivial empirical and theoretical considerations. First, the assumption that the ordering options of adverbials are a function of their appearing in the specs of alternative functional heads with an alternative semantic value is not evident, especially for a language with a considerable wealth of rearrangeable items, as for instance illustrated by German with examples as those cited above. Secondly, the cascade of functional heads must be projected separately for every V-projection. This becomes obvious with perception verb complements, as in the following example: (6)

2

a.

Er hat sie wieder (nicht) wieder * (singen) gehört 'he has her again (not) again (sing) heard' He has again heard her sing again

b.

Er hat sie leider (nicht) wieder singen gehört 'he has her deplorably (not) again sing heard'

"PRT" stands for "particle". In German, there axe various particles whose semantic contribution could be rendered into English only by means of paraphrases. They code, for instance, for speaker- or hearer-oriented background assumptions.

56

Hubert Haider

c.

*Er hat sie wieder (nicht) leider singen gehört 'he has her again not deplorably sing heard' He has not heard her sing again deplorably

d.

*Sie hat wieder (nicht) leider gesungen 'she has again (not) deploralby sung'

In (6.a), each occurrence of 'again' relates to one of the main verbs. If there is only one main verb, the repetition becomes unacceptable. So, there must be two occurrences of the same functional head, hence two potential cascades of F-projections. But, the two projections cannot be exploited separately, as (6.c) illustrates. (6.c) is unacceptable for the same reason as the simple clause (6.d), in this respect: A sentence adverbial cannot appear in the sope of 'again'. Technically, however, the two adverbial projections are separated, so they could in principle project separately. Third, if adverbs occupy functional spec-positions, these spec-positions can in principle host constituents of arbitrary complexity. Empirically, this is contradicted by the fact that adverbials that can occur between the subject and the VP in VO-languages are subject to restrictions which are absent in OV-languages: (7)

(8)

a.

He has [(much more) carefully (*than anyone eise)] analyzed it

b.

He has [(much less) often (*than I (thought))] rehearsed it

a.

Er hat es [sehr viel sorfältiger als jeder andere] analysiert (= 7a) 'he has it [much more carefully than anyone else] analyzed'

b.

Er hat es [viel weniger oft als ich dachte] geprobt (=7b) 'he has it [much less often than I thought] rehearsed'

Descriptively, the restriction illustrated by (7) - the head of the adjunct must not have material to its right - is paralleled by the restriction known from adnominal attributes: (9)

a.

eine [viel größere (*als ich dachte)] Summe

b.

ein [unzufriedener (*damit)] Syntaktiker

'a much bigger (than I thought) sum' 'a unsatified (it-with) syntactician'

The descriptive generalization that covers the two instances is this: Adjunction to a head-initial constitutent is subject to the restriction that the head of the adjoined projection P must be adjacent to the node to which P is adjoined. Note that the crucial factor is not the headedness value of the adjoined element. 3 What matters is the headedness-value of the adjunction site. Whatever theoretical principle is responsible for this fact, it is independent of the problem under discussion: Elements in uncontroversial spec-positions are not subject to this restriction, neither in cases of spec-head agreement, as in the case of spec-I, or in cases of spec-C, as in the top spec-position in V2-languages. 3

AP-attributes are head-initial in English but head-final in German. Nevertheless the restriction is identical for adnominal attributes.

Adverbials at the Syntax-Semantics

Interface

57

The restriction on intermediate adverbial positions found in English is paralleled by Danish (cf. 10), which is representative of Scandinavian languages. Dutch (cf. 11) patterns like German, however. (10) a.

Han er forlaengst kommet hjem fra universitetet [Da] 'he has long-ago come home from university-the'

b.

*Han er [for en time siden] kommet hjem fra universitetet 'he has [since an hour ago] come home from university-the'

(11) a.

als niemand luider (dan Jan) kan roepen [Du] 'if noone louder than Jan can shout'

b.

omdat ik het vel sneller (dan J a n (denkt)) in een zakje kan pakken since I it much feister (than Jan (thinks)) in a bag can pack'

The restriction illustrated above is problematic for the spec-F analysis of adverbials in at least two aspects. First, it is unexpected for a spec-position. It is indicative of genuine adjunction structure. Second, the contrast between OV and VO languages is unexpected as well: A specific restriction for adverbial specs should apply to both language types. Only in an adjunction analysis, the headedness property of the adjunction site could matter. Fourth, and most important, the spec-F analysis of adverbials implies t h a t phrases preceding an adverbial must be either in a spec-position or adjoined to the Fprojection t h a t hosts the adverbial. Constituents adjoined to F-projection are opaque for extraction and so are F-specs. (12) a.

*WhOi was [a picture of e»] sold?

b.

* Which windowsillj did [across e*] walk an ant?

c.

* Which subject* has [about e*] he never talked?

d.

*Whoj would [to be able to talk to e*] be more fun?

Irrespective of the analysis - adjunction or substitution into an F-spec - the facts are robust: Extraction out of a constituent in spec-I or higher is ungrammatical. The evidence from German, however, contradicts the implication that constituents preceding adverbials are either adjoined to F-projections or occupying spec-F positions: (13) a.

Wenj hat [e* damit zu überzeugen] schon jemals einer versucht? 'who has [it-with to convince] already ever someone tried?' who has someone ever tried to convince with this?

b.

Wenj ist [e^ damit zu überzeugen] leider kaum jemandem gelungen? 'who has [it-with to convince] deplorably hardly someone succeeded?' who did deplorably hardly anyone succeed to convince with this?

The relevant opacity conditions are of course operative in German. This can be verified with clear cases of spec-positions, as the spec-position preceding the position of the finite verb in V2-clauses (cf.l4.b,d):

58

Hubert Haider

(14) a.

Worauf; glaubst du, daß sie gesagt hat, daß er geglaubt habe, daß ich e, warte? 'what-for do you think that she has said that he believes that I had waited' What do you think that she has said that he believes that I had waited for?'

b.

* Worauf, glaubst du, daß sie gesagt hat, [{daß ich e t warte] habe er geglaubt]? 'what-for do you think that she has said [[that I had waited] has he believed]'

c.

Wasj glaubst du, daß sie gesagt hat, daß er ej-für Leute kennen würde? 'whatj believe you that she has said that he ej-(kind) of people would know?' What kind of people do you believe that she said that he would know?

d.

"Was, glaubst du, daß sie gesagt hat, [ei-für Leute] würde er kennen? 'what, do you think that she has said e,- kind of people would he know?'

Extraction out of a constituent in the top spec-position of an embedded V2-clause is ungrammatical in German and Dutch, both, in case of extraction out of a fronted clause, as in (1.4b), or with splitting of a fronted D P (was-für-split), as in (1.4d). In German, extraction out of constituents in the middle field ( = the domain between the base position of the verb on the right and the V2-position on the left) - scrambled or un-scrambled - is fully grammatical, as (1.3) illustrates. Hence, it is safe to conclude that these positions generally cannot be analyzed as spec-positions or positions adjoined to F-projections. A fortiori, adverbials cannot be assigned to spec-positions. The facts follow if both adverbs and scrambled constituents are contained in the Vprojection (cf. Haider 1997). Fifth, VP-topicalization would have to be reanalyzed as topicalization of a functional projection containing the VP. This, however, is in conflict with a restriction on the topicalized constituent: It must not contain the trace of the finite verb. (15) a.

[Fast immer (e* ) irgenwann umfallen (ej)] könntej j a nur ein Auto« ej '[almost always sometimes fall-over (e^)] couldj indeed only one car'

b.

*[Fast immer (ej ) irgenwann um-e^] fielj ja nur ein AutOj ej '[almost always e^-over] fellj indeed only one car'

c.

[Immer wieder Fehler NO M unterlaufen] pp sind dabei auch Syntaktikern '[always mistakes happened] have with-this also (to) syntacticians'

If, as in (15.c) the nominative and a higher adverbial is contained in the fronted constituent, this constituent must be part of the cascade of functional projections and it must contain Agr-S, according to the minimalist logic of derivations endorsed by the positional approaches. The functional head Agr-S is a head the finite verb passes through on its way to the top functional head-position in a V2-clause. The fronting of a subtree with a trace across the antecedent of the trace results in a crossing violation. This can be read-off from (15.5b): If a finite verb leaves its base position and it is a particle-verb, the particle is stranded. Thus, the position of the trace can be identified. The ungrammaticality of the resulting constructions as in (15.5b) is evidence against any analysis that would have to posit a trace of the finite verb in the fronted extended V-projection.

Adverbials at the Syntax-Semantics Interface

59

Having cast enough doubt on the viability of an absolute positional approach to the syntax-semantic interface on the evidence of adverbials, let us backtrack and explore another road. The following section sketches a relative domain mapping approach to the interface.

3

A relative interface? - Mapping c-command domains

Syntactic structures partition a linguistic expression in hierarchically organized constituents. This, together with the lexically stored semantic properties of terminals, is sufficient input to the semantic construction algorithm. There is no need for a syntactic structure to be prefabricated to fit a semantic blueprint. Consider the following case: (16) a.

(Some) linguists often were treated (badly)

b.

They often treated (some) linguists (badly)

c.

(Some) linguists they often treated (badly)

As a bare plural, the term linguists in (16.b), compared with (16a,c), is interpreted differently. This is true for the indefinite DP some linguists, as well. Note, that the difference vanishes in (16.a), but not in (16.c), if we drop the frequency adverb. In (16.a,c), as it stands, the bare plural cannot be interpreted as denoting an arbitrary subset of linguists in the universe of discourse, an interpretation appropriate for (16.b). The adequate interpretation is, as has been frequently noted, rather the generic one: If one had the property of being a linguist, one was often (badly) treated (by them). Similarly for the indefinite DP: In (16.a,c), the approriate interpretation is a specific subset of linguist, while in (16.b) an unspecific subset suffices. In Diesing's (1992) account, this is a positional effect: The VP is the default domain of existential quantification. DPs outside VP cannot be existentially quantified in a default manner. This prima facie accounts for the difference between (16.a,c) and (16.b). The fact that in the absence of a frequency adverb, an indefinite subject can get an existentially bound interpretation called for an amendment: Subjects can be lowered into the VP-internal base position in some cases. Things get more complicated in a language like German. The various serialization options provide a good testing ground for Diesing's hypothesis, and for alternative accounts, as well. Consider the constrasts in (17): (17) a.

Er hat ja immer Bücher aus dem Fenster geworfen (exist.) 'he has PRT always books out of the window thrown'

b.

Er hat ja Bücher immer aus dem Fenster geworfen (gen.) 'he has PRT books always out of the window thrown'

c.

Er hat ja bereits Bücher [fast aus allen Fenstern] geworfen (gen.) 'he has PRT already books almost out of all window thrown'

60

Hubert

d.

Haider

Er hat ja bereits [fast aus allen Fenstern] Bücher geworfen (exist.) 'he has PRT already almost out of all windows books thrown'

The fact that the existential reading is not available in (17.b,c) is problematic for Diesing's (1992) view, since she takes the position of the particle ja to be indicative of the left boudary of the VP, and hence the domain of existential closure. 4 (17.c) is particularly instructive: The bare plural object and the P P occur in their normal order, namely object-before-directional PP. In this order, an existential interpretation is available, as (17.a) illustrates. The crucial difference is the occurrence of quantificational elements in the P P in (17.c,d). In both cases, elements with quantificational force are involved: The frequency adverb, as in (17.a,b), is quantifiying over properties of events. In (17.c,d), the quantified P P induces the same effect, namely quantification over properties of events: The events of throwing something out of windows are such that almost all windows under discussion have been involved in such events. The general situation is this: The semantic representation of a clause with a quantifier expression consists of at least three parts, the quantifier, the restrictor and the nuclear scope. The relative position of a quantified adverbial in syntax partitions the clause. The c-command domain of the adverbial quantified expression is mapped onto the nuclear scope, the elements c-commanding the quantified expressions go into the restrictor part. Existential closure applies in the nuclear scope part. Therefore, an indefinite DP c-commanding the quantified expression is mapped into the restrictor part and cannot be subject to existential closure. So, an indefinite outside the c-command domain of an adverbial quantifier gets a different reading, independent of the 'derivational history', that is, in a scrambled position as well as in a non-scrambled position. Note that the verb order in (18) is identical. It is the difference in the quantificational force of the manner adverbial that matters. (18) a.

Er sollte ja Fische auf schwäbische Art zubereiten (exist.) 'he should PRT fish in Suabian style prepare'

b.

Er sollte ja Fische auf mehr als 5 Arten zubereiten (gen.) 'he should PRT fish in at more than 5 styles prepare'

Let us proceed now to the general question of how the interface conditions influence the interpretation of adverbs and thereby their distribution. The claim is this: 4

There is also evidence that Diesing's splitting hypothesis is not appropriate for even smaller parts of VP: (a)

daß ja heute [Privatpatienten, [der Oberarzt e, operiert]] (generic) 'that PRT today the senior consultant private patients operates'

(b)

daß ja heute [der Oberarzt Privatpatienten operiert] (existential)

Scrambling within the domain of modal particles has the same effect as scrambling across. What is crucial is not the VP in the sense of a maximal projection but rather the "minimal argument complex" (MAC), that is the minimal V-projection that contains all arguments. Diesing's account covers VO-languages without scrambling, though, because in these languages VP = MAC. Scrambling, in my view is adjunction to VP. Hence the domain of scrambling is the VP, and the domain of scrambling contains the MAC. MAC is the domain of existential closure of indefinites.

Adverbials at the Syntax-Semantics

Interface

61

Adverbials are modifiers of different types of semantic expressions, like process denotations, event denotations, or propositions. These are ordered by inclusion relations (19.a): The process-denotations include the event denotations and these include the propositions. In other words, event denotations result from process denotations (= verbal denotation) by integrating material of the VP. Propositions as situated event denotations in terms of time and place result from event denotations. (19) indicates the principal interface organization: P-related adverbial material must c-command E-related adverbial material, which c-commands L-related adverbial material, that is, adverbs that relate to the (verbal) lexical element as its denotation is that of a process/state. Typical representatives of P-related adverbials are the so-called sentential adverbs. E-Related adverbials comprise among others temporal and local adverbials. The prototypically L-related adverbials come from the class of manner-adverbials. The examples in (19) illustrate typical distribution restrictions that follow from what has been said above: (19) a.

*Er hat alle [völlig [häufig verwirrt]] 'he has everyone completely often confused'

b.

*Er hat [häufig [angeblich was verwechselt]] 'he has often allegedly something confused'

c.

*Er hat sie [völlig [angeblich entnervt]] 'he has her completely allegedly enervated'

[Häufig verwirrt](= often confused) is an E-predicate since the frequency adverbial quantifies over properties of events. Therefore the adverb völlig (= completely) cannot be interpreted as an L-related adverbial, that is, as a predicate applied to verbal denotations like processes or states. If, however, it would be interpreted as an Lrelated adverb, the frequency adverb in its sope could not be interpreted as an Erelated adverbial. Hence, either move is blocked. In (19.b), a sentential adverbial is in the scope of an E-related adverbial. Since sentence adverbials are applied to propositions, the E-related adverbial cannot be mapped properly anymore. Analogous considerations apply to (19.c): Either the manner adverbial is applied correctly, then the sentence adverbial cannot occur in its scope, or the sentence adverbial is applied correctly, then the manner adverbial cannot be applied correctly. (20) a. b.

semantics: Proposition C Event C Process/State structure: ['P-related' ['E-related' ['L-related']]]

The isomorphic relation between (20.a) and (20.b) constitutes an interface effect, if the interface is strictly monotonic and incremental: Semantic expressions are structured by the inclusion relation, and when the syntax-semantic mapping has reached a higher type this forbids going back to a property of a lower type: (21) Interface criterion Syntactic c-command domains are mapped monotonically on incrementally structured semantic type-domains. 5

62

Hubert

Haider

(22) Corollary: If an adverbial Aj is included in the c-command domain of an adverb Aj which is an attribute of a predicate m , and Aj is an attribute of a predicate P„, then the type of P„ must be equal to or lower than the type of P m . In the next section, the implications of (21) and (22) will be analyzed in more detail with respect to the properties of temporal adverbials. Before we focus in on this particular type of adverbial, let us consider one more example that highlights the effects to be captured by (22): (23) a.

Er hat wen/alle [fast geheilt] (Rl, *R2) 'he has someone/all almost cured'

b.

Er hat fast [wen/alle geheilt] (*R1, R2)

c.

Er hat wen/alle fast [auf Anhieb geheilt] (*R1, *R2) 'he has someone/all almost at the first go cured'

d.

Er hat wen/alle fast kostenlos geheilt (*R1, *R2) 'he has someone/all almost freely cured'

e.

Fast, hat er (ej) [wen/alle [(?e,) geheilt]] (?R1, R2) 'almost has who/all cured'

(24) a. b.

Rl = Reading 1: the process of curing was almost successful R2 = Reading 2: a curing-event almost happened to come about

If, as in (23.a), the adverb immediately precedes the verb, the process reading is modified. If, on the other hand, the adverb is attached to a projection of V, the narrow process-modifying reading is not available, but only the event-modifying reading (as in (23.b), unless the projection is itself a projection of a verb with only a process modifying adverb. In this case the denotation resulting from the application of the c-commanded manner adverb and the verbal denotation is modified, as in (23.c,d). Topicalization of the adverb in (23.e) is preferrably construed with a short chain. The objects in (23) are chosen as either a scope sensitive element, that is a quantifier, or an indefinite w-pronoun. The latter element is unlikely to scramble. In both cases, the resulting denotations are similar, however, with respect to process versus event reading. If, on the other hand, the adverb c-commands an argument, this means that semantically a property of an event structure is already present, hence the adverb must be modifying a property of an event structure. In all cases, the function of the adverb is identical, but its position may vary. The preliminary conclusion is: Syntactic principles of clause structuring determine what is a possible position for adverbials. The actual occurrence of an adverbial in one of these positions is a function of their semantic type and the interface condition. Unlike arguments, adverbials do not have a unique structural base position. An apparent but misleading impression arises if an adverb, due to its constrained semantic 5

If the semantic representation is chosen to be a discourse representation structure (DRS; cf. Kamp & Reyle 1993), the mapping of c-command domains on DRS-domains becomes particularly transparent in notation: 'box-command' relations in DRT are a subset of c-command relations on the syntactic structure.

Adverbials at the Syntax-Semantics

Interface

63

type, is excluded from potentially accessible structural position because of the interface condition. In this case, the adverb cannot occur anywhere else but in a narrow domain. To call this a base position would be misleading. Not the position is unique, but only the domain. An illustration for this type of distributionally restricted adverbials is the difference between selected manner adverbials and free ones: The selected manner adverb must be in the licensing domain of the verb, that is, to its left in a VO-language like English. (25) a.

*They (have) badly treated him

b.

They (have) treated him badly

c.

"They (have) frugally lived

d.

They (have) lived frugally

e.

They very much missed her

f.

They missed her very much

Semantic differences between the preverbal and the postverbal position of manner adverbials, as those described by Quirk et al. (1986:595f, § 8.109), are a function of the difference of the c-command domain of the manner adverbial: (26) a.

He dissected the animal completely (i.e. into all the described parts)

b.

?

c.

*He completely [j< has dissected the animal]

He completely dissected the animal (i.e. completed the dissection task)

For the narrow process modifying reading of (26.a), the manner adverbial should not c-command the whole VP, because in the case of (26.b), where the whole VP is ccommanded by the manner adverbial, the denotation of the VP, that is, the event, not the process, is characterized as complete with respect to animal-dissecting. (26.c) is ungrammatical, because the position of the adverb is too high to be interpretable as a manner adverb: The functionally extended VP, that is I' in (26.c), is mapped on an event-denotation, hence it cannot be modified by a process-modifying element. So, (26.c) is ruled out on semantic grounds, incured by the interface-criterion.

4

Temporal adverbials, for the sake of demonstration

What is commonly called a temporal adverbial covers at least three different temporal relations: specification of time reference, of duration, and, of frequency. Time reference contributes to situating a specific event, whereas duration and frequency are specifications of properties of events. If an adverbial contributes to situating an event, it must c-command the finite verb or its trace, since the finite verb is the exponent of tense. For event-specification, on the other hand, it suffices that (a link of the chain of) the head that provides the event-variable, which is in most cases the main verb,

64

Hubert

Haider

is c-commanded by the adverb. This difference becomes immediately evident in VPtopicalization patterns. (27.a) is unacceptable, because the time-reference adverbial does not c-command the finite verb. (27.b) is marginal. The adverbial must not be interpreted as a time-reference adverbial, for the same reason as in (27.a). (27) a.

* [Heute morgen gratuliert] hat er 'today (in the) morning congratulated has he

b.

[Vor Sonnenaufgang gratuliert] hat er 'before sunrise congratulated has he'

c.

[Minutenlang applaudiert] hat er 'for-minutes applauded has he'

d.

[Stündlich gratuliert] hat er 'hourly congratulated has he'

The adverbial in (27.a) is deictic. If the adverbial P P in (27.b) is interpreted as deictic (= at the time before sunrise on the relevant day), the status of the clause should be as degraded as (27.a). It's improved status is due to reinterpreting the adverbial as a manner-like specification (before sunrise = in the darkness of the night). Duration (27.c) and frequency (27.d) are wellformed specifications for an event. The pattern in (27) is reconfirmed by sentential adverbials: (28) a.

* [Gratuliert, nachdem wir gegangen waren], hat sie 'congratulated, after we had left has she'

b.

?

c.

[Applaudiert, solange sie konnte] hat sie 'applauded eis long as she could has she'

d.

[Gratuliert, wenn jemand vorbeidefilierte] hat sie 'congratulates when someone marched-past has she'

[Applaudiert, ehe die Sonne aufging] hat sie 'applauded, befor the sun rose has she'

The VP-topicalization data are robust and indicative of the c-command requirements at the interface. Time reference adverbials share the c-command requirement with P-related adverbials, as in (29), for different reasons though: The time reference adverbial needs to c-command the exponent of tense, the P-related adverbial needs to c-command the functional features that bring about the functional closure of the V-projection. In both cases this happens to be the tensed verb. (29) a.

* [ Vermutlich gratuliert] hat sie zu früh 'presumably congratulated has she too early'

b.

Gratuliert hat sie vermutlich zu früh 'congratulated has she presumably too early'

Adverbials that are to be interpreted as predicating over propositions as in (29), or situated events as in (28.a,b), need to occur in positions c-commanding the finite verb or its trace in order to guarantee a successful mapping between syntactic structure

Adverbials at the Syntax-Semantics

Interface

65

and semantic structure. So they are excluded from topicalized VPs. Event related adverbials can occur in fronted VPs since the verb that supplies the event structure occurs in the VP as well. There is an additional domain effect that is reflected in the serialization options of a temporal adverb with respect to other material in the clause. The adverb precedes or follows an object, depending on the relation between adverbial and object: (30) a.

Er befürchtete Streiks (nur) im Februar 'he feared strike (only) in February

b.

Er befürchtete (nur) im Februar Streiks 'he feared (only) in February strikes'

c.

Er hoffte auf Regen (auch) im Mai 'he hoped for rain (also) in May'

d.

Er hoffte (auch) im Mai auf Regen 'he hoped (also) in May for rain'

In (30.a,c), the temporal specification relates to the respective intensional objects, that is, events concerning the object. In (30.b,d), however, the temporal adverbial relates to potential events in general as objects of fear or hope at a given time. The fact that there are different relations involved becomes perspicuous in sentences with two occurrences, each relating to a different aspect: (31) a.

[Regen erwartet im Mai] hat er schon im März 'rain expected in May has he already in March'

b.

In March, he foresaw rain in May

Let us clarify now a crucial issue for structure and serialization, namely the question as to whether there are structural base positions for adverbials in general and temporal adverbials in particular. As briefly mentioned above, there are two possible approaches: Adverbs may be assumed to be individually tied to unique positions (cf. for instance Cinque 1997), or, adverbs occur freely in syntactically available positions, with an interface requirement: The particular occurrence is subject to interface effects, depending on the semantic property the adverb carries. These effects are effects of relative domain inclusion requirements only. Frey k, Pittner (1998), in a thorough study on the syntax of adverbials in German, opt for a position compromising between the two approaches mentioned: Adverbs, in their opinion, occur in various adjoined positions, but with specific primary targets of adjunction corresponding to the respective semantic function of an adverb. Temporal adverbs are assumed to be adjoined to the MAC (cf. Fn.4), whereas local adverbials are assumed to occur in a MAC-internal position. This predicts a basic structure with local adverbials lower than temporal ones. The following contrast seems to support this assumption. (32) a.

"In Peters, Garten ist [Peter/Mr]« der Hund entlaufen 'in Peter's garden is PeterDAT the dog escaped' In Peters garden, the dog escaped Peter

66

Hubert b.

In Peters* B ü r o hat der Chef [ P e t e r ¿ c c ] ¿ kritisiert 'in Peter's office has the boss Peter criticized' In Peter's office, Peter criticized the boss

c.

A n Peters, G e b u r t s t a g hat m a n Peter¿ im B ü r o gefeiert 'on Peter's birthday has one Peter in the office celebrated' On Peter's birthday, they celebrated Peter in his office

Haider

Frey & Pittner interpret the contrast b e t w e e n (32.a) and (32.b) as evidence for different base positions: T h e base position of the local adverbial must be lower t h a n the dative object in (32.a) and higher t h a n the a c c u s a t i v e object in (32.b): T h e trace of the fronted element m u s t be in or out of t h e c - c o m m a n d d o m a i n of the respective binder in order t o derive the principle-C effects in (32.a). Their conclusion is appropriate as t o the positioning of the trace, but the position is not a base position in the technical sense of t h e only p o s i t i o n at which the adverbial could b e generated in the clause. Consider t h e following contrast: (33)

a.

* [In Heides¿ Version der Geschichte]^ hat sie¿ einen Fehler e¿ entdeckt 'in Heide's version of the story has she a mistake dedected'

b.

[In HeideSj Version der Geschichte]^ hat dj sie¿ die Tat nicht b e g a n g e n 'in Heide's version of the story has she the deed not committed'

In (33.a), the adverbial specifies the location of the d e n o t a t i o n of t h e object. So, the object should c - c o m m a n d the adverbial, which is the case w i t h t h e trace indicated in (33.a). In (33.b), however, the P P is a frame adverbial for the proposition. Frame adverbials c - c o m m a n d the material that belongs t o t h e proposition, so their position is one t h a t c - c o m m a n d s all other material in t h e clause. 6 A n a l o g o u s considerations apply t o temporal adverbials EIS well: (34)

a.

[An Heides¿ Geburtstagjj hat e j sie¿ sich sehr g u t amüsiert 'at Heide's birthday has she amused herself very well'

b.

*[An HeideSj Geburtstag]j hat sie¿ ein E r d b e b e n e j prophezeit 'at Heide's birthday has she an earthquake prophecied' She has prophecied an earthquake (to happen) on H's birthday

T h e temporal specification in (34.a) is interpretable as a specification of the entire event, hence its trace is higher t h a n the pronoun position. So, principle C is not violated. In (34.b), the adverbial can be interpreted as s p e c i f y i n g t h e point of t i m e 6

Frame adverbials, unlike 'lower' adverbials, can occur between C° and fronted pronouns: (a)

daß unter solchen Umständen man/er keine Wahl hat 'that under such circumstances one/he no choice has'

(b)

daß (* leider/" heute/" sorgfältig) man/er Sätze analysierte 'that (unfortunately/today/carefully) one/he sentences analyzed'

As Frey & Pittner point out, the elements in the position between C° and the fronted pronouns in German must be possible topics in order to fuction as frame adverbials. The adverbials in (b) cannot serve as topics, so they cannot be interpreted as frame adverbials.

Adverbials at the Syntax-Semantics

Interface

67

of the prophecy to become true. In this case, the adverbial is a specification of the prophecied event. So the trace of the adverbial is in the c-command domain of the object, and a fortiori also of the subject, whence the principle C violation. If, however, the adverbial in (33.b) is interpreted as situating the event of prophecying something, the position of the trace can be posited higher than the pronoun. Under this interpretation, (33.b) is wellformed with respect to binding. The corresponding non-fronted orders are given in (35). Note that (35.b) and (35.c) are not synonymous. In (35.c) the adverbial specifies the time of the phrophecying event, in (35.c) the timing of the prophecied event: (35) a.

daß an Heidesj Geburtstag sie* sich sehr gut amüsiert hat 'that at Heide's birthday she herself very well amused has'

b.

* daß sie* ein Erdbeben an Heidesj Geburtstag prophezeit hat 'that she an earthquake at Heide's birthday prophecied has'

c.

daß an Heidesj Geburtstag siej ein Erdbeben prophezeit hat 'that on Heide's birthday she an earthquake prophecied has'

Frey and Pittner (1998: sect.5) point out that the class of adverbials that do not trigger a principle-C effect is identical with the class of adverbials that, according to Browning (1996), are base generated in front of the subject position in English and block the that-¿-effect (cf. 36.a). (36.b) is ungrammatical because extraction out of a clause with a constituent moved to the position preceding the subject is illicit (cf. 36.c): (36) a.

W h o j do you think that in Ben'Sj office ej is a real dictator?

b.

*Whoj do you think that in Ben'si office e» lay on his desk?

c.

* W h o do you think that this book he gave?

Next, it is worth emphasizing that the interface properties of adverbials interact with the interpretation of indefinite DPs, in particular bare plurals. The effects described by Diesing (1992) are accounted for immediately, without particular structural commitments, if the domain of existential closure is taken to be the event-internal domain. The following examples are chosen to contain a bare plural sandwiched between a particle and an event-specifying adverbial. According to Diesing, the particle marks the left VP-boundary and the V P is the domain of existential closure. This is a problematic assumption for various reasons. First, scrambling within this domain produces the same effects as scrambling across the particle: (37) a.

daß ja Hunde jemand füttern müßte 'that PRT dogs someone feed should'

b.

daß Hunde j a jemand füttern müßte 'that dogs PRT someone feed should'

Secondly, a constituent scrambled across the particle would be scrambled out of the VP and therefore should become opaque for extraction, which is clearly not the case, as (38) demonstrates, or (13) above:

68

Hubert

Haider

(38) Welche*/diese¿ Tür hat [damit e¿ zu öffnen]j (denn/ja) jeder e_, versucht? 'which/this door has [with it to open] (PRT) everyone tried' These data do not pose any problem if it is assumed that the domain of scrambling is the VP and that positions within VP in the canonical licensing direction of the verbal head are transparent for extraction. The syntactic domain that corresponds to the semantic domain of existential closure is the MAC, that is, the minimal V-projection that contains all A-positions discharged. In (37), the scrambled object has left the MAC, hence the loss of the existential reading for the bare plural. In (39.a), the existential reading of the bare plural preceding the temporal adverbial is lost in those cases in which the adverbial predicates over the event and hence over the semantic domain of existential closure. Since the bare-plural is outside the ccommand domain of the adverbial it is outside the domain on which the c-command domain is mapped in the semantic representation. In (39.b) the existential reading is appropriate. Reordering the object and the adverbials as in (39.a) would cancel this reading: (39) a.

Hat denn wer Vögel heute morgen/täglich/3

Stunden lang beobachtet?

'has PRT someone birds today (in the) morning/daily/ for 3 hours watched?'

b.

Es hat ja wer heute morgen/täglich/3

Stunden lang Vögel beobachtet

'there has PRT someone today in the morning/daily/for 3 hours birds watched'

If the temporal event specification is such that the temporal adverbial is interpreted distributively across potential instantiations of event types, as in (40.a), and the bare plural is interpreted distributively as well, the resulting interpretation amounts to the interpretation with existential closure on the indefinite DP. (40) a.

Hat denn wer Linguisten* an deren¿ Geburtstag photographiert? 'has PRT someone linguistsi on theiri birthday photographed?'

b.

?

Es hat ja wer an dereni Geburtstag Linguisten* photographiert 'there has PRT on their; birthday linguists; photographed?'

The sketch of a relative interface descibed above is incomplete without a remark on the implications for the clause final occurrencies of adverbials in VO-languages. The postverbal order in VO-languages like English is a mirror image of the preverbal order in VO-languages like German: Definition 4.1 English: V - respect -< process -< space -< time (s. Quirk et al. 1986:§ 8.87) German: time -< space -< process -< respect - V The order relations indicated above are illustrated in the examples under (41): (41) a.

She has worked [on her hobby] [with great care] [in the garden] [the whole time] [today]

b.

Sie hat [heute] [die ganze Zeit] [im Garten] [mit großer Sorgfalt] [an ihrem Steckenpferd\ gearbeitet

Adverbials

at the Syntax-Semantics

Interface

69

The standard account - layered adjunction to VP on the right - captures both the mirror image order and the compositionality requirements for the semantic interpretation. But it is insufficient in various other respects. For instance, it cannot account for binding relations as illustrated in (42): (42) a. b.

He photographed every j linguist with great care on his* birthday He photographed the menj with great care on each others* birthday

This paradoxical situation has vexed Pesetsky (1995) into proposing a non-standard solution: Each clause is to be assigned a dual structure, a layered one and a cascading one. The layered structure is to account for the mirror image compositionality, the cascading one is to provide the c-command relations for binding. Rich enough though this system may be descriptively, it does not provide satisfactory insights into the grammatical causalities that give rise to the puzzling combination of properties in a VO-projection. There are of course various other non-standard attempts that one might imagine. Suffice it to say that the two cardinal properties - mirror image word order and ccommand sensitive left-to-right dependencies - still are wanting satifactory theoretical modellings.

5

Outlook and Conclusions

The syntax and semantics of adverbials has not been a specific focus of interest of grammar theory in the past. Adverbials have played a circumstantial though crucial role in various contexts concerning the distribution of verbs (e.g. cascading of functional projection) on the one hand and the distribution of arguments on the other hand (e.g. 'scrambling' in Dutch). But a comprehensive account of the structural and interpretative characteristics of adverbials in a crosslinguistic perspective is still missing. The theoretical position sketched in this paper argues for a specific strategy: The specific patterns of the distribution and interpretation of adverbials are epiphenomenal: They are the composite result of a set of structural conditions (What is a possible position for an adverbial in the given clause structure?), an interface condition, and a set of semantic conditions (What is a possible interpretation for an adverbial in a given syntactic context?). The interface condition relates the syntactic and the semantic structures by a monotonic mapping relation. In an OV-langage like German, adverbials are adjuncts of VP. The complex serialization options can be shown to be the result of the interaction of simple parameters: actual syntactic c-command domains of the adverbial, semantic type of the constituent which the adverbial is adjoined to, possible semantic type of the adverbial. The analysis of the structure of postverbal adverbials in VO-languages is left on the agenda.

70

Hubert

Haider

References Alexiadou,A. 1994. Issues in the Syntax of Adverbs. Unpubl. Dissertation, University of Potsdam. Browning,M.A. 1996. CP-recursion and 'that-t'-Effects. Linguistic Inquiry 27:237-255 Chomsky,N. 1995. The Minimalist Program. Cambridge, Mass.: MIT Press Cinque, G. 1997. Adverbs and Functional Heads. A Cross-linguistic Perspective. Ms. Univ. Venezia. (1999 Oxford, Oxford University Press) Diesing,M. 1992. Indefinites. Cambridge, Mass.: MIT Press Frey,W. & K.Pittner. 1998. Zur Positionierung der Adverbiale im deutschen Mittelfeld. Linguistische Berichte 176: 489-534 Haider,H. 1997. Projective Economy. On the Minimal Functional Structure of the German Clause. In: W.Abraham & E. v.Gelderen eds. German: Syntactic Problems - Problematic Syntax. Tübingen: Niemeyer. Kamp,H.& U.Reyle. 1993. Prom Discourse to Logic. Dordrecht: Kluwer. Pesetsky,D. 1995. Zero Syntax. Experiencers and Cascades. Cambridge,Mass.: MIT Press Quirk,R. & S.Greenbaum & G.Leech & J.Svartvik 1986 4 . A Comprehensive Grammar of the English Language. London: Longman.

Klaus

Baumgärtner

halten und die Einheit des verbalen Aspekts

1

Eine Fallstudie

Es kommt sicherlich nicht oft vor, daß die originelle Semantik eines einzelnen verbalen Prädikats imstande ist, das gewohnte System der Aspekte (oder Aktionsarten), das allgemein als wohlfundiertes Instrument gilt, an zentralen Punkten in Frage zu stellen. Das deutsche Verb halten ist offenbar ein bisher unbeachtetes einschlägiges Exempel dieser Art. Die logischen Eigenschaften seines Gebrauchs scheinen jedenfalls genügend Anlaß dafür zu geben, bestimmte Prinzipien in der Theorie der Aspekte neu zu überdenken. Das Folgende sollte dazu als eine Fallstudie verstanden werden. Sie besitzt in den empirischen Fakten des Verbs halten ihren einzigen (und einzigartigen) Angelpunkt und versucht vor allem abzuklären, wie dessen aspektuale Probleme mit den hierher gehörigen Auffassungen seit Dowty 1979 bis hin zu Kamp und Reyle 1993, Bartsch 1995, ter Meulen 1995 oder Pustejovsky 1996 in Einklang zu bringen sind. Wenn halten für das Folgende als empirischer Angelpunkt fungiert, ist es zunächst ganz nützlich, einen Blick auf die neutralen linguistischen Daten zu werfen (zumal keines der großen Wörterbücher des heutigen Deutsch in der Lage ist, den spezifischen semantischen Ansatz des Verbs und die spezifischen Abhängigkeiten seiner verschiedenen Verwendungen auch nur annähernd präzise vorzuführen 1 ). Wir lassen hier alle gesonderten Entwicklungen beiseite und begnügen uns für die Zwecke dieser Studie mit der folgenden lexikalischen Grundsystematik: (1)

(2)

(3)

a.

Das Seil hält. Der Nagel hält. Der Ballon hält. Das Wetter hält.

b.

Emil hält das Seil [straff]. Anna hält Otto das Buch vor die Nase.

a.

Der Wagen hält. Die Karawane hält. Der Ballon hält über dem See.

b.

Peter hält den Wagen. Der Mann im Tor hält [den Ball].

a.

Anna hält Schafe, Ziegen und Hühner.

b.

Otto hält [sich] einen Pudel, eine Zeitung, eine Diva und einen Stadtrat.

Die beiden grundlegenden semantischen Strukturen sind (1) und (2), und zwar jeweils in (.a) intransitiver oder (.b) transitiver Version. Für beide (.a) kann offenbar Referenz auf einen 'Zustand' - entsprechend 'state' bei Dowty (siehe dafür und für 1

Die großen Wörterbücher zeigen frappierende Ähnlichkeiten in der - teilweise kaum nachvollziehbaren - Paraphrase, Zerlegung, Einteilung, Anordnung und Streuung der syntaktischen und semantischen Grundfakten des Lemmas über (rein durchgezählt) 29, 34 oder 28 Einträge und sind außerstande, ein systematisch verständliches Bild zu vermitteln (AKADEMIE-WB. III, 1967; BROCKHAUS-WAHRIG III, 1981; DUDEN-GROSSWB. III, 1993).

72

Klaus

Baumgärtner

weitere Terme: Dowty 1979,180-186) - angenommen werden, jedoch so, daß bei (l.a) genau derselbe Zustand bereits im Vorbereich gegeben, dagegen bei (2.a) der Vorbereich durch einen Vorgang besetzt sein muß. Das läßt sich hier fürs erste begrifflich und zunächst spekulativ festhalten mit den Etiketten: durch (l.a) wird 'ZustandsAndauer' bezeichnet, durch (2.a) 'Zustands-Beginn'. Die transitiven Varianten (.b) referieren dann auf die 'Handlung', die es jeweils erlaubt, Fälle der Art (.a) herbeizuführen: das ergibt für (l.b) 'Kausation' von 'Zustands-Andauer', dagegen für (2.b) 'Kausation' von 'Zustands-Beginn' - bei Dowty gegeben als Gegensatz von 'agentive interval Stative' vs. 'agentive change of State'. Es versteht sich, daß Zustandssätze unter (.a) nicht etwa die formal eindeutigen Ergebnisse von Handlungssätzen unter (.b) repräsentieren. Sätze wie (3), die Sprechern des Deutschen beim Erwähnen von halten wohl als erste in den Sinn geraten, vermitteln die ursprüngliche Basis von (1) und (2). Es ist interessant zu verstehen, daß das Verb historisch auf die komplexe Semantik von 'hüten' und 'weiden' zurückzuführen ist, wie sie das Beispiel (3.a) noch heute konserviert, und zu Anfang, wie Dialekte zeigen, geregelte Abfolgen bezeichnet haben muß der Art: 'Hintreiben' (an den Weideplatz), dort 'Anhalten' und 'Verbleiben' (zum Abgrasen), zuletzt 'Weitertreiben' 2 . Die Alternativen 'Zustands-Beginn' vs. 'Zustands-Andauer' sind also nichts als generalisierte Abstraktionen der beiden zentralen Komponenten 'Anhalten' und 'Verbleiben' aus dem vielschichtigen 'Hüten'-Abfolge-Muster mit Tätigkeitsaspekt (Dowtys 'activity'), das in (3.a) noch zuweilen ablesbar, in (3.b) schon lange verblaßt ist. Man betrachte aber: (4)

a.

Peter hält den Wagen, [situationell: der gerade erst auf ihn zugerollt war]

b.

Peter hält den Wagen, [situationell: der sonst wieder weiterrollen würde]

c.

Peter hält den Wagen, den er soeben gehalten hat.

Beispiel (4.a) ist eine Handlung der Art (2.b) und vom Effekt 'Zustands-Beginn', Beispiel (4.b) ist eine Handlung der Art (l.b) und vom Effekt 'Zustands-Andauer'. Nehmen wir an, die Intervalle der Ereignisse zu den Aussagen (4.a) und (4.b) folgen dicht aufeinander: dann gilt (4.c), und die beiden Abstraktionen halten repräsentieren in ihrem unmittelbaren Kontakt auch heute noch korrekt den komplexen Kern ihres historischen Ursprungs. Die Wortsemantik von halten erscheint beim ersten Hinsehen durchaus nicht ungewöhnlich. Wenn es nicht das erstaunliche Verhalten geben würde, das den Anlaß dieser Studie bildet, könnte man das Verb als einen bloßen nächsten Kandidaten für die schon klassischen Prototypen von (un) verursachtem Zustands-Wechsel verbuchen3 2

3

Mögliche Rückführung bieten ig. Wurzeln mit genereller Bedeutung 'treiben, zur Bewegung antreiben' (Pokorny I, 1959, 548) und 'drehen, sich herumbewegen' (Seebold 1970, 248). Gebote oder Verbote der Art 'das Vieh an gewissen Stellen, zu gewissen Zeiten [nicht] halten sollen/dürfen' im Sinne von offiziellen geregelten Aktivitäten sind vielfach belegt z.B. bei Fischer (FISCHER WB. III, 1911, 1083). So könnte der Zusammenhang der drei Aspektstrukturen des Modells open (zur Systematik der lexikalischen Einträge siehe zuletzt Bierwisch 1996, 241 ff.) suggerieren, halten vom Fall (1) einfach unter dieses Paradigma zu ziehen. Paradigmen noch näher an halten

halten und die Einheit des verbalen Aspekts

73

und im übrigen auf sich beruhen lassen. Tatsächlich existiert nun aber dieses Verhalten. Tatsache ist darüber hinaus, daß sich die beiden systematischen Varianten (1) und (2) gerade in den Erscheinungen und Konsequenzen ihrer aspektualen Problematik in beträchtlicher Weise unterscheiden. Sie werden deshalb von hier an gesondert als die Alternativen haltertA (mit 'Zustands-Andauer') und haltenß (mit 'Zustands-Beginn') behandelt. Fürs erste wird ausschließlich die sehr komplexe Problematik von halten a im Vordergrund des Interesses stehen und in den nächsten Abschnitten - beginnend mit: »Das Erwartungsproblem« - die Diskussion bestimmen. Die ganz anders gelagerten Probleme der Aspektstruktur der Alternative haltenß werden dagegen zunächst nur kurz angedeutet. Die später anstehende Problematik von haltenß läßt sich im ersten Zugriff in dieser Weise skizzieren: der Aspekt 'Zustands-Beginn' ist wohl zunächst problemlos der elementaren aspektualen Deutung des Verbs beginnen anzuschließen, wenn auch mit Auflagen (zumal es für beginnen gerade nicht die Norm ist, daß damit ein Zustand eröffnet wird und dann auch noch mit Vorgang im Vorbereich); ernste Probleme ergeben sich aber daraus, daß diese aspektuale Struktur auch in ganz andere Strukturen übergehen kann, ohne daß dafür jedesmal eindeutige Auslöser im Kontext verantwortlich zu machen wären. Man vergleiche die folgenden Paraphrasen zu oben (2.a) Der Wagen hält: (5)

a.

Der Wagen steht da und fährt nicht (egal, was vorher war und danach ist).

b.

Der Wagen beginnt zu stehen (indem er aufhört zu fahren).

c.

Der Wagen steht da (nachdem er gefahren ist und bevor er weiterfährt).

Die Paraphrase unter (5.b) kann normalerweise als der Standard von haltenß angesehen werden. Der Originalsatz (2.a) erlaubt aber in den meisten Fällen, zugleich im Sinn von (5.a) oder (5.c) interpretiert zu werden. Das heißt, daß das Verb hier zwischen 'Zustand', 'Zustands-Beginn' (nach Vorgang) oder 'Zustand' (zwischen Vorgängen) changiert. Im letzteren Fall (5.c) wäre es dann nicht länger an die aspektuale Erklärung von beginnen anzuschließen, sondern an die von unterbrechen. Es erübrigt sich fast zu sagen, daß im Unterschied dazu ein typisches Inchoativ für Zustände mit Vorgängen im Vorbereich, wie es z.B. erwachen darstellt, nur eine Paraphrase wie die unter (5.b) gestattet, nämlich: Beginnen von Wachsein oder Aufhören von Schlafen, jedoch keine andere. Die Problematik von haltenß besteht vor allem in der Behandlung dieser gleitenden Aspektstruktur. Wir kommen gegen Ende darauf wieder zurück.

könnten fälschlich gesehen werden in der formalsemantischen Explikation der intransitiven statischen Positions- und Kontaktverben stehen usw. und der transitiven Positionsverben stellen usw. durch Kaufmann (Kaufmann 1995, 98 ff., 121 ff.). Im folgenden wird sich zeigen, daß diese Muster gerade nicht übertragbar, sondern indirekt vorauszusetzen sind.

74

2

Klaus

Baumgartner

Das Erwartungsproblem

Das erste der Probleme, die durch halten (genauer, bis auf weiteres: durch halteriA) aufgeworfen werden, kann als Erwartungsproblem angesprochen werden. Es besteht darin, daß mit halten durchweg die Erwartung des Nicht-Haltens verbunden ist. Das bedeutet, etwas ausführlicher formuliert, daß jede Situation, in der irgendeine Aussage mit halten über einen Gegenstand X - also: das Halten oder Nicht-Halten von X getan werden soll, die Bedingung erfüllen muß, daß die Erwartung besteht, daß X jedenfalls nicht hält. Betrachten wir dazu die Beispiele (6): (6)

???

a.

Wir betraten den Pavillon.

An der Wand hielt ein Gemälde.

b.

An der Wand des Pavillons hing ein Gemälde. *Das Gemälde hielt.

c.

An der Wand hing ein Gemälde an einem dünnen Faden. Es hielt.

Beim Satz (6.a), der referentiellen Einführung des Gegenstands Gemälde, ist fast unmöglich, halten zu gebrauchen (mit der marginalen Ausnahme eines bestimmten, selber schon halb konventionellen Ironiestils, in dem sich sogar noch mit diesen Mitteln andeuten läßt, daß das Gemälde aus wer weiß was für Gründen nahe daran zu sein scheint, von der Wand zu fallen). Eine korrekte Einführung des Gegenstands liegt vor in (6.b) mit dem Ausdruck hing statt hielt, das halten im unmittelbaren Anschluß daran ist jedoch schlichtweg falsch gebraucht ohne weiteren Kontext, der die nötige Gegen-Erwartung etabliert. Exakt das leistet nun (6.c) mit der Phrase an einem dünnen Faden. Ein dünner Faden, alltagsgemäß verrechnet mit dem gewöhnlichen Gewicht von Gemälden in Pavillons, liefert durchaus die erforderliche Disposition für die Erwartung, daß dieser Gegenstand wohl nicht mehr lange hängen zu bleiben droht. Der Witz der Aussage (6.c) ist insofern: Entgegen allem Anschein bleibt das Gemälde hängen. Es sind derartige (durchaus auch äußerst indirekte) Problem-Indizien, die über den richtigen oder falschen Gebrauch des Verbs entscheiden und selbst für fragwürdige Fälle die Existenz des Erwartungsproblems bestätigen. Es ist interessant, bereits hier kurz festzuhalten, was für verbales 'Material' bei dem Versuch benutzt wird, die Zia/ien-Situationen (6.a) bis (6.c) passend zu verdeutlichen. Oben sind halten mit hängen+bleiben und nicht+halten mit herabfallen gleichgesetzt worden. Sobald man darangeht, Bedeutungspostulate für das Verb aufzustellen, wird es demnach am ehesten wohl einerseits durch die Klasse der lokalen Zustandsverben (verknüpft mit bleiben), andererseits durch die große Klasse der lokalen Vorgangsverben (dominiert von der Negation) umschrieben. Zugleich ist abzusehen, daß die jeweiligen Prädikate, die in Umschreibungen von Aussagen mit halten auftreten können, durch die Charakteristik der Gegenstände (und ihrer Umgebungen) bedingt sind, von denen gerade gesagt wird, daß sie halten (oder nicht): Wenn im gedachten Pavillon sich etwa eine Skulptur befindet, von der sich vorhersagen läßt, daß sie nicht hält, dann löst das schwerlich die Vermutung aus, daß sie von der Wand fällt, vielmehr eher, daß sie umkippt oder wegrutscht. Strukturelle Beziehungen - oder enger: Selektionen - dieser Art werden als Abrundung und Endpunkt der ganzen Diskussion im Kapitel über »Das Vorwissensproblem« wieder aufgegriffen. Vermerkt sei hier nur als äußerste sachliche Ambiguität, daß ein Satz wie (l.a) Der Ballon hält bedeuten kann, daß der

halten und die Einheit

des verbalen

Aspekts

75

Ballon sich lediglich nicht losgerissen hat, oder aber, daß er schlichtweg nicht geplatzt ist. Folglich kann das Verb (in seiner intransitiven Version), was die Gegenstände an der Subjektposition angeht, entweder allein deren Lokalisierung betreffen oder aber deren komplette Existenz. Das Moment der Gegen-Erwartung ist für viele Vorkommen von halten alles andere als plausibel und erscheint dann für die Erklärung des Prädikats nicht nur entbehrlich, sondern als Verfälschung. Das gilt weniger für die intransitiven, dagegen für viele der transitiven Verwendungen. Der Eindruck täuscht jedoch, und es macht die Sache nicht besser, daß dieser Täuschung bereits die Lexikographen erliegen, die der Meinung sind, man könnte halten systematisch über einen Ansatz mit Verben wie ergreifen oder gar mit den Händen fassen beschreiben 4 . Die Bedeutung des Verbs sollte deswegen (und auch unabhängig davon) noch prinzipieller erfaßt werden als lediglich mit ProblemIndizien, die nur auf mehr oder minder indirekte Weise zur Identifikation der jeweiligen Gegen-Erwartung und semantischen Grundfunktion hinführen. Man vergleiche etwa: (7)

a.

Hilf mir mal bitte und halte mir doch hier mal kurz die Zange.

b.

Du hast die Zange nicht gehalten, du hast sie bloß genommen und hingelegt.

c.

Du hast die Zange nicht gehalten, du hast sie zu Boden hängen lassen.

d.

Du hast die Zange nicht gehalten, du hast sie in die Höhe gestreckt.

Die Sätze (7.b) bis (7.d) sind offenbar hinreichend begründete, also selber kaum wieder abweisbare Zurückweisungen von drei möglichen Handlungen in Reaktion auf die Bitte (7.a). (Es ist keine Frage, daß man sie mühelos durch beliebige weitere, deskriptiv noch differenziertere Fälle ergänzen könnte). Sie lassen bereits auf intuitive Weise erkennen, daß halten auf gar keinen Fall mit Verben wie greifen, fassen, packen oder etwa stellen, legen, setzen in eine Reihe gebracht werden darf, bei denen durchgängig ein Zustands-Wechsel mit mehr oder minder reichem Sachgehalt vorliegt. Eine sorgfältige Analyse von (7.a) im Vergleich mit den Fehlhandlungen (7.b) bis (7.d) erweist vielmehr: halten eines X heißt grundsätzlich, dieses X in eine Disposition 5 zu bringen (oder, falls X sich schon in dieser befindet, dafür zu sorgen, daß es in ihr bleibt), - und zwar in die jetzt eben gewollte Disposition, die jedenfalls keine der üblichen natürlichen Lagen von X darstellt und von dem Gegenstandstyp, zu dem X gehört, normalerweise nicht von selber eingenommen wird. Genau in dieser nichtnatürlichen Disposition besteht nun das spezifische Risiko des X, in die nächstbeste seiner natürlichen Lagen zurückzufallen, ist mithin die Gegen-Erwartung begründet, daß X zu ebendiesem Rückfall tendiert. Soweit die lexikalische Semantik überhaupt imstande ist, ihre isolierten Objekte in sprachlichen Definitionen zu erfassen, können 4

5

Dieser Ansatz, durchweg gebraucht von den großen Wörterbüchern (vgl. Fußnote 1), verfehlt leichfertig die abstrakten Spielräume natürlicher Bedeutungen und unterschlägt einfach Verwendungen wie den Stift zwischen den Lippen halten, das Zelt mit dem Rücken halten, die Türe mit dem Fuß halten (oder mit dem Zeh oder Knie oder Bauch oder Kopf). 'Disposition' steht hier alltagssprachlich für 'geeignete, zweckdienliche Lage oder Anordnung' und nicht für den LogikbegrifF, der Dispositionsprädikate erklärt. Die nachfolgende 'nicht-natürliche Disposition' - generell zu verstehen als 'passende günstige, nicht dauernd sichere, von Wechsel bedrohte, also risikobehaftete Ruhelage' - läßt sich aber zum logischen Dispositionsbegriff durchaus in Beziehung setzen.

76

Klaus

Baumgärtner

die beiden letzten Sätze als die zentrale Definition von halten angesehen werden. Sie umfaßt neben der transitiven Version des Verbs auch ohne weiteres die intransitive; man braucht dafür nur abzuziehen, daß transitive riskante Dispositionen von X auf dauernder Kausation beruhen. Aufgrund dieser Definition wird nun klar, daß von Satz (7.a) nichts Substantielles besagt wird außer eben rein dispositional, daß die Zange geeignet parat 'gehalten' werden möge, und daß die Sätze (7.b) bis (7.d) dann nur beschreiben, daß diese Disposition der Zange nicht wie gewollt geleistet worden ist. Geeignete Modifikationen lassen natürlich auch die nicht-natürlichen Lagen solcher Beschreibungen von Fehlhandlungen wie vollgültige Erfüllungen von Haltens-Bitten erscheinen; man vergleiche nur: (7)

c'/d'.

Du hast mir die Zange perfekt zu Boden / in die Höhe gehalten.

Der Effekt ist allerdings wieder, daß solche Aussagen über gelungene Dispositionen wie (7.c'/d') nun nicht mehr die speziellen deskriptiven Sachgehalte von hängen+lassen oder strecken vermitteln; den ursprünglichen Sachgehalt sichern dann gerade noch die Direktivphrasen zu Boden oder in die Höhe. Freilich ist das präzis der Punkt, an dem sich leicht der falsche Eindruck einstellt, daß beim transitiven Gebrauch des Verbs das Moment von Gegen-Erwartung zunehmend verlorengeht: Gewissermaßen in Proportion zum Umfang der Argumentstrukturen - bis zu oben (l.b) Anna hält Otto das Buch vor die Nase - entwickeln sich immer reichere Situationsstrukturen, die geeignet sind, die beinahe leere Bedeutung von gelungener Disposition (und sonst nichts) zu überdecken und damit den mitgesetzten Risiko-Aspekt scheinbar verschwinden zu lassen. Tatsächlich aber wird die Definition des Verbs in keiner Weise eingeschränkt: Falls nicht andauernd - vom Agentiv oder den Umständen - dafür gesorgt wird, daß die einmal erreichte nicht-natürliche Disposition weiter besteht, ist halten für die betreffende Situation, ob nun deskriptiv arm oder reich, schlichtweg nicht zu gebrauchen. Das Erwartungsproblem ist im heutigen Deutsch keine exklusive Eigenschaft von halten (und den meisten seiner Derivationen). Es scheint wenigstens - natürlich abgesehen von erwarten selber, bei dem dies offenliegt - noch zwei weitere Elemente zu geben, nämlich ausbleiben und unterbleiben, deren Bedeutung im wesentlichen auf der Erwartungskomponente beruht 6 . Bei beiden kommt sogar eine deutliche Verallgemeinerung, wenn nicht Verschärfung der Problematik zum Tragen, und zwar dadurch, daß die Erwartung hier auf eine überhaupt erst eintretende Existenz gewisser Entitäten zielt. Die Entitäten sind notwendig Ereignisse, indem als Subjekte nur Ausdrücke zulässig sind, die sich selber wieder auf Sachverhalte beziehen. Im übrigen unterscheiden sich beide Verben noch nach der Möglichkeit von nicht-intentionaler oder intentionaler Deutung dieser Ereignisse beim Subjekt-Argument. Betrachten wir die Fälle (8): (8)

6

a.

Ich glaube, wir bekommen eine Sturmflut.

b.

Die / Eine Sturmflut ist ausgeblieben / * unterblieben.

Gut bestätigt wird das von den ausführlichen, logisch orientierten Feldanalysen durch Schumacher et al.; das Metten-Feld resümiert ausbleiben von X mit der (im Formalen gesondert motivierten) Paraphrase: 'es hört wider Erwarten nicht auf, daß es X nicht gibt' (Schumacher 1986, 95 ff.); halten ist nicht verzeichnet.

halten und die Einheit des verbalen Aspekts

77

c.

Ich glaube, wir bekommen eine Erklärung.

d.

Die / Eine Erklärung ist unterblieben /

?

ausgeblieben.

Die Sätze (8.b/d) sind zunächst nur passende Feststellungen zu den Annahmen der Sätze (8.a/c). Es versteht sich, daß ein Diskurs schwerlich mit ihnen eröffnet werden kann. Sie reagieren insofern, jedenfalls mit definiter Nominalphrase, genau nach dem Muster (l.a) Das Seil hält. Ebenso wie dort für den vorerwähnten Gegenstand Seil verstanden werden soll, daß er entgegen der Erwartung hält, soll nun für die vorerwähnt eingeführten Ereignisse Sturmflut oder Erklärung verstanden werden, daß es zu ihnen, entgegen der Erwartung, effektiv nicht gekommen ist. Mit indefiniter Nominalphrase dagegen (ein darf also nicht zählen) tritt ein kritischer Unterschied zutage: Während (l.a) Ein Seil hält nun einen eindeutig generischen Satz darstellt, werden mit (8.b/d) lediglich die indefiniten Vorerwähnungen von (8.a/c) erneut in Wiederholung gebraucht. Das bedeutet: Ein bloß erwartetes Ereignis, das nicht eingetreten ist, muß nicht unbedingt als 'das' erwartete Ereignis erscheinen, sondern kann durchaus auch weiterhin als 'ein' beliebiges erwartetes Ereignis behandelt werden. Sätze mit positivem ausbleiben oder unterbleiben bezeichnen somit im Extremfall, wenn sie indefinit formuliert sind, nur unbestimmte Erwartungen der Art, daß es eigentlich im normalen Lauf der Dinge zu gewisser Zeit eine Existenz gewisser Sachen wie Sturmfluten oder Erklärungen hätte geben müssen. Daneben (und gegen die definite Annahme eines irgendwie schon fixierten Ereignisses) besagen sie überhaupt nichts. Das Ergebnis der Diskussion in diesem Kapitel ist zweifellos, daß der Semantik solcher Verben wie halten (oder auch ausbleiben oder unterbleiben) eine implizite Komponente der Erwartung zugeschrieben werden muß. Das Verhalten der Daten erlaubt außerdem den Schluß, daß diese Komponente nur zutreffend behandelt wird, wenn sie den Status einer Präsupposition bekommt: Im einen Fall gibt es eine Erwartung bezüglich X, das X könnte nicht halten - gleichgültig, ob X dann hält oder nicht, und im anderen Fall gibt es eine Erwartung bezüglich X, das X könnte eintreten gleichgültig, ob X dann ausbleibt oder nicht. Die beiden Typen von Ausdrücken arbeiten dabei, wie erwähnt, mit konträren logischen Beziehungen: Während halten eine Erwartung der Nicht-Existenz (oder wenigstens der Aufhebung der Lokalisation) voraussetzt, setzen ausbleiben und unterbleiben umgekehrt eine Erwartung des Eintritts von Existenz voraus. Darüber hinaus unterscheiden sich beide Typen noch grundlegend darin, daß sie als Prädikate fungieren entweder für Gegenstände oder Ereignisse. Damit dürften sämtliche wichtigen Bestandteile für eine adäquate Beschreibung dieses Konzepts versammelt sein. Allerdings streben wir dafür in diesem Kontext ohne eigene Basis keine spezielleren formalen Notationen an, zumal die Phänomene, die gerade hier anstehen, im gewohnten verbreiteten Standard von Semantischen Formen verbaler Prädikate nur schwer zu behandeln sein dürften. Wenn man versucht, die diskutierte aspektuale Problematik von halten zunächst mit dem geringsten formalen Aufwand explizit zu machen, gelangt man (hier nur für die intransitive und lokale Variante) zu dem folgenden Ausdruck (9):

78

Kl aus Baumgärtner

(9)

Gegeben seien eine Zeitlogik mit den Intervallen Äußerungszeit to, Referenzzeit t j und t j (sodaß ti < i j ) , ein Gegenstand X und eine Menge von Dispositionen {..., Dfc,...} derart, daß die Satzform halten(X), geäußert zur Zeit t 0 in bezug zur Zeit t j , wahr ist genau dann wenn a.

Djt(X) zur Zeit t¿,

b.

erwartbar ist zur Zeit t j , daß -> Dfc(X) zur Zeit t j ,

c.

Dfc(X) zur Zeit t j .

Dabei beschreibt Bedingung (9.a) den Vorbereichs- und Anfangszustand des Haltens von X und Bedingung (9.b) die Erwartung des Nicht-Haltens von X. Bedingung (9.c) umfaßt als Fortdauer von (9.a) den eigentlichen Haltens-Begriff; sie ist für das Erwartungsmoment allein nicht erforderlich, dient also dazu, (9) zu komplettieren zu einer vollen Bedeutungsangabe (von dieser Variante) des Verbs. Die existentielle Variante ist entsprechend zu formulieren. Die transitive Variante erhält man, indem man (9.c) in der üblichen Form unter Rückbezug auf (9.a) als verursacht beschreibt. Ein mit gleichen Mitteln formulierter Ausdruck für ausbleiben würde sich weit problematischer gestalten, insofern die Funktion dieses Verbs j a gerade darin beruht, daß die Existenz von Ereignissen X nicht von vornherein, sondern allein für den Bereich der Erwartung angesetzt werden kann. Das wichtigste Merkmal von (9) ist die durchgehende Gültigkeit der Erwartungskomponente. Die Bedingung (9.b) gilt wie auch der Anfangszustand (9.a) in jedem Fall. Sobald halten in einem Satz unter Negation gelangt, springt zwar die Bedingung (9.c) auf den Gegenwert entsprechend der Erwartung unter (9.b), was aber gerade nichts weiter heißt, als daß die Erwartung erfüllt ist: (9)

c'.

D fc (X) zur Zeit t j .

Die Erwartungsbedingung (9.b) hingegen bleibt weiter (sozusagen: nun erst recht) eben dieselbe. Dabei ist dennoch eine spezielle problematische Differenz möglich. Die Erwartung des Nicht-Haltens erstreckt sich eher darauf, daß ein Ereignis eintritt, das den Ausgangszustand von X beendet, egal was dann folgt. Die Bedingung (9.c') dagegen ist nur so zu verstehen, daß X effektiv zurückfällt (natürlich, weil das anders nicht geht, durch irgendeine Art von Ereignis) auf einen neuen Zustand, dem selber keine nicht-natürliche Charakteristik zukommt. Es könnte folglich sein, daß sich die Referenzen von (9.b) und (9.c') nur partiell auf ein und dieselbe Sachlage beziehen. Was alles für Sachlagen von Nicht-Dispositionen abgedeckt werden, bedürfte daher noch der Klärung. Als problematisch könnte die Formulierung von Bedingung (9.b) erscheinen: 'erwartbar' ist leicht zu kritisieren als ein Ausdruck, der das eigentliche Moment der Bedingung, also daß es Subjekte gibt, die etwas erwarten, nur verschleiert. Erwartungen haben jedoch in erster Linie - anders als Hoffnungen oder Befürchtungen, die vor allem in individuellen mentalen Zuständen bestehen - einen bestimmten generellen, nämlich graduell objektivierbaren Status. Es scheint daher nicht nötig und wäre wohl nicht einmal sinnvoll, die Erwartung in (9.b) etwa den in äußerungssituationen involvierten Personen (Sprechern, Adressaten und Zuhörern) oder Personengruppen zuzuteilen, und das vielleicht noch in wechselnden Konstellationen. Stattdessen müßte man

halten und die Einheit des verbalen

Aspekts

79

hier eher an eine Paraphrase denken, die ausführlich darlegt, daß sich Erwartungen bilden, indem aus der Gesamtheit der jeweils vorliegenden Tatsachen und Umstände die richtigen Schlüsse gezogen werden. Doch genau dies soll der Ausdruck 'erwartbar' besagen. Am kritischsten verhält es sich vermutlich mit dem Begriff der Disposition, sobald er eine derartig grundlegende Funktion übernimmt wie gerade im Ausdruck (9). Hier liegt der Einwand nahe, daß Dispositionen in der vagen Umschreibung, mit der sie notwendigerweise eingeführt worden sind, wohl kaum für elementare Einheiten der Semantik geeignet sind. Der Einwand ist zunächst in formaler Hinsicht leicht auszuräumen: für die Klasse der Dispositionen gibt es gar keine andere Begründung, als daß sie eine echte Teilklasse der Einheiten sind, die von vornherein durch den Aspekt Zustand beschrieben werden müssen. Zu den Dispositionen als denjenigen Einheiten, von denen sich herausgestellt hat, daß sie als nicht-natürlich, risikobehaftet, von Wechsel bedroht usw. zu gelten haben, gibt es dann eine komplementäre Teilklasse der restlichen natürlichen Einheiten - und beide zusammen bilden, jedenfalls für die Explikation von halten, die Klasse der Zustände insgesamt. Die Problematik verschiebt sich damit jedoch letztlich in die Wahrnehmung und Definition zum jeweiligen einzelnen Gebrauch. Die speziellen Zustände der Disposition, ohne die es halten in dieser Bedeutung überhaupt nicht gäbe, lassen sich nicht etwa sortieren als unabhängige Sachlagen aller beliebigen Gegenstände und lassen sich auch nicht, soweit man die nötigen Differenzierungen im Gegenstandsbezug einbezieht, etwa aufsammeln als Sachverhalte von Umständen, die unabhängig beobachtbaren Kriterien unterliegen würden. Stattdessen sind verwickelte Interpretationen der involvierten Personen anzunehmen, wie sie z . B . die Fälle (6) und (7) belegen sollen. Der Dispositionsbegriff erhält damit einen Status nahe dem Erwartungsbegriff in der zuvor diskutierten Weise. Das wirft die Frage auf, ob nicht beide Einstellungen zum Sachverhalt - seine Einschätzung als nichtnatürliche Disposition und die Erwartung seines Nicht-Haltens - als zwei Seiten ein und derselben Faktenbewertung aufzufassen sind. Womöglich wäre es dann unsinnig anzunehmen, daß ein Anfangszustand, der sowieso zur Bedingung hat, Disposition zu sein, eine eigens auf ihn gerichtete Erwartung des Nicht-Haltens überhaupt noch determiniert. Die beim Ausdruck (9) unterstellte Parallelität von Erwartungskomponente und Komponente der Tatsachenreferenz wäre damit erheblich gestört. Bei der Alternative wären demgegenüber die Bedingungen von (9) als (noch problematische) Folgebeziehung betrachten etwa der Art, daß etwa die Erwartungsstruktur (9.a-b) einer Aktualitätsstruktur (9.a-c) vorausgeht, obwohl beide auf demselben Anfangszustand (9.a) beruhen. W i e eine endgültige Lösung für (9) aussehen könnte, hängt allerdings auch noch vom Weiteren ab und bleibt somit an dieser Stelle offen.

80

3

Klaus

Baumgartner

Das Vorwissensproblem

Das zweite der Probleme, die halten (bis auf weiteres wiederum: haltenA) für das Aspektsystem hervorruft, wird am besten als Vorwissensproblem beschrieben. Es besteht darin, daß die Bezeichnung der Sachverhalte, die das Verb mit seinen Argumenten liefert, nicht in üblicher Weise geschlossene Sachverhalte darstellt, sondern defizitäre Sachverhalte, die durch spezifische Informationen komplettiert werden müssen. Diese Eigenart ist bereits verschiedentlich zutage getreten. Die Schwierigkeit ist vor allem, daß das Verb allein nicht das mindeste hergibt für die konkrete Art der Vorzustände, deren Andauer es bezeichnet. (Von einem Vorkommen mit halten kann man auf gar keinen Fall auf einen Vorzustand wieder mit halten selber schließen, aber leider eben auch auf sonst kein konkretes Verb). Unter den üblichen Prozessen der Semantik findet sich nun offenbar keiner, mit dem sich diese prinzipiell offenstehenden Informationen schließen ließen. Da es also nicht einmal darum gehen kann, etwa gewisse Präsuppositionen auszuwerten, scheint die Problemlage nur noch vom pauschalen VorwissensbegrifF angemessen erfaßt zu werden. Es ist leicht zu sehen, daß dieses Vorwissensproblem ebenso wie zuvor schon das Erwartungsproblem hinausläuft auf die jeweilige Konkretisierung der Dispositionen in der Ausgangsbedingung (9.a). Hinzu kommt nun jedoch, daß das Vorwissendefizit, das zunächst für die Vorzustände des Verbs zu registrieren ist, in derselben Weise auch seine Nachzustände betrifft. Der Zustands-Wechsel, der von der Erwartungsbedingung (9.b) angezeigt und bei Negation des Verbs gemäß (9.c') behauptet wird, schafft ein Vorwissensproblem in bezug auf den Nachbereich, das dem für den Vorbereich in prinzipieller Weise gleicht. Der Unterschied liegt allein darin, daß das Wissensdefizit sich nunmehr auf die jeweiligen konkreten Vorgänge erstreckt, die aus dem erwarteten oder effektiven Nicht-Halten resultieren. Wir konzentrieren uns fürs erste auf die Problematik im Vorbereich, erweitern sie aber dann auch wieder auf den Nachbereich. Der Ansatzpunkt des Problems ist die Feststellung vom Anfang, daß bei halten offenbar ebenderselbe Zustand, auf den das Verb referiert, bereits im Vorzustand bezeichnet sein muß. Genauer zu betrachten ist somit das semantische Verhältnis zwischen dem aktualen Bereich, den die Bedeutung des Verbs ausmacht, und diesem Vorbereich. Das Problem läßt sich am besten konkretisieren durch Rückbezug auf das klare aspektuale Verhalten der beiden Verben öffnen und schließen, die im allgemeinen als Prototypen für die systematischen Alternativen des Aspektsystems gehandelt werden. Vergleichen wir dazu die folgenden teilweise sequentiellen Sätze: (10) a.

Die Tür ist offen / öffnet sich / wird von Hinz geöffnet.

b.

Die Tür ist geschlossen / schließt sich / wird von Kunz geschlossen.

c.

Die Tücher sind trocken / trocknen / werden von Anna getrocknet.

d.

Die Tür hält, wenn sie wie jetzt ein wenig offen steht.

e.

Die Tür hält nicht, auch wenn sie wie jetzt geschlossen ist.

In den Fällen (10.a) und (10.b) fungieren öffnen und schließen in ihren drei aspektualen Varianten, und zwar - (i) als Zustand ('state') und als solcher nicht weiter

halten und die Einheit des verbalen

Aspekts

81

spezifiziert, (ii) als Vorgang ('achievement'), der den Wechsel vom einen zum anderen Zustand anzeigt, und (iii) als Handlung ('accomplishment'), mit der die Kausation eines derartigen Vorgangs bezeichnet wird, agentiv oder nicht. All das ist bestens (und wohl zur Genüge) bekannt. Wichtig ist hier für diesen Kontext, daß sich (10.a) und (10.b) in Kombination als geschlossenes semantisches System verhalten: Die Negation der drei Sätze in (10.a) führt ohne Unterschied zum ersten Satz von (10.b) Die Tür ist geschlossen. Auf dieselbe Weise führen die Sätze (10.b) zum ersten Satz von (10.a) Die Tür ist offen. Die beiden Satzsequenzen erhellen damit ein simples Bezugssystem mit genau zwei Zuständen, nämlich offen und geschlossen. Deswegen gibt es hier keinerlei Probleme, für die Vorgangssätze und Handlungssätze jeweils die Präsupposition des Zustandssatzes mit dem Gegenwert anzusetzen: Ob die Tür sich nun öffnet (oder nicht) oder gar von jemand geöffnet wird (oder nicht) - die Tür ist im Vorbereich der Ereignisse (und ihrer Negation) in jedem Fall geschlossen. Es ist müßig zu spekulieren, ob diese Binarität als Eigenschaft der Gegenstände (also der Türen, Klappen, Riegel usw. oder der Dinge mit Türen, Klappen, Riegeln) oder als Effekt der Verben öffnen und schließen zu interpretieren ist. Ein Fall wie (10.c) läßt erkennen, daß schon eine nur leicht erweiterte Kategorie dafür ausreicht, Mehrdeutigkeiten im Verhältnis zum Vorbereich hervorzurufen: Ein Vorgangssatz oder Handlungssatz mit dem Lemma trocken erlaubt keinesfalls die eindeutige Präsupposition des Zustandssatzes Die Tücher sind naß, weil hier die Gegenposition wenigstens dreifach besetzt ist, nämlich mit den Adjektiven naß, feucht, klamm, mit denen sich in gradueller Weise abnehmende Nässe bezeichnen läßt; der Vorbereich könnte deswegen ebensogut mit der Präsupposition Die Tücher sind klamm besetzt sein. Man braucht hier nur an die Farbkategorie zu denken, deren Spektrum je nach Bedarf verschieden fein in Farbnamen aufteilbar ist, und dazu etwa an ein Prädikat wie schwärzen, um zu ermessen, von welcher Unbestimmtheit Präsuppositionen schon im Rahmen von Kategorien betroffen sein können. Schließlich ist an Sätzen wie (10.d) problemlos vorzuführen, daß sogar die vermeintlich klaren Sachverhalte von Türen manchmal mit Zustandsangaben zu verbalisieren sind - also: wirkliche Türen aus Zuständen zu öffnen oder zu schließen sind für die es überhaupt keine geregelten Kategorien und Gegenbereiche mehr gibt. Betrachten wir nochmals kurz die Fälle (10.a) bis (10.e) im Vergleich. Während für die vermeintlich prototypischen Sätze unter (10.a) und (10.b) zweifellos zutrifft, daß sie sich kontradiktorisch verhalten und damit wechselweise eindeutige Präsuppositionen liefern, sind Sätze wie die unter (10.c) lediglich imstande, Zustandssätze mit einer Variablen zu präsupponieren. Für deren Ersetzung wird verlangt, daß diese nicht durch beliebige Ausdrücke erfolgt, sondern durch ein Element des konträren Bereichs derselben Kategorie, die vom Satz (10.c) vorgegeben ist, - oder, in der Terminologie der hierfür zuständigen Adjektivsemantik: ein Element des antonymen Skalenabschnitts derselben Dimension (Bierwisch 1987, 108 ff.). Von einem allgemein gültigen Modell für das Aspektsystem kann danach sicherlich nicht mehr die Rede sein. Die beiden folgenden Fälle (10.d) und (10.e) bieten entsprechend - ganz abgesehen davon, daß sie anders als die bisherigen Fälle keine Vorgänge oder Handlungen betreffen - noch nicht einmal diese Möglichkeit, das Verb zumindest in die eigene spezifische Kategorie zu relationieren. Der Begriff der Disposition, der für halten anzusetzen ist, beschreibt ja gerade die Freiheit, auf Zustände zu referieren, die nicht

82

Klaus Baumgärtner

auf Alternativen und Gegenwerte von bestimmten Kategorien zu reduzieren sind. Der Fall (10.d) ist als ein exemplarischer Beleg für nicht-natürliche, also labile, nicht auf Dauer gesicherte Ruhelagen kaum besser denkbar: Eine Tür in einer gewissen gering geöffneten Stellung, die ständig das Risiko einschließt, endlich gänzlich zuzufallen oder aufzuspringen. Deskriptiv unpräzise Positionen sind eine Art Einladung für die Verwendung des Verbs. Auf eine stabil geöffnete oder geschlossene Tür wäre das Verb überhaupt nicht anzuwenden (möchte man meinen). Wenn das nun aber trotzdem geschieht wie in (10.e), dann bestätigt das bloß den Begriff der Disposition: Eine Tür, die droht, (bei korrekter Wortwahl) nicht zu 'halten', obgleich sie geschlossen ist, muß sich ganz einfach in der prekären nicht-natürlichen Position befinden, plötzlich aufzuspringen, zusammenzufallen, auseinanderzubrechen oder sonstwas. Beide Beispiele zeigen gerade im Vergleich zueinander, daß es zuletzt überhaupt nicht darauf ankommt, was für triviale oder kuriose Sachverhalte es objektiv sind, auf die das Verb bezogen wird, sondern allein, daß die Bedingung einer riskanten Disposition erfüllt ist. Das Resultat ist, daß vom Mechanismus des Modells öffnen und schließen, die Referenz für den Vorbereich zu fixieren, bei halten nichts bleibt als die allgemeinste Präsupposition des äußersten Kategorietyps, also gerade noch, daß die Referenz prinzipiell vom Aspekt 'Zustand' zu sein hat. Allerdings ist daraus nicht abzuleiten, daß das Verb dann in bezug auf den Vorbereich (strukturalistisch gesprochen) einfach als eine Art Archilexem für die Klasse der Zustandsverben betrachtet werden kann und deswegen keine weiteren Fragen mehr aufwirft. Das hierfür vielleicht naheliegende Argument, daß archilexemische Verben wie tun, machen, herstellen usw. in ähnlicher Weise beträchtliche Teilklassen der Handlungsverben vertreten, ohne daß dabei Vorwissensfragen entstehen, trifft nicht den Punkt. Tatsächlich existiert hier eine vielleicht intuitiv nicht unmittelbar greifbare, aber durchaus systematische Differenz. Ein Lexem wie machen, das unter Suspendierung aller einzelnen Varianten deskriptiven Eigenschaften (der Lexeme, die es vertritt) auf eine Bedeutung nahe der Aspektstruktur verallgemeinert erscheint, mag in vielen Fällen viel zu allgemein verwendet werden, erfüllt aber problemlos seine Funktion; dazu gehört, daß die Aspektlogik in keiner Weise beeinträchtigt ist: machen besagt, daß das X, das jemand macht, im Vorbereich nicht existent, aber im Nachbereich existent ist. Ein Lexem dagegen, dessen Bedeutung ausschließlich darauf beruht, die weiter bestehende Andauer eines Zustands zu bezeichnen, von dem jedoch sonst nichts bekannt ist, verliert jedwede Funktion und wird sinnlos, wenn keine Kontexte da sind, die den jeweiligen Zustand perzeptiv oder deskriptiv präzisieren. Das ist der Grund dafür, warum isolierte und auf Dauer kontextlose Sätze von der rudimentären Form Das Ding hält - deren Funktion sich in ihrer Aspekthülle erschöpft, nämlich: daß Andauer irgendeines gewissen Zustands besteht - in jedem Fall defizitär, also letztlich Nonsense sind. Durch die Diskussion ist bestätigt, daß die Bedeutung von halten in grundlegender Weise mit der klaren Problemlage eines Vorwissensdefizits (und nicht sonstiger Beziehungen) belastet ist. Die Frage ist danach, ob es womöglich sprachliche Strukturen gibt, dieses Defizit auf systematische Weise zu beheben, also das nötige Wissen für den Gebrauch des Verbs aus dem gegebenen Sprachwissen zu ergänzen. Bevor wir jedoch dazu übergehen können, sind zunächst einmal ganze Gruppen von Kontexten zu registrieren, in denen sich dieses Vorwissensproblem überhaupt nicht mehr stellt

halten und die Einheit des verbalen

Aspekts

83

oder aber dermaßen verletzt, verschoben oder übertragen erscheint, daß die bisher unterstellten einheitlichen Aspektstrukturen außer Kraft gesetzt sind. Betrachten wir die folgenden Kontexte (11) für eine genauere Analyse: (11) a.

Die Wand wird hier immer dünner, aber jetzt hält der Nagel.

b.

Anna hielt Otto ihr neues Buch zugeklappt mit dem Titel vor die Nase.

c.

Die Tür steht offen - halte sie doch bitte, wenn der Sturm kommt.

d.

Der Haken hält sicher ewig, nachdem du ihn diesmal angeschraubt hast.

e.

Paula hat den Pudel den ganzen Weg lang an der Leine gehalten.

Beginnen wir mit der Tatsache, daß das Vorwissensproblem von bestimmten Kontexten ganz zum Verschwinden gebracht wird, nämlich dann, wenn es gar keinen Vorbereich mehr gibt, der mit Vorwissen spezialisiert werden müßte. (11.a) ist hierfür ein Beleg aus dem intransitiven Bereich, (11.b) aus dem transitiven Bereich. Bei intransitiven Sätzen sind durchweg komplexere Sachverhalte anzunehmen. Der Fall (11.a) ist, soweit er mit abweichender Interpretation, also ohne präsupponierten Vorzustand verstanden werden soll, überhaupt nur möglich am genauen Verlaufspunkt (oder am Endpunkt) eines solchen Prozesses: der Nagel wird sukzessive in die Wand geschlagen ohne schon festzusitzen, sitzt jetzt eben gerade fest, worauf sich die Aussage von (11.a) bezieht (und wird danach unter Umständen weiter in die Wand geschlagen). Im Zusammenhang solcher Prozesse und Aussagen sind nun offenbar keine Vorzustände mehr präsupponiert, die eine Präzisierung verlangen würden, mit Ausnahme der etwas absurden Konstruktion, hier für den Gebrauch von halten auf irgendeine physikalistische Weise den 'Halte'-Punkt des Nagels als höchst flüchtigen Ausgangszustand vorauszusetzen. Falls hier überhaupt Vorbereiche zum Verb gehören, bietet sich noch am ehesten an, Präsuppositionen von unspezifischen Ereignissen (Handlungen, Tätigkeiten usw.) anzunehmen, die mit Zustands-Beginn enden oder unterbrochen sind. Das würde aber heißen, daß sich haltenA (in solchen Kontexten) auf wundersame Weise als aspektuale Parallele von haltenß herausstellt. Beim transitiven Gebrauch ergeben sich prinzipiell dieselben Sachverhalte: Eine Aussage ( l l . b ) - ganz gleich, wie man die Nominalphrasen oder small sentences der Struktur regulär reduziert oder ergänzt - kann auf keinerlei Vorzustand hin gedeutet werden, suggeriert vielmehr bestenfalls eine vorlaufende deskriptive Handlung der Sorte Anna STRECKT Otto das Buch vor die Nase, die aber keine Präsupposition mehr darstellt. Transitive Fälle der Struktur ( l l . b ) sind sogar, anders als die intransitive Version, völlig konkurrenzlos in dieser Interpretation. (Deswegen steht der Fall (ll.c) hier lediglich im Kontrast zu ( l l . b ) und gegen die Vermutung, daß die transitive Version des Verbs die Präsupposition von Vorzuständen prinzipiell verwehrt, was aber, wie man sieht, durchaus nicht der Fall ist). Das Ergebnis ist, daß für halten in keiner seiner Verwendungen eine sichere Garantie für Präsuppositionen existiert: Als faktives Verb fungiert es nur meistens und gehört insofern in die allgemeine negative Diagnose von Reis (Reis 1977, 158 ff.). Das Ergebnis könnte daher, nebenbei bemerkt, auch von Bedeutung sein für die generelle Geltung der formalen Bedingung (9.a). Im Ausdruck (9) steht ein und derselbe Anfangszustand (9.a) für die Erwartungskomponente und für die Aktualitätskomponente des Verbs. Er hat als Präsupposition, was das Erwartungsmoment angeht, nach

84

Klaus Baumgärtner

wie vor uneingeschränkte Gültigkeit: bei keinem der Fälle (ll.a-c) stehen etwa die risikobehaftete Disposition und die Gegen-Erwartung zum Halten infrage. Dieselbe Präsupposition eines Zustands existiert aber, wie eben vorgeführt, durchaus nicht mit verläßlicher Garantie für die Tatsachenreferenz. Das könnte die Überlegungen stützen, den Bedingungen unter (9) auch eine stark veränderte Organisation zu geben. Die Erscheinungen bis zu (ll.c) vermitteln dabei noch lange nicht die volle Flexibilität der Standardmuster der Aspektlogik (hier bei halten, wohl genauso im allgemeinen). Die folgenden Fälle (ll.d) oder (ll.e) sind eher zufällig gewählte Belege aus einer größeren Vielfalt abweichender Strukturen. So scheint der Satz (ll.d) nur mit Referenz auf einen doppelten Sachverhalt erklärbar zu sein. Jedenfalls ist er kompatibel mit dem Sachverhalt, daß der Gegenstand Haken wenigstens schon einmal befestigt war und nicht gehalten hat, aber nun erneut sicherer befestigt ist und deshalb eigentlich nicht länger als riskanter Fall beurteilt werden muß; das Verb wird somit angewandt auf eine vorzeitige Vergleichssituation und deren Erwartungsergebnis in einer späteren Situation, die selber schon gar keine Grundlage mehr bietet für den Gebrauch des Verbs. Der Satz (ll.e) scheint dagegen den Modellfall darzustellen für alle möglichen Verschränkungen des Halten-Aspekts mit weiteren Aspektstrukturen. Jedenfalls referiert er auf das parallele Vorkommen der beiden Ereignisse Paula hält den Pudel an der Leine und Paula und der Pudel [laufen] den Weg, die zwar beide intakte Aspekte zeigen, aber natürlich eine formale Koordination verlangen: entweder dominiert der eine Aspekt den anderen, was etwa heißen könnte, daß der Halten-Aspekt im Vorgangsaspekt unendlich feinkörnig mitläuft, oder es kommt zu einem kontextbewirkten Aspekte-Wechsel etwa gemäß den Vorgaben von Pustejovksy für reguläre Prozesse von 'Coercion' (Pustejovsky 1995, 105 ff.). Die korrekte Lösung solcher Fälle kann hier (und wohl noch für eine ganze Weile) nicht das Thema sein. Es muß aber grundsätzlich registriert werden, daß einheitliche Aspektmuster wie ZustandsAndauer oder Zustands-Beginn, wie sie unsere Diskussion von Anfang an bestimmen, angesichts der enormen empirischen Gelegenheiten zu ihrer Überschreitung nur als grobe vorläufige Idealisierung zu akzeptieren sind. Wir kommen damit auf die Frage zurück, ob Möglichkeiten bestehen, das Vorwissensdefizit des Verbs mit bestimmten systematischen Strukturen des Sprachwissens zu beheben; die Kategorie Zustand etwa des Vorbereichs bestimmt noch nicht die Klasse der möglichen Elemente. Damit ist wohlgemerkt nicht gemeint, was die Sprecher und Adressaten alles an Mitteln nutzen können, um das allfällige Wissensdefizit beim Gebrauch des Verbs auszugleichen. Ein Beispiel wie (10.d) hat gezeigt, wieweit dies durch passende Deutung der Wahrnehmung des Kontexts erfolgen kann. Beispiele wie (6.c) oder (10.e) führen vor, wie sich verbale Kontexte, vom Sprecher geeignet angelegt, mit der richtigen Verrechnung bestimmter kritischer Problem-Indizien vom Adressaten auswerten lassen. Die Fragestellung für das folgende ist wesentlich bescheidener. Zwischen dem absolut defizitären Gebrauch Das Ding hält und den unabsehbaren alternativen Verbalisierungen, die sich nutzen lassen, um beliebiges Vorwissen beim Gebrauch von halten zu umschreiben, besteht die Möglichkeit, primäre verbale Kontexte zu isolieren, aus denen sich ermitteln läßt, welche generelle Charakteristik den Defizitstrukturen des Verbs schon mittels Elementen aus der direkten Nachbarschaft des Lexikons gegeben werden kann. Die nächstbesten Kontexte dieser Art sind zweifellos die Gegenstände an den kritischen Positionen von Subjekt und Objekt in den

halten und die Einheit des verbalen

Aspekts

85

Argumentmustern des Verbs. Wir begnügen uns hier mit dem Subjekt im intransitiven Muster. Vergleichen wir die folgenden Fälle: (12) a.

Der Ballon auf unserem Dach hat nicht gehalten.

b.

Die Kiste drüben im Gang hat nicht gehalten.

c.

Das Seil über den Bach hat nicht gehalten.

d.

Die Rosen von dir haben nicht gehalten.

e.

Das Wetter gestern hat nicht gehalten.

Sachverhalte wie in den Sätzen (12), insbesondere die zu (12.a) bis (12.d), die konkrete Einzel-Objekte betreffen, werden interpretiert auf der Basis der abstrakten konzeptualen Objektmuster, über die wir offenbar kognitiv unabhängig verfügen. Die Begründung der betreffenden Theorie, die das strukturalistische Konzept von Selektionsrelationen ablöst, indem sie es auf eine eigenständige Grundlage stellt, ist Lang im Anschluß an Bierwisch zu verdanken (Lang 1987, 296 ff.; Bierwisch 1987). Die Interpretation der Sachverhalte, in denen sich die Gegenstände der Fälle (12) befinden können, wird demnach hauptsächlich durch die Raumkonzepte - daneben sicher auch durch Zweckkonzepte und ähnliche Konzepte - gesteuert, die den Bedeutungen der sprachlichen Ausdrücke korrespondieren. Weis die möglichen Vorzustände angeht, kann von beliebigen Dingen zuerst einmal gesagt werden, daß sie sich irgendwo befinden (oder, bildhafter, irgendwo ruhen). Zur Diskussion steht nunmehr aber nicht allein das Vorwissen in bezug auf Bedingung (9.a). Das Verständnis gerade der in (12) gegebenen Negationen wird keinesfalls erreicht ohne zusätzliches Wissen über die konkreten Nicht-Haltens-Ereignisse, die den Gegenständen der Fälle (12) geschehen sind. Von diesem Vorwissensdefizit, das die Erwartung (9.b) oder die Negation (9.c'), in bezug auf den Nachbereich des Verbs hervorruft, war schon oben die Rede. Ein Aushilfswissen zum Vorbereich und zum Nachbereich zusammen liefern nun für Gegenstände bei halten doppelte verbale Paradigmen folgender Art. Wenn etwa ein kugelförmiges Objekt Ballon vorliegt, so kann dieses: auf dem Dach liegen, am Dach hängen, über dem Dach stehen oder schweben, auf dem Dach angeseilt sein usw., zum anderen: platzen, aufreißen, zusammenfallen, sich losreißen, wegfliegen, abstürzen usw. Für quaderförmige Objekte Kiste gelten: stehen, liegen, auch hängen, zum anderen: platzen, aufspringen, zusammenbrechen, wegrutschen usw. Ein Objekt Seil kann hängen, gespannt sein, dann platzen, reißen, abreißen usw. Das Objekt Rose wird, wenn es nicht hält, am wahrscheinlichsten welken, kann aber auch brechen, knicken, usw. Wesentlich ist für Gegenstände generell, daß sie, um körperlich halten zu können, zuerst einmal ganz sein, heil sein oder in Ordnung sein müssen. Das gilt von Ballon bis zu hin Milch, Frisur oder Wetter. Fraglich ist, ob wenigstens solche Phänomene wie Wetter dem binären Verhalten des Modells öffnen und schließen unterliegen: doch wenn Wetter nicht hält, wechselt es nicht notwendig in sein Gegen-Extrem. Dieses Bild wäre noch durch vieles zu ergänzen. Zum Beispiel sind Ballons oder Kisten, die platzen, als diese Gegenstände nicht länger existent, doch ein Seil, das reißt, oder eine Rose, die knickt, können auch danach noch meistens als ebenderselbe Gegenstand behandelt werden. Es ist sicherlich nicht nötig, diese Betrachtung hier fortzusetzen. Die Paradigmen zeigen, d^ß (ungeachtet der großen Unterschiede in den Objektmustern) die elementaren Prädikate des Lexikons, mit denen halten an seinen logischen

86

Klaus

Baumgartner

Defizitpositionen in Beziehung steht, beträchtliche Übereinstimmung aufweisen. Obwohl das Verb im Vorbereich nichts weiter als den kategorialen Aspekt 'Zustand' präsupponiert, erfolgt die Besetzung nahezu einheitlich mit denselben Grundlexemen. Das von Objektmustern determinierte Vorwissen scheint sich in erster Linie - wenn man absieht von den wenigen Prädikaten, die für körperliche Ganzheit zur Verfügung stehen - auf die unmittelbaren Klassen der Zustandsverben zu erstrecken. Sekundär eingeschlossen sind dann natürlich, über den grammatischen Zugriff auf Zustandspartzipien, auch kaum noch abschätzbare Bereiche von Vorgangsverben und Handlungsverben, die jedoch im Resultat indirekt auf die primären Zustandsverben zurückführen müssen. Das heißt aber, daß das Vorwissensproblem im Vorbereich des Verbs sich auf prinzipielle Weise potentiell einschränken läßt durch die Klassen der Positions- und Kontaktverben entsprechend den grundlegenden Beschreibungen durch Kaufmann (Kaufmann 1995, 98 ff.). Auf der anderen Seite impliziert das Verb für die Erwartungskomponente ganz allgemein und für die Tatsachenreferenz im Fall der Negation, daß Gegenstände (wie die von (12)) einem Zustands-Wechsel unterworfen sind entweder in das Ende ihrer Existenz oder in einen neuen Zustand (der, wie öfters erwähnt, wohl die nächstbeste natürliche Ruhelage für das betreffende jeweilige Objekt darstellen muß). Wir haben damit an dieser Stelle, wie das vom gewohnten Standard des Aspektsystems auch unterstellt wird, eine Abfolge von 'Vorgang' und 'Zustand' anzusetzen. Die Bedeutung von halten bewirkt hier jedoch in grundsätzlichem Unterschied zu den gewohnten Annahmen und insofern auch anders als beim Modell von öffnen und schließen, daß keine Inkorporation des vom 'Vorgang' herbeigeführten jeweiligen 'Zustands' in diese Aspektstruktur erfolgt. Von Relevanz ist offenbar ganz allein der implizierte 'Vorgang'. Das ist leicht mit den empirischen Fakten zu belegen. Das Verständnis der GegenErwartung und der Negation etwa im Fall (12.a) sowie entsprechende Nachfragen sind ausnahmslos darauf gerichtet, ob der Ballon nun platzt, aufreißt, herabfällt, wegrollt, davonfliegt oder sonstwas, dagegen - z.B. für den Fall, daß er geplatzt ist - nicht im mindesten, ob sein nächstbester natürlicher Zustand zuguterletzt darin besteht, daß er im Baum hängt, an der Mauer klebt oder in der Gosse liegt. Das heißt aber: Der Nachbereich ist systematisch besetzt mit den elementaren Klassen der Vorgangsverben, die dann beliebige nicht-systematische Nachzustände (als die eigentlichen parallelen Gegenstücke zur Zustands-Andauer bei der Affirmation) zur Folge haben. Sofern für die eigentliche Position des Nachzustands eine Zustandsform gebraucht wird, kann höchstens auf das Zustandspartizip des Vorgangsausdrucks zurückgegriffen werden, der den Nachbereich eröffnet, insofern er den Wechsel bezeichnet. Den Schlußzustand dieser Aspektstruktur bildet somit bestenfalls das abgeleitete Vorgangsresultat. Ein Seitenblick auf haltenß zeigt übrigens, daß die aspektuale Variante des Verbs hier im Grundsatz die gleichen Eigenschaften aufweist. Es genügt, sich die Beispiele (2.a) anzusehen und für den Fall Der Wagen hält die übrigen Subjektnomen wie Karawane, Ballon usw. durchzuspielen. Wenn man den Vorbereich betrachtet, ist zunächst nichts präsupponiert als die Aspektkategorie 'Vorgang', die höchstens zu 'Fortbewegung' spezifiziert werden kann. Für diese Variable gilt wiederum, daß sie erst mit den Kriterien der genannten Objektmuster zu substituieren ist, etwa, der Reihe nach, durch fahren, laufen, fliegen. Bei Satznegation, wenn der Aspekt 'ZustandsBeginn' im ganzen suspendiert ist und das Verb nur auf die Andauer des Vorgangs vom

halten und die Einheit

des verbalen

Aspekts

87

Vorbereich referiert, sind diese Elemente sogar der einzige Anhalt für die durchgängige Bedeutung des Verbs. Dieselbe Situation ergibt sich auch wieder im Nachbereich, für den das Verb dann lediglich den rein kategorialen Aspekt 'Zustand' impliziert. Es ist die Frage, ob nicht auch dieser Zustand gleichfalls nach den genannten Objektmustern zu differenzieren ist. Die beiden halten^/ß zeigen insofern, trotz aller Gegensätze, nahezu gleiche Abstraktionen im Aspekt verhalten. Nach allen diesen Klärungen sind nun ansatzweise die Klassen von Prädikaten aufzuführen, die zur Beseitigung der Wissensdefizite im Vorbereich und Nachbereich von halten eine erste Charakteristik liefern. Es ist nicht mißzuverstehen, daß damit natürlich keine effektiven Präsuppositionen des Vorzustands oder Implikationen des Nachzustands angegeben werden (zumal sich nichts daran geändert hat, daß hierbei die bloßen aspektualen Kategorien 'Zustand' oder 'Vorgang' nicht unterschreitbar sind). Die Auflistungen (13) und (14) repräsentieren insofern nichts anderes als (möglicherweise noch unvollständige) Disjunktionen der - (.a) existentiellen oder (.b) lokalen - Zustände und Vorgänge, aus denen sich die beiden Paradigmen des potentiellen Wissens bei halten primär zusammensetzen: (13) a. b.

(14) a. b.

(generell:) ganz sein, heil sein, in Ordnung sein, ... stehen, liegen, sitzen, knien, hocken, kauern, lehnen, stecken, hängen, schweben, haften, kleben, ..., (generell:) s.befinden, ruhen, ... platzen, bersten, springen, explodieren, reißen, brechen, knicken, ..., (generell:) entzwei gehen, kaputt gehen, ... fallen, steigen, stürzen, rollen, rutschen, kippen, wackeln, ...

Die Disjunktionen (13) gegen (14) sind völlig unabhängig voneinander und bilden keinerlei Korrespondenzen jenseits der Zuordnungen, die durch Objektmuster determiniert sind. Ständige Korrespondenzen sind nur im Gedankenexperiment möglich: Man könnte sich z.B. ein Verb * stützen erfinden, das bei Negation exakt den Zustandswechsel von stehen zu geplatzt sein bezeichnet und sonst nichts. Der einzigartige Witz von halten ist jedoch gerade, daß es wie eine Spinne im Netz des Lexikons alle erdenklichen Paare von Verben solcher Art in Aspektstrukturen von (13) zu (14) vereinigt. Für jeden konkreten einzelnen Gebrauch des Verbs läßt sich natürlich, wie anfangs erwähnt, die lexikalische Bedeutung oberflächlich erfassen durch Bedeutungspostulate, die entweder aus der Verknüpfung der Elemente von (13) mit bleiben7 oder durch Negation der Elemente von (14) entstehen: ganz+bleiben, stehen+ bleiben, nichts-platzen, nicht+fallen, usw. Die Ergebnisse, die damit vor uns liegen, können ohne besondere Schwierigkeiten in den formalen Ausdruck (9) übertragen werden, wenn man an diesem Ausdruck, 7

Die Konstruktion Positions/Kontaktverb+Mei&en ist nicht nur Entsprechung zu haltenA, sondern ebenso zu haltens, und zwar durchweg von stehen+bleiben bis kleben+bleiben. Es existieren somit ein bleibenA und ein bleibens mit den Aspekten Zustands-Andauer und Zustands-Beginn in frappierender Parallele zu halten. Worauf diese Ubereinstimmung zurückzuführen ist, läßt sich hier im einzelnen nicht verfolgen, zumal bleibens von allen großen Wörterbüchern (vgl. Fußnote 1) nicht einmal registriert wird, folglich als Alternative weder erkannt noch erklärt ist.

88

Klaus

Baumgärtner

trotz einiger Bedenken zu seiner Struktur, zunächst keine taktischen änderungen vornimmt. Mit dem Ausdruck 'PAR/Vi' wird das Prädikat wiedergegeben, das aus dem Zustandspartizip eines Vorgangsverbs hervorgeht. Danach ergibt sich der modifizierte Ausdruck (15): (15) Gegeben seien eine Zeitlogik mit den Intervallen Äußerungszeit to, Referenzzeit t j und t j (sodaß t» < tj), ein Gegenstand X, eine Menge von Zuständen Z°, eine Menge von Dispositionen D° = {..., D^, ...} (sodaß D° c Z°) und eine Menge von Vorgängen V° = {..., Vi, ...} derart, daß die Satzform halten(X), geäußert zur Zeit to in bezug zur Zeit t j , wahr ist genau dann wenn a.

Dfc(X) zur Zeit tj,

b.

erwartbar ist zur Zeit tj, daß PAR/Vi(X) zur Zeit t j ,

c.

Dfc(X) zur Zeit t j - oder alternativ unter Negation:

c'.

PAR/Vi(X) zur Zeit t j .

Der Ausdruck (15) besagt nunmehr noch differenzierter als Ausdruck (9), daß unter der Haltens-Bedeutung verstanden werden muß: Für den Gegenstand X im prekären Zustand einer Disposition besteht zu der Zeit, in der das Prädikat halten auf ihn bezogen wird, gegen die gegebene Erwartung, daß der Dispositionszustand über einen Vorgang in denjenigen Zustand wechselt, der mit dem Vorgangsresultat bezeichnet wird, tatsächlich der Sachverhalt, daß X sich weiterhin im gegebenen Dispositionszustand befindet. Falls alternativ dazu jedoch nicht+halten auf den Gegenstand X bezogen wird, besteht als tatsächlicher Sachverhalt der Inhalt der Erwartung. Zu dieser Beschreibungsform sind schon zu Ende des Kapitels »Das Erwartungsproblem« und im Verlauf dieses Kapitels zwei Einwände erwogen worden. Der eine betrifft die Frage, ob sich die Erwartung unter (15.b) überhaupt schon auf den spezifischen Vorgang (und damit auf dessen Resultat) beziehen kann, der womöglich erst im Fall eines effektiven Endes der Disposition D^ unter (15.c') entschieden ist. Das ist letztlich die Frage, ob Gegenständen in einer riskanten Disposition im Prinzip immer nur eine einzige Ereignisweise zur Verfügung steht für ihr tendenzielles Herausfallen aus dieser Disposition. Sie wird vermutlich besser verneint. Der andere Einwand betrifft das Verhältnis zwischen riskanter Disposition und Erwartung des Nicht-Haltens überhaupt und läuft auf die Frage hinaus, ob für halten nicht eins von beiden genügt: die Auszeichnung eines Anfangszustandes als nichtnatürliche Disposition oder die Voraussetzung einer Gegen-Erwartung zum Halten. Als strikte Alternative wäre dies wohl nur schwerlich zu rechtfertigen. Es würde ein semantisches Weltkonzept unterstellen, in dem jedes Vorkommen eines irgendwie prekären Sachverhalts ausnahmslos die dazugehörige Erwartung seiner Beendigung determinieren müßte. Schwächere Alternativen in der Beziehung zwischen der Erwartungsstruktur (15.a-b) und der Aktualitätsstruktur (15.a-c) sind dagegen, wie oben schon kurz angedeutet, ohne weiteres vorstellbar. Wir machen hier aber nicht den Versuch, solche möglichen Varianten in genaueren Paraphrasen oder Notationen festzuhalten: Nachdem sich bei der Interpretation von Beispielen wie (11) erwiesen hat, daß der Ausgangszustand (15.a) für die Aktualitätsstruktur (im Gegensatz zu der durchweg

halten und die Einheit des verbalen

Aspekts

89

stabilen Erwartungsstruktur) unter gewissen Umständen völlig entfallen kann, dürfte es am zutreffendsten sein, die beiden Komponenten in der Bedeutung des Verbs möglichst getrennt zu behandeln, wie das in (15) geschehen ist.

4

Konsequenzen

Die abschließende Frage ist, inwieweit das einzigartige Verhalten, das für halten (von nun an wieder insgesamt) analysiert werden kann, mit den Grundzügen und neueren Diskussionen des Aspektsystems in Einklang zu bringen ist. Dabei sind nicht alle dargestellten Erscheinungen vom gleichen Interesse. Von Wichtigkeit sind insbesondere: (i) haltenA ist nicht Zustands-Andauer von halten, (ii) haltenA ist nicht ZustandsAndauer von genau stehen oder liegen oder sitzen usw., (iii) haltenA präsupponiert rein kategorial den Aspekt 'Zustand', (iv) nicht+haltenA impliziert rein kategorial den Aspekt 'Vorgang', (v) haltenß ist nicht Zustands-Beginn nach genau fahren oder laufen oder fliegen usw., (vi) haltenß ist nicht Zustands-Beginn von genau stehen oder sitzen oder liegen usw., (vii) haltenß präsupponiert rein kategorial den Aspekt 'Vorgang', (viii) haltenß impliziert rein kategorial den Aspekt 'Zustand'. Auf das Erwartungsmoment, so wesentlich es auch ist für die Bedeutung des Verbs, wird hier nicht mehr zurückgegriffen. Die letzte Untersuchung des Aspektsystems im kompletten Zusammenhang von Beginn, Fortdauer, Abbruch oder Unterbrechung fortlaufend verbalisierter Ereignisse stammt von ter Meulen (ter Meulen 1995). Die Problemstellung könnte erwarten lassen, daß gewisse Einsichten auf die Erklärung der Strukturen von haltenA/ß zu übertragen sind. Tatsächlich findet sich bei ter Meulen ein ausgefeilter Formalismus, der in der Lage ist, Analysen für ganze Textfolgen in großer Variation zu liefern und die Bedeutungen von Einzelsätzen in die komplexen Aspektstrukturen ganzer Diskurse umzusetzen. Grundaspekte wie Beginn, Fortdauer, Abbruch oder Unterbrechung werden dabei jedoch in genereller Form für die Verhältnisse von Satzbedeutungen vorausgesetzt. Der Apparat (etwa einer kubusförmigen Struktur aller Textaspekte) läßt jedenfalls nicht erkennen, welche zusätzlichen Argumente für die Strukturierung der impliziten Aspektmuster verbaler Prädikate zu gewinnen wären. Dazu gehört, daß Prädikate wie beginnen, fortfahren, beenden usw., die für die Assertion von Aspekten (und daraufhin auch zur externen theoretischen Ordnung von Sätzen) zur Verfügung stehen, ohnehin nicht mit komplexeren Prädikaten zu identizieren sind, deren Bedeutung womöglich äquivalente Binnenstrukturen aufweist; darauf haben insbesondere Kamp und Reyle hingewiesen (Kamp & Reyle 1993, 559). Andere spezielle Arbeiten, die mit dem Verhalten von Zustands-Andauer oder Zustands-Beginn und ähnlichen Aspekten intensiv befaßt sind, scheinen nicht zu existieren. Das Grundkonzept, das Hypothesen oder Ergebnisse zu den Aspekten generell zur Voraussetzung haben, ist nach wie vor die Logik der Verbklassifikation von Dowty und, nach dessen eigenem Verständnis, der ganzen Tradition von Aristoteles bis zu Ryle, Vendler, v.Wright usw. (Dowty 1979). Ein Kernthema von Dowty ist, den Satzoperator BECOME (mit Derivaten wie START, REMAIN usw.) semantisch zu präzisieren.

90

Kl aus

Baumgärtner

BECOME gilt zurecht als Herz des Aspektsystems, ist aber keineswegs unproblematisch als Operator für Zustands-Wechsel aller Art (weshalb auch zunächst vermieden worden ist, BECOME in den Ausdruck (15.c) einzuführen). Dowty behandelt BECOME - von gewissen sonstigen Varianten abgesehen - auf zweierlei logischer Grundlage, innerhalb (i) einer Zeitpunkte-Logik und (ii) einer Zeitintervall-Logik. Wir nehmen nur bezug auf die Explikation (ii), begnügen uns aber auch hier mit einem formalen Minimum8: BECOME (p ist wahr am gegebenen Zeitintervall, sofern am vorangehenden Zeitintervall der Satz -> wahr ist und am nachfolgenden Zeitintervall der Satz wahr ist. BECOME bezieht sich somit primär auf ein Intervall, das unausgedehnt oder verschieden ausgedehnt sein kann, aber auf beiden Seiten an Grenzintervallen überlappt sein muß vom Intervall davor und vom Intervall danach. Für die Anschauung läßt sich hier gut Dowtys eigene Zeitpfeilfigur gebrauchen; wir ergänzen sie nur um das Modell des Verbs öffnen (das von ihm selber an dieser Stelle mehrfach als prominentes Beispiel benützt wird): (16) a. BECOME gegolten hat. Tatsächlich aber dient die Konstruktion dazu, zwei weitere Bereiche abzudecken: zum einen gezielt den Bereich gradueller Prozesse verschiedenster Abstufung (für Verben wie platzen, sinken, welken), wofür BECOME ein Intervall braucht, das ausgedehnt sein kann; zum anderen eher stillschweigend den Bereich von konträren oder kontradiktorischen Folgerungen (wofür die Prototypen öffnen vs. schließen stehen). Daraufhin erst ist BECOME der eigentliche Operator für Zustands-Wechsel von einem ' zu einem " derselben Kategorie. Für weitere primitive Operatoren wird hier keine entsprechende Logik angeboten. Aber natürlich verlangen bereits Negationen wie -> öffnen ein Gegenstück zu BECOME. Die erwähnte Zeitpunkte-Logik, von der Dowty ausgeht, stellt dafür einen Operator REMAIN bereit mit der Definition: -i BECOME -*(>. Formal abgebildet auf die Struktur von (16.a), also weiterhin mit 'geschlossen' als negativer Basis, bekommt REMAIN diese Gegenfigur: 8

Dowtys formale Definition für den Zustands-Wechsel umfaßt keine Bedingungen über unsere sprachliche Version hinaus, wird aber durch Auflagen zusätzlich abgesichert gegen etwaige laufende Iteration der beschriebenen Intervallfolge in sich selber (Dowty 1979, 139-145). Die modifizierte Definition für einen sukzessiven Zustands-Wechsel vom Typ von A nach B laufen bleibt hier beiseite.

halten und die Einheit des verbalen

(16)

91

Aspekts

b. REMAIN -.0 ist wahr

-i ist wahr

geschlossen

öffnen

geschlossen

Es versteht sich, daß für REMAIN kein mittleres Intervall entsprechend BECOME gegeben sein kann; daher die Striche in der Figur (16.b). Das wirft jedoch die Frage auf, ob REMAIN als kaum strukturierte Funktion der allerprimitivsten Fortdauer des Gegebenen überhaupt aus der Negation von BECOME abgeleitet werden soll. Wie leicht zu sehen ist, taugt Schema (16.b) nicht für haltenA und das Gegenbild (16.a) ebensowenig für haltenß• Gesetzt, nur an Intervall 3 wahr sein, aber nicht, wie erforderlich, schon an Intervall 1. Im anderen Fall, wenn man BECOME etwa auf (2.a) Der Wagen hält anwendet, ist das Ergebnis kaum anders. Gleichgültig, wie hier die Negation aussieht: da Nicht-Halten jedenfalls nicht dazu geeignet ist, einen Gegenvorgang zu definieren, sind Der Wagen hält nicht oder Der Wagen hat nicht gehalten lediglich gleichermaßen kurios an Intervall 1. Das heißt aber: Eine zutreffende Beschreibung von haltenA/b ist nicht zu bewerkstelligen ohne Ersetzung der einfachen Satzbewertung durch eine Bewertung, die selber schon für Aspekte vorgenommen wird. An Intervall 1 muß es möglich sein, die Kategorie Zustand zu identifizieren für haltenA, die Kategorie Vorgang für haltenß. Die Rahmentheorie von Dowty und allen anderen, die ähnlich verfahren, ist so offenbar nicht zu halten oder nur dann, wenn sie durchgängig auf problematische Sachverhaltskategorien umgestellt wird. Damit kommen die Resultate im Resümee vom Anfang unmittelbar ins Spiel. Außerdem ergeben sich sogleich zwei weiter reichende Probleme: zum einen die Begründung einer einheitlichen kategorialen Struktur für das Aspektsystem und zuvor schon ein prinzipielles theoretisches Dilemma. Die Kategorienfrage wird zum Schluß noch kurz gestreift. Das Dilemma entsteht bei Durchsetzung unserer Intuitionen über Zustände und Andauer von Zuständen. Betrachten wir REMAIN , Folgen von Intervallen 1+3 (wie in (16.b)) und das Verb haltenA- Andauer und Einheit von Zuständen wären problemlos gesichert, wenn es überhaupt sinnvoll wäre, ein und denselben Satz mit haltenA erst aktual an Intervall 3, dann präsupponiert an Intervall 1 zu bewerten (was aber unmöglich ist). Und umgekehrt: Es wäre völlig problemlos, den Satz 3 mit haltenA und einen dazugehörigen Satz 1 etwa mit liegen als ein und denselben Zustand zu deuten, wenn man dabei nicht zweierlei Zustände bekäme (die eigentlich nichts anderes sind als Strecken ein und desselben ununterbrochenen Zustands) 9 . Die eingehende Untersuchung, die Bartsch der Identifikation 9

Entsprechend leicht stößt man auf Formulierungen wie die, daß Zustands-Andauer gegeben sein soll bei Verben in Sätzen, die eine Folge von gleichen Zuständen bezeichnen (Schumacher 1986, 16), in denen das Dilemma wohl nicht einmal reflektiert ist.

92

Klaus

Baumgärtner

von Situationen im Aspektsystem widmet, bringt eine ganze Taxonomie der Kriterien zutage, unter denen Sachverhalte von Prädikaten (über Individuen, Konzepten oder Rollen durch Zeiten, mögliche Welten oder Perspektiven) identifizierbar sein müssen (Bartsch 1995). Das Problem der Zustands-Andauer fällt unter solche Identifikation (sogar als Grenzfall immer derselben Kriterien), solange sich nichts daran ändert, daß ein Zustand, der kontinuierlich wahrnehmbar und z.B. mit halteriA beschreibbar ist, in lauter logische Zustände zerfällt, sooft die Haltens-Aussage getan wird. Das Dilemma tritt nicht bloß bei halten A auf oder bei nicht+halt eng, sondern natürlich auch schon bei nicht+öffnen und damit in allen Sätzen, in denen die Negation des Wechsels einer Aspektkategorie vorkommt. Skizzieren wir damit die Ergebnisse in einer extensionalen Form, die nichts weiter voraussetzt als Klassen von Verben (wie oben bei (13) zu (14)). Die Analyse von haltenA wird allein dann korrekt erfasst, wenn für den Vorbereich an Intervall 1 nun eine Menge $ von Zustandssätzen vorgegeben ist. REMAIN ist wahr genau dann, wenn ein Zustandssatz wahr genau dann, wenn ein Vorgangssatz tph € ® wahr ist (jedoch, wie vorgeführt) nicht an Intervall 3, sondern nur am Intervall von BECOME selber: im eigentlich leeren Nachbereich steht höchstens das Zustandspartizip PAR/Vi des Vorgangsverbs, das zum Gegenstand X paßt. Die Figuren (17) und (18) verweisen auf solche Elemente des Ausdrucks (15). (17) a. REMAIN tpg G $ ist wahr

vsem_J arbeiten(e.y) : = l e , ,akt(hom),MTV

arbeiten(e) agent i C t

c o n a

(e) = x

arbeiten is a function from pairs of referents into labelled DRSs of t y p e verb semantics, where the one argument is the event variable and the other the agent of the event which, by the compositional semantics, will be identified to the distinguished referent

German French Machine

Translation

115

of the subject. We note that there is some Aktionsart typing of the event variable (which is the distinguished referent of the verb representation) and of the structure as such (seen as an event predicate), and in connection with this of the thematic role, where hom means homogeneous and cons means constant. We come back to this in the section after next. Note also that the event introduced is bound to some time t, which we call the focus time and which will be determined by the context as an available, presently focussed reference time of the context. The decoration of the distinguished event referent, besides the Aktionsart information, will contain tense information also, by MTV. The evaluations of the flat noun- and verb-representations are so called basic representations (i.e. conventional, but decorated DRSs). jung(npsem_l) =^npsem_l

jung(x, Ay P(y))

jung(L) := l l x : CXpresp

(Pj

jung is a function from labelled structures of type NP semantics into structures of the same type. It percolates the distinguished referent of the argument by determining it as its own distinguished referent. The evaluation is structurally different from the evaluations of the noun- and verb-representations. It puts its (NPsem-type) argument L at a distinguished position and stipulates this argument to be dominated by the adjective contribution proper in the UDRT-sense. We call such structures pDRS structures. They are typical for modifiers like adjectives and adverbs. In disambiguations of FUDRSs the chosen argument of the modifier will be unified to this argument-L (and with this the argument referent to the percolated referent). In case, the argument is detpsem_l q d e r ( L ) := lie,!,!:

der \ L2e

{13 < 12} L3

det def,ag (npsem_l;qpsem_l) =>-detpsem_l detdey(Sp( P

— L x ) :— lle,x,x*

detrfe/^j( qpsem —'L

1 a;ia;) :— lie,!,! X,X •

{12 < 11} L2e "icf (*.••»(!•))

{12 < 11} L2£

der is a function from NPsem labelled structures and QPsem labelled structures into

German French Machine Translation

117

DetPsem labelled structures, where QPsem labelled structures are representations of (quantized) phrases like viele Männer which allow for a (further) determiner and where DetPsem labelled structures are determiner phrases which don't accept a further determiner or quantifier. The result is a pDRS-representation whose argument is a VP-representation and whose label is annotated by three distinguished referents, where the first is the percolated (or modified) event variable, where the second is the referential argument of the D P as accessible from outside the D P representation (in case of summation, as with viele, it is the sum referent), and where the third is the referent which is accessible from the argument VP-position (the referent which, by the syntax-semantics interface, is identified as (one of) the complement argument(s) of the VP). The evaluations of der render the referential determiner use and the attributive use as in derjenige welcher/the one who. The sentence representation itself, (11 pudrs) 5 structurally is of a third (and last) type of representation: it is a so called functor set representation, consisting of a set of functors - the structures labelled by 11, 12 and 13, of a set of ordering conditions (which constrain the order of application of the functors to the argument in a disambiguation of the representation), and, finally, of the argument, that is, the minimum of the partial order described. For more details about this represesentation formalism and a discussion about the similarities and differences between FUDRSs and UDRSs refer to Eberle 1997b, also refer to this study for the disambiguation algorithm that turns such structures into classical DRSs and for the FUDR-model theory. Adopting the referential reading of der (the attributive reading is not supported by the context), and accommodating the D R F of the corresponding description at the main DRS-universe, (11 fudrs) can be disambiguated to the following two DRSs: x t now mann(x) jung(x, Ay. mensch(y)) t -< now arbeiten(e) agent(e)=x stundenlang(e) e Ct

x t now mann(x) jung(x, Ay. mensch(y)) t -< now stundenlang(t) e arbeiten(e) agent (e)=x e Ct

Now, what is the relation between the dependence analyses, as depicted by the dependence trees, and FUDRSs? The leaves of the dependence structures are easily identifable as lexical FUDRs with evaluations of the basic type and of the pDRS-type respectively. With regard to (11/?), we identify as follows (remember that the nodes of the dependence structures carry names (position numbers), which are not rendered in the visualization, but which, nevertheless, we can use as labels of the corresponding FUDRSs):

118

Kurt

Eberle

1 - s(stundenlang,423619) adv(_,adv)

lie :stundenlang(Ll')

2 - s ( a r b e i t , 4 1 9 9 9 ) mtv(ind:dcl:nwh,tf(past,0,_),a)

I

3 4 5 6

axbeiten(e,x) 13£ '.y.^dertLS') 14 2 :jung(L4') 15-r:mann(x) 16,:nicht(L6'i

-

19 ~ ^,e,mtv(ind:dcl:nwh,tf(past10,_),a)

s(der,d) det(nMgdFSgP,def) s(jung,367349) adj(nMnaFNSnaP,_) s(mann,1062123) n o u n ( c n , [ n o m , e g , m , k d a | d e t ] , n w h ) s(nicht,neg) adv(_,nwh)

•

It should be clear, that, up to renaming of variables, and more generally, up to isomorphism, of course, for each leaf node of the dependence structure there is exactly one FUDRS describing the node. This is so because lexical ambiguity is passed to the evaluations of the flat representations. For further illustration: the article der not only is ambiguous between the referential and the attributive meaning (as shown by the representation above), it also accepts plural referents (case genitive plural) and its set of evaluations will therefore also contain the corresponding collective and distributive reading. The further constraint of case nominative singular, which is provided by the sentential agreement information, effects the respective constraining of the evaluations of der (to the treated det^e/.sg-cases - by requiring the identity and atomicity of the referential indices). However, there is no parallel decrease of the set of interpreting FUDRSs. The different meanings are summarized by the one initial 'der'-representation der. The difference completely resides in additional constraints (which, in the case at hand, are about the decoration of der) and the corresponding impact on the admissible kinds of evaluation. We already have mentioned that slot grammar analyses don't really make a difference between verbs and their projections (and nouns and their projections etc.). However, FUDRSs do. The verb is represented as a FUDRS of the basic type, whereas the VP is represented as a FUDRS of type functor set, consisting of a basic FUDRS, for the verb, at its bottom, of a set of functor representations and of a set of specifications about the order of applying the functors to the argument. We solve this problem by considering the dependence structure as a set of nodes, a set of two place relations (which are the slot relations), and inferrable from these, a set of (sub)trees: Definition (preliminary): D e p e n d e n c e structure A dependence structure DS is a structure:

where Edges are pairs of nodes, Subj-Edges etc. are these subsets of Edges which are subj-, obj- relations etc., and where the trees are either the singleton sets over the set of nodes or these sets of nodes which are dominated by exactly one node, w.r.t. Edges, and which are maximal in this respect. 1 {TFUDRS(T)m

[ | ), . . . , T F C / D f l s ( D N n ^

) }

According to this, for each DS, there is a one-to-one-mapping M TOax which, to a node N, assigns the maximal tree which is dominated by N, and there is a second one-to-onemapping M m i n which assigns the set consisting of N to N. We can use these mappings in order to get our projection problem right: we do not map the nodes of a DS into FUDRSs, but the corresponding trees: The singleton set trees (meaning nodes as such)

German French Machine

119

Translation

of a DS are interpreted by basic FUDRSs or pDRS-FUDRSs as illustrated above. In case a node N of a DS is modified, M m a a ; (N) is different from {N} ( = M m i n ( N ) ) . We map M m Q I (N) onto a functor set FUDRS which uses the FUDRS of {N} as its bottom element, and the FUDRS-interpretations of the maximal subtrees of M m a a ; (N) as the functors. Calling this mapping T F U D R S (translation into FUDRSs), the relevant recursion step easily is rendered as follows: T f ( / C f l s ( M N [DN1 ^ , . . . , DNn ^ ] ) := IMJV:

{TFc/Dfls(DNl [ j ) , . . . , T FC/DRS (DNn j j ) } Tfi/D«S({MN})

where MN [DN1,..., DNn] stands for the tree dominated by MN with maximal subtrees dominated by DN1, . . . , DNn, and Tf(/£>fls(DNx j j ) for the translation of the tree dominated by DNx. Considering this recursive translation of dependence structures into FUDRSs, the specific edge-information is often redundant. For instance, vadv says that the daughter applies adverbial modification to the (syntactic) verbal mother. This, however, is clear already, knowing that the daughter is an adv(_,_) and the mother a verb (mtv(_,_,_)). The edge-information is relevant to distinguish adjunct modification from subcategorized roles however, and, with regard to the latter, the naming of the edges represents the linking information of the syntax semantics interface. (Since the case-roles of the dependence structures are deep cases in the Chomskian sense, there is no risk, however, that the syntax style naming of the roles conflicts with the assumption that this naming represents an underspecified thematic classification of the roles of the considered verb). With subcategorized roles, therefore, the naming of the edges causes the distinguished referent of the role to be correctly identifed to the corresponding verb argument. We skip formally working out the translation function TFUDRS in this respect. We also skip extending T F U D R S to all types of dependence descriptions as can be generated from the German slot grammar (or the grammar of some other language). It should be clear how this working out proceeds with regard of other parts of speech (conjunctions etc.) and corresponding syntactic structures. What is more relevant to our subject here is to have shown how lexical Aktionsartinformation and tense information is available as decorations of the FUDRSs, where the percolation of the information from decorations follows the (HPSG-like) principles formulated in Eberle 1997b. As it stands, the dependence structures are identifiable as FUDRSs, as shown. However, they always represent maximally weak information with respect to the partial ordering statements that can come with the FUDRSs. The only ordering constraints stem from the dependence hierarchy as such and from type information: The functors of a relative clause, for example, are bound within the local domain of the modified NP (note that FUDRSs treat wide scope readings of embedded definites as a matter of presuppositional projection and not as a matter of quantifier scope, for this, again, see Eberle 1997b). (11) presents an example of how type information can constrain the partial order of the functor representations, der and jung are modifiers of mann. Since both modify NPs and since only jung outputs structures of the same type, in each disambiguation, jung must be applied before der. Therefore, it holds:

Kurt Eberle

120

{13e ,j/,y:der,14x :jung(L4')} {14 verb [obj(n | fin | ob),iobj(n >rn: ecrire

\ p( [anfycc] ))]

122

Kurt

Eberle

(12 d ) Dependence t r e e . > — subj(n) s(er,206261) noun(pron(pers3),[nom,sg,m,kda|X3],nwh) o top s(schreib,615154) mtv(ind: d e l : n w h , t f ( p r è s , 0 , 0 ) , a ) ! ' - ndet s(dem.d) det(dMNS,def) ' iobj(n) s(mann,1062123) noun(cn,[dat,sg,m,kda|det],nwh)

( 1 2 TD) Target dependence t r e e . o— '— '

top subj(n) iobj(p(['a|dat])) ndet

e'erire mtv(ind:del:nwh,ti(près,0,0),a) il noun(pron(pers3),nom,pers3-sg-m,X2) homme noun(cn,dat,pers3-sg-m,X3) d det(dat,pers3-sg-m,X3)

Since the lexical entry of schreiben does not provide particular constraints for the transfer, the structure is preserved and the naming of the nodes and edges follows t „ and r s . Note that each structure that is computed at intermediate levels of the recursive process is a well defined target dependency structure (in the sense of the target grammar) that renders a substructure of the final complete target dependency structure. Next to formal exactness, this means that at each level of processing, inference mechanisms which are defined for the expressions generated by source and target grammar respectively are applicable to the actual input and output structures. Translation cannot always output target structures which are homomorphic pictures of the source. With regard to the mother node and its daughters, renaming slots, exchanging slots for each other, deleting and adding slots (and fillers), and more generally, renaming, deleting and adding paths (and path values), in this order, increases the structural difference between source and target. Dorr 1992, Kuhn and Heid 1994 present examples and classifications of (such structural) translation problems. Similar to suggestions like Kaplan et al. 1989 (for the LFG framework), we assume that the lexical item can stipulate specific transfer correspondences about the arguments which will be assigned to it in the sentence, or, more generally, about the substructure which it is the head of in the sentence. These particular transfer statements override the general translation routine with respect to the node in question. Consider the following example and its French translation: (13) Er gedenkt des Mannes. Il évoque le souvenir de l'homme. Here, the German genitive object must be deleted and a new obj-role must be introduced that incorporates the source genitive at an embedded level. We provide the following lexicon entry that handles this case:

German French Machine

Translation

123

(13 L) gedenk >verb [gobjl ] > r '.évoquer [item(obj(n), [souvenir|m] .det(def,sg), [comp(p(de)):r(d(gobj))] ),e]

The item introduces an instruction for the recursive transfer which says the following: The translation of the slots which are less oblique than the genitive slot follow the general recursive translation routine using r s (this is the information of not mentioning them, as in (12/,) -here, this only relates to subj). The target does not know a genitive slot, however, (the e refers to gobj and means empty, the '1' of gobj requires the slot to be obligatory). In addition, there is a new slot that has no counterpart in the source (designated by (new) item) and which is an obj(n)-slot with head word souvenir realized in the definite singular variant. In addition, this obj-role comes with a complement which is a rfe-PP and whose content is the translation of the source genitive object. Note that further embeddings would be allowed by assigning itemliterals as values of the slot-list of target-items, if necessary. With this, we obtain (13TD) as r-correspondence of the dependence structure (13D) of the sentence (13): (13D) , o ! '

(13

subj(n) top ndet gobj (n)

s(er,l) noun(pron(pers3),[nom,sg,m,kda|Xl],nwh) s(gedenk,l) mtv(ind:dcl:nuh,tf(près,0,0),a) s(des,d) det(gMNS,def) s(mann,l) noun(cn,[gen,sg,m,w|det],nwh)

TD)

0 > ! , >-+ ! ,'

top subj(n) ndet obj(n) ndet comp(p(de))

e'voquer il d souvenir d homme

mtv(ind:dcl:nwh,tf(près,0,0),a) noun(pron(pers3),nom,pers3-sg-m,X2) det(acc,pers3-sg-Xl,X2) noun(cn,acc,pers3-sg-Xl,X2) det(dat,pers3-sg-m,X3) noun(cn,dat,pers3-sg-m,X3)

We will not go into the details of the lexicon formalism here. The example should make clear however, that the so-called r-s/ois-statements which, as illustrated by (13¿), describe path equivalences of the source- and target-structures which are dominated by the considered item and its translation respectively, certainly are expressive enough to solve the transfer problems already mentioned. Also, it should be clear how such statements fit with the recursive translation strategy and how they direct the general transfer algorithm by presenting specific translations for particular nodes (or, more precisely, for the trees dominated by such nodes). However, there are structurally more complex translation mismatches, which cannot be dealt with by the transfer algorithm as it is defined at present. These problems are posed by the so called head, switching phenomena, where functor and argument exchange their positions for each other. The following (14) presents an example of this type of structural change:

124

Kurt Eberle

(14) Er schwimmt gerne. II aime nager. H E LIKES T O S W I M .

In the lexicon, we treat this by a (type changing) r-slots statement within the entry of gerne: (14 l)

gerne >adv [] >T:

(u-cat(verb) —• aimer [item(subj(X'),r(u-d(subj(X)))),item(obj(binf),r(u2^ ) ) ) ]

Here, u and d stand for up and down respectively, where the corresponding path statements relate to the edge structure of the dependence tree. v£ld is like u , except that besides id (that is, the lexical structure itself) it does not contain the modifiers which are assumed or known to be outside the scope of id. (14^) states that provided gerne modifies a verb, it is translated into aimer, where the subject of aimer will be the translation of the subject of the modified verb (where the subclassification of the target subject, X ' , is identical to the subclassification of the source subject, X modulo the TS default correspondences, that is, r s ( s u b j ( X ) ) = s u b j ( X ' ) ) , and where the object of aimer will be the bare infinitive variant of what r ( u 2 t d ) designates. Here, r ( u 2 i d ) designates the translation of the verb together with the (subcategorized and free) modifiers of the verbal complex that are analysed as being in the scope of gerne. For (14) we obtain the following: (14 D)

subj(n) s ( e r , l ) o- top s(schreib,l) ' - vadv s(gerne,l)

( U

T D

noun(pron(pers3),[nom,sg,m,kda|X5],nwh) mtv(ind:dcl:nwh,tf(pres,0,0),a) adv(X6,X7)

)

o— ' — >—

top subj(n) obj(binf) subj(n)

aimer mtv(ind: d e l : n w h , t f ( p r e s , 0 , 0 ) , a ) il noun(pron(pers3),nom,pers3-sg-m,X2) e'erire mtv(inf,tf(pres,0,0),a) empty coref(l)

Note that, here, in order to avoid cyclic structures, we exploit the fact that our source structures are (unresolved) semantic structures and, therefore, allow for scopal ordering of the functors of an argument such that, in the transfer definition of gerne, the structure u can be restricted to which otherwise would contain id itself such that the recursive formulation of the translation would trigger an infinite regression. It remains to reformulate the transfer routine such that it accounts for the order of the transfer steps as prescribed by the scopal interpretation. As a prerequisite, we

German French Machine

Translation

125

assume the input to be rewritten as a sequence of application steps in accordance with the condition set OC. With this, we obtain the following basic routine: modi: Modi, mod n : Modn

k OC) = r(modi(Modi,... mod„(Mod n ,Mother)...) )

:= r s (modi)(r n (Modi),...r g (mod n )(r n (Mod n ),r n (Mother))...) & OC As before, specific r ' s from lexicl entries may stipulate specific conditions about the (syntactically or semantically) subcategorized complex. Often it is argued that head switching is a syntactic phenomenon that disappears if transfer is formulated for the semantic level. There, gerne, for instance, takes the VP-semantics as argument and applies some modal relativization to it just like aimer faire qc or to like to do s.th. do. Most elegant in this respect is the interlingua approach to translation, which always assumes identical representations of source and target sentence at the level of universal semantic description (cf. Kuhn and Heid 1994 for a modern HPSG-based example of this approach). Here, translation is just a matter of semantic construction and generation from semantic representations. Until now, nobody has discovered the broad coverage interlingua however. It is not so clear, whether the more modest approach to translation, of which our suggestion is a representative and also others like Dorna et al. 1998 and which assumes semantic representations that may be language dependent in certain respects, can truly assume that there is no head switching on the semantic level. We will not go into detail with this question here, we just want to emphasize that our approach can deal with head switching phenomena, independently on whether they are purely syntactic or semantic in nature. (For a discussion of this question and details of our lexicon formalism compare Eberle 2000a). It can do this, we repeat it, because the representations to be processed are structured into argument and functor semantics related by the scopal constraints of the dependence structure. It depends on the nature of the modification whether the argument is characterized relationally -adjuncts of temporal location do this for instance- or whether the argument falls in the scope of an embedding operator - a s is the case with gerne. In the presence of the latter true scope bearing modifiers, meaning preserving translation necessitates a procedure which decides the scopal relations in these cases which axe not predicted by the syntactic constraints of the source sentence and which, nevertheless, are relevant to get the target dependence structure (and word order) right. Still adhering to the underspecification philosophy, the computation of this information is postponed to the time, however, where, within transfer or generation, this information is needed. (Head switching slots or adjuncts will trigger scoping such that the transfer can correctly evaluate the of the r ( u 2 ld )-statements as referring to appropriate structures consisting of the representation of the syntactic head together with a number of modifiers). This also follows the lazy expansion strategy, as stipulated in Kay et al. 1994 and elsewhere. The "-»"-prefixes in the lexical transfer statements, like "u - cat(verb) - V of the T-value of (14¿), are lexical examples of contextual constraints (in the sense of Eberle et al. 1992). They have to be evaluated to get the target dependence structure right.

126

Kurt

Eberle

Others will be formulated at later stages of processing and/or will be evaluated later like those that are relevant for target word order and those that are relevant for the task of determining the target surface tense, to which we will turn now.

5

Evaluation

Besides the use of the contextual constraints that the lexical entries stipulate for particular disambiguations and/or translations respectively, which we have mentioned in the last section and which are mainly syntactic and semantic type tests applied to the slots of the considered item or to nodes that are reachable via some particular path, we make use of another type of contextual filtering: the antecedent information about pronoun anaphora. The corresponding module, the nominal resolver (that is taken over from the LMT-implementation), makes weighted suggestions about the identity of antecedents, mainly by exploiting information structural knowledge (parallelism etc.) and type information. It does this on the basis of accessibility constraints as provided by the syntactic structure, see Lappin and McCord 1990, Leass and Schwall 1991. This device is optional. Depending on the translation mode, it can be switched on or off. Currently, the algorithm is extended in order to make better use of semantic constraints - like the semantic accessibility relation as provided by the semantic representations of the sentence and the preceding text. Note that semantic accessibility and scope relational setting are interdependent (pronouns must be in the scope of their antecedents). Since the determination of the target tense can depend on the Aktionsart and the temporal localization of the considered event, which, as we tried to motivate in the introduction, in turn also depend on the scope ordering, the different types of resolutions should be interleaved. This is work in progress. At present, when it is switched on, nominal resolution in fact precedes tense and aspect evaluation, which is sensitive to the scopal consequences of the antecedence decisions in this case. Consider the following example: (15) Seinen Lehrer lernt jedes Mädchen bei dessen Amtseinführung kennen. E V E R Y G I R L B E C O M E S ( / W I L L B E C O M E ) A C Q U A I N T E D W I T H H E R ( / H I S ) T E A C H E R AT HIS I N I T I A T I O N .

In case the possessive pronoun seinen is resolved to the different girls (solution with her) the representation of the accusative NP must be in the scope of the representation of the subject NP, otherwise, because of the necessary distributive reading of the quantifier jed, there could not be the corresponding anaphoric link. Now, in this case, the PP-representation must be in the scope of the subject NP also, provided dessen is correctly related to the teacher. From this, we obtain that the temporal modification bei dessen Amtseinführung relates to the single situations of becoming acquainted. In contrast, resolving seinen to some other (wide scope) accessible DRF of the preceding text the wide scope accommodation of the initiation event becomes available and is preferred. In accordance to what has been said in the introduction (when considering example (5)), this wide scope modification would trigger the fu-

German French Machine

Translation

127

turate translation of the German Präsens, whereas, in the first case of narrow scope modification, the quantificational statement as such is not bound to some specific time from an adverbial and allows for the present translation also. Whether this translation is preferred to the futurate translation will depend on whether the sentence is interpreted as describing a (homogeneous) situation rather than a (heterogeneous) event (also in accordance with the motivation of the introduction). This decision, in turn, heavily depends on whether the recipient assumes the quantifying jedes Mädchen to be bound to some implicitly given contextual reference set or not. In the latter case, the sentence will obtain a reading of generic quantification and, with this, it will be attributed homogeneous Aktionsart. In the first case, it will obtain the existential interpretation, and will be attributed heterogeneous Aktionsart. Note that this decision is also a matter of exploiting the results of the nominal resolution (of nominal descriptions in this case). The results of the nominal resolution are maintained in the dependence structures of the text via coreference statements similar to those of the representation of (10) in section 2 for syntactically empty reentrancies. Depending on the different results of the resolution, we obtain different dependence structures from which the tense/aspect heuristics computes the different target tense suggestions, as used in the corresponding translations: (15di) Dependence tree.

— 0 —— 1 — 1
verb [obj(n \ fin)] >T:

croire

verb

>T:

[obj(n)grad]

manger

• offn [obj(n)char]

>verb

>T:

ouvrir

• seh >verb

[obj(n)Const]

>T: voir Of course, we must also classify the more oblique slots, if there are any. Note however, that most of these additional slots in slot grammar analyses are (optional) PP-complements, which in other grammar theories would be treated as adjuncts, mainly certain (collocational) temporal and spatial localizations. In such cases the relevant slot-type can be inferred from the type of the preposition and of the argument (such t h a t the lexical entry is discharged and can do without corresponding explicit statements). Similarly, as sketched further above in section 3, the free modifiers

132

Kurt Eberle

themselves carry the information about their potential for changing the Aktionsart. In Eberle 1995a, we have spelled out an Aktionsart-calculus which, on the basis of this, can compute the Aktionsart for the classical case of disambiguated (DRT-) sentence readings (DRSs). The advantage of the DRS case is that the scope relations are known and the reading of the different roles and modifiers. How about the underspecification case with which we are faced in our translation scenario? First, plural roles can be read collectively or distributively. We represent distributive readings via DRS-duplex conditions. Duplex conditions from quantified verb roles change event descriptions into more complex descriptions (into descriptions of the event sums that are abstractable from the duplex condition). Such changes can be accompanied by Aktionsart changes. Typically, for instance, bare plurals bring about homogeneous event types, as in (17.a). However, they don't always do this, compare (17.b). (17) a.

Rocky öffnete die Tür. Frauen gingen vorüber, jede mit einem Kind. ROCKY

OPENED THE DOOR.

WOMEN

WENT

BY,

EACH ACCOMPANIED

BY A

CHILD.

b.

Rocky öffnete die Tür. Reporter schössen auf ihn zu, jeder mit einem Micro. R O C K Y O P E N E D THE D O O R .

R E P O R T E R S RUSHED OVER T O HIM, EACH WITH A

MIKE.

The accompanying floated quantifier guarantees the distributive reading of (17.a) and of (17.b). However, whereas, in (17.a), the event sum gets the (more or less default) reading, where the atoms of the sum axe ordered successively (according to what a control modifier like nacheinander/successively makes explicit), in (17.b), background knowledge (about the considered event type and the related scenario) says, that the reading with simultaneous subevents should be preferred. We obtain: u v W el t l t2 rocky(u) tiir(v) öffnen (el) agent(el) = u object(el) = v el C t l ti X n e2 c vorübergehen(e2)

(17.a rep )

mit(w,c) e2 C t2 successive(E2) t2 X n el C E2 E2 C t2

133

German French Machine Translation u v R el tl t2 rocky(u) tiir(v) öfFnen(el) agent(el) = u object(el) = v el Ç tl tl ^ n

e2 m zuschiessen(e2) agent(e2) = r micro(m) mit(r,m) spat_goal(e2) = u e2 Ç t2

(17.b r e p ) E2 R ::

simultaneo us(E2) t2 -< n el t2 E2 Ç t2 (17.a) and (17.b) should exemplify what, here, we can only informally motivate: Simultaneous quantification percolates the Aktionsart of the argument event type to the result (het -thet, in (17.b)). This corresponds to the impact of the collective reading of a plural role on the Aktionsart. In the case of temporally distributed quantification, bare plural roles will generate homogeneous event types (—>hom, in (17.a)), 'quantized' roles (definite descriptions and most quantifier expressions) will generate heterogeneous event types however (exchange die Frauen for Frauen in (17.a)). We account for these data as follows: The quantifiers are classified into: • always distributive quantifiers, DISTR_QU,

• quantifiers which prefer the distributive reading, D _ D I S T R _ Q U (distributive per default), • quantifiers which prefer the collective reading, D _ C O L L _ Q U (collective per default), • always collective quantifiers, COLL_QU

where: •

D I S T R _ Q U encompasses the singular universal quantifier (jeder/every) variants (fast jeder/ almost every etc.): UNIV_QU & SG_QU

•


' s described in the very same way. The structure of enumeration that is recognized by the interpreter when he comes to the sentence Alors venait le tour des allumeurs de réverbères de Russie et des Indes, is thus the following. 8

Its acceptability does, however, not alter the argument we are going to make.

Enumerations

and the Semantics

r ¡„venir. e

partiri t

I L z®A ' zeA

r venir. e

partir-i -i e

J' l c®s > ces

This structure then gets extended to r ¡„venir. e

179

of puis and alors

partir-i. r venir.

J J

partir-i, r venir] 1

L l Z@A ' CZ®A J> leC®S > e C®S J' L H®/ J J which demonstrates a lack of parallelism and yields the marginal interpretation. To extend the parallelism the interpreter has recognized up to now he will easlily accommodate the leaving event, e i j^ t j i r , for the groups of Russia and India. I.e., he will resolve the elliptical sentences on the basis of f ¡„venir. K

partir 1. r venir.

partir i r venir.

partir-i l

L L Z®A > eZ®A J' l e C®5 ' e C®S J' lcR®I > e fl®/ J J which uniquely determines the correct interpretation. Note that the last three occurrences of puis completely fit this picture. They mark the enumeration of pairs of coming and going events. But what about the occurrences of alors? They occupy the very same slot. But this is not a contradiction to what we said above. Any use of alors locates the event described in the sentence it occurs in within the state that is determined by the preceding sentence. This state is characterized as the result of the night watchers having left and thus the scene being free for the turn of the new ones. Alors being a binary relation doesn't care about the structure of enumerations. It thus may occur where puis can occur or should occur from the structural point of view. But if it does, then no contradiction to our proposal arises. And that it actually does occur in (15) is thus indeed a matter of stylistics. 9

6

Conclusion

The possibilities to use puis and alors overlap, but aren't identical. Alors is compatible with contingency relations like enablement and result, which puis is not. Alors has a consistency presupposition with respect to the state introduced by the previous sentence, which puis has not. Both, puis and alors may occur in contexts that trigger temporal succession and in contexts that don't. But alors has a temporal meaning 9

It is instructive t o compare t h e G e r m a n translation, in which t h e occurrences of dann completely coincide with t h e introduction of a new item in t h e complex enumeration. For t h e ease of reference I insert (puis) and (alors) at t h e places where t h e y occur in t h e original. Von einiger Entfernung aus gesehen, wirkte das prächtig. Die Bewegungen dieser Armee waren gedrillt wie die eines Opernballetts. (D'abord) Den Reigen begannen die Anzünder der neuseeländischen und australischen Laternen. (Puis) Hatten sie ihre Lampen angezündet, gingen sie schlafen. (Alors) Dann traten die Anzünder von China und Sibirien zum Tanze an. (Puis) Auch sie verschwanden hinter den Kulissen. (Alors) Dann kamen die russischen und indischen Anzünder an die Reihe. (Puis) Dann die von Afrika und Europa. (Puis) Dann die von Südamerika. (Puis) Dann die von Nordamerika. Und niemals irrten sie sich in der Reihenfolge ihres Auftrittes. Es war grossartig.

Uwe Reyle

180

of its own, and puis does not. It may inherit a temporal interpretation in narrative contexts. Furthermore puis cannot be used to combine the goal of a plan to the sequence of sentences describing how the agent intends to achieve this goal. We explained these facts on the basic assumptions that puis may only occur within enumerations, and that alors temporally locates its eventuality within the state determined by the preceding text.

7

References

Adam, Jean-Michel and Françoise Revaz. 1994. Aspects de la structuration du texte descriptif: Les marqueurs d''¿numeration et de reformulation. Langue Franç aise 81. Asher, Nicholas. 1993. Reference to Abstract Objects in Discourse. Dortrecht: Kluwer Academic Press, Caenepeel, Mimo. 1989. Aspect, Temporal Ordering and Perspective in Narrative Fiction. PhD thesis, University of Edinburgh. Grevisse, Maurice. 1986. Le Bon Usage. Duculot, Paris-Gembloux. Kamp, Hans and Christian Rohrer. 1983. Temporal Reference in French, ms. Stuttgart. Roßdeutscher, Antje. 2000. Lexikalisch basierte formale Textinterpretation. Habilitationsschrift, Universit"at Stuttgart. Sandström, Gorel. 1993. When-Clauses and the Temporal Interpretation of Narrative Discourse. PhD thesis, University of Umea. Shiffrin, Deborah. 1994. Making a list. Discourse Processes, (17):377-406.

Hans Kamp

and Michael

Schiehlen

Temporal Location in Natural Languages

Summary This paper deals with the semantics of various types of noun phrases which refer to time, as well as some phrases of related categories, such as common noun phrases and prepositional phrases. A prominent place play "calendar expressions", expressions such as year, month, day, hour, February, Wednesday, Easter, etc. The semantics we pursue in this paper is a model-theoretic one and its first part (Section 2) is devoted largely to the development of a model concept that can serve as the foundation for what is to come after. Since calendar terms take centre stage in our explorations, it is crucial that they can be given plausible interpretations in the models we use, models which reflect the typical speaker's knowledge about time in general and about the meanings of those terms in particular. This makes it necessary to look quite carefully into the models' underlying time structures, which serve as the substratum for the model-theoretic interpretation of calendar terms and other temporal expressions. Besides, the model theory of temporal concepts developed here is also intended as the semantic basis for a conceptually plausible definition of "time-logical" consequence, which incorporates general knowledge about time and about the calendar. The second part (Sections 3, 4, and 5) deals with the interpretation mechanisms of a variety of time-referring expressions, mostly at an informal or semi-formal level. A good deal of attention is paid to the context-dependent aspects of various temporal expressions. One of our aims in this part is to attain a better understanding of the different kinds of context sensitivity, and in particular of the difference between anaphora and indexicality. We have had to abandon our original plan of concluding the paper with an explicit model-theoretic treatment of a representative "fragment" of temporal NPs and PPs (generated by a set of phrase structure rules from a reasonably sized lexicon). We hope that the informal analyses in the second part will give a fairly clear idea of how such analyses work in general.

1

Introduction

The observations reported in this paper originated in the speech-to-speech translation project Verbmobil, in which German utterances were translated on-line into English. The dialogues which Verbmobil aimed at translating were, for most of the VM project, of a quite restricted kind: Their exclusive aim was to fix time and place of one or more future meetings between the dialogue participants. In these dialogues reference to (future) times was - obviously - a central issue, and not surprisingly, "calendar"

182

Hans Kamp / Michael

Schiehlen

expressions like der siebte Oktober (the seventh of October), die dritte Maiwoche (the third week in May), Montag zwischen zwölf und zwei (Monday between twelve and two), nächster Mittwoch (next Wednesday), an diesem Mittwoch (that Wednesday), etc. play a particularly prominent part. The translation of such expressions from German into English is by and large not especially problematic. For the semantic intricacies of the German phrases and their English counterparts usually match, so that a translator need not worry about them. Nevertheless there are a number of intricacies connected with these expressions of which we became aware only when Verbmobil led us to think more carefully about the systematic aspects of their interpretation. Salient among these - in our own current state of mind and in the following pages which reflect it - are in particular two points: (i) The difference between purely semantic and epistemic factors in our use of such expressions - it is common for us to be unsure whether, say, the third of next month is a Wednesday (and, by implication, the first Wednesday of next month) or some other day of the week, although the denotations of the terms the third of next month and the Grst Wednesday of next month are fully determined and we know in principle how to compute them; and (ii) The special character, within the temporal domain, of context dependence, which shows a blend of anaphora and indexicality not known from the by now well-known studies of third person as opposed to first and second person pronouns. In order to account for these and other aspects of time-referring expressions we found it necessary to provide ourselves with an explicit and detailed model-theoretic description of time as it informs human discourse. An important part of this description is concerned with the ways in which the temporal axis is partitioned by calendar predicates such as year, month, week, day and the like. We are aware that this (earlier) part of the paper makes for uncommonly tedious reading. But the model-theoretic set-up is indispensable for what follows, which we hope will not seemm tedious to quite the same degree. One pervasive aspect of the semantics of the NPs with which we will deal is contextsensitivity. This is especially obvious, and has been extensively discussed, for expressions like tomorrow or last week. But the phenomenon is much more pervasive than that; it attaches equally to the first of September, Sunday or half past ten. An occurrence of the Grst of September denotes a particular day only insofar as the context fixes which year is concerned; in the case of half past ten the context must determine the day to which the referent belongs; and to fix the reference of an occurrence of Sunday the context must determine a temporal domain containing a unique Sunday. We will look at the principles which govern these kinds of contextual resolution in Section 3. For now we only note that the context sensitivity of many temporal NPs is directly connected with the categorial ambiguity of expressions such as Sunday or Christmas. We have seen that these terms can be used as (context-sensitive) NPs. But they can also be used as common nouns, as in next Sunday, every Sunday, the Christmas before last, etc. One obvious connection between these two uses is that the referent of an NP-occurrence always belongs to the extension of the common noun. For instance, whatever day the NP Sunday denotes, it will always be a Sunday. So the semantic relationship between NP-occurrence and common noun is the same as that between occurrences of the NP the first of September and the complex common noun first of

Temporal Location in Natural

Languages

183

September. This latter relationship is one that has long been seen as crucial to any compositional semantics of these and other definite descriptions: the denotation of the first of September should be analyzed as determined by what the singular definite article the does with the meaning of the common noun first of September. The semantics we will present in this paper endorses this principle. And it will account, in an analogous way, for the reference of O-determiner NPs such as Sunday on the basis of the extension of the corresponding common noun. It is for this reason that we will begin our exploration of temporal expressions with a discussion of such common nouns and turn only after that to the problem of context-dependent NP reference. But before we can start any part of these investigations, it will be necessary to establish the general framework they presuppose. While the impetus for writing this paper came from our involvement in Verbmobil, in which Christian Rohrer has played a central part throughout, many of the paper's actual concerns go back to issues on which he and the first author worked in the eighties (and to which an entire chapter was devoted in the still unpublished work on temporal reference in French which they then wrote). Thus the following pages would seem a suitable tribute to someone in whose earlier work tense and time played so prominent a part and who then focused increasingly on the computational concerns of which his prominent role in Verbmobil has been one of many manifestations.

2

Model-theoretic Framework

2.1 Temporal Order and Metric Our approach follows the familiar tenets of model-theoretic semantics. Each expression of the part of natural language we will study has a denotation. This denotation depends on the one hand on the meaning of the expression and on the other on the factual properties of the world. By abstracting over the latter it is possible to focus on the former. One does this by treating the world as instantiating one from a class of possible models. Each possible model M represents one possible world (or a cluster of possible worlds between which it is unnecessary to distinguish in the context of the semantic investigation at hand). Given such a class of possible models we can identify the semantic context of an expression e with the function which maps each possible model M to the denotation of e in M. A precise characterization of this function presupposes a precise characterization of its domain - i.e. of the class of models to be considered. It is important that this class be "just right". On the one hand it should include among its models a counterpart to each of the possible contingencies that might have their effect on what the denotation of various expressions is. But on the other it should not be too big. Models which are inconsistent with semantic properties of and relations between expressions of the language under consideration, ought to be excluded.1 1

For instance, it seems reasonable that the English adjectives dead and alive are mutually

184

Hans Kamp / Michael Schiehlen

In actual practice model-theoretic treatments of natural language have usually followed the strategy of approximating the intended model classes "from above": When developing a model theory for a language or language fragment L one starts out with a model class C0 that includes more than it should and then restricts this class by adopting "meaning postulates", principles which capture certain semantic properties of and relations between expressions of L. This is more or less the way in which we will proceed. We will formulate various conditions that our models will have to satisfy. As more such "meaning postulates" are added, some of the models that were previously admitted will be eliminated. The question what models should be included in the model class that is needed for our purposes can be divided into several parts. The first and most basic question concerns the structure of the temporal order and its metric. Here we encounter a difficulty which will confront us equally when we come to address the remaining questions: How are we to determine "the" way in which language users conceive of time? It may be expected that the conceptions that different speakers have of time do not fully coincide, so that the best we can hope for is to capture something like the "conception of the typical speaker". But who qualifies as the typical speaker? That is hard to tell. And in the absence of extensive empirical investigation as to what it is that different speakers assume about time, one is likely to identify what one takes to be the typical speaker's conception with one's own intuitions. We realize that more likely than not we are victims to such prejudice ourselves and that the proposals we will make about what should be regarded as a conceptual truth about time (and thus about what holds in all models of our class) are probably not free from some measure of arbitrariness. We do not think, however, that this is an issue that seriously affects what else we will have to say. The first assumption we make about the conceptual properties of time is that it is a dense, linearly ordered medium. Perhaps the intuitively most satisfying way of describing this is as a structure consisting of "periods" which are linearly ordered by complete precedence and overlap and in which all periods (excepting the minimal, "punctual" elements of time, if these are to be called periods at all) are infinitely divisible. One can also think of the periods as collections from an antecedently given point set. This is arguably more in keeping with mathematical tradition. But for our present purposes the choice between these two approaches does not really matter. (See e.g. van Benthem 1984 or Kamp 1979 for discussion of how the two approaches are related.) What will be important, however, - no matter which approach is chosen - is the distinction between the non-minimal periods and the minimal ones. We will refer to the latter as moments of time and to the former simply as periods. More precisely, we will assume, as a basis for further model-theoretic considerations, a topological time structure consisting of both moments and periods, in which the moments are arranged in a dense linear ordering Sl% ~ A and hence h P

(y)S* ~ A

a contradiction to the consistency of p. In this case let A 0 = {P,S^A}

an d let R0 = { ( ( « i , . . .

2. P is not existential. In this case let A 0 = {P} and R0 = 0 .

is consistent.

Temporal Logic and Quantifier Translations CASE n

249

+ 1

Assume A„, Rn have been defined and A n and Rn are consistent and finite. We now define A n + i , R n + i . Consider a „ and ~ a n . One of these two is consistent with A n . Denote by fi n a one which is consistent. We distinguish several subcases, according to the form of (5n. 1. Case (3n = (Ey)S£A(y,x). We let u be a new variable not in A n or Rn or ¡3n and let A n + 1 = A n U {/3n,S^%A} U T n where r„ = W e l e t Rn+1

| S*(y)

1 and n > 1.

A Oa, if m > l,n = 0

= 0(a A ¡3), if n = m = 0.

In any event, the above case analysis shows that Of\Ym^aiJ j is equivalent to a conjunction of pure formulas

•

Temporal Logic and Quantifier Translations

259

E x a m p l e 5.6 Let us put in normal form the formula a =

• ~ D 0 ( Y 2 ~pi VY ~ p 2 V ~ p i ) o • ~ • ( O J ' ' 2 ~pi VQY ~ p 2 v O ~pi) •H- • ~ D ( y ~ piV ~ p2 V 0 ~ pi) D 0 ( V p i A p 2 A Dpi) D(pi A 0 ( p 2 A Dpi)) = a'

a

Note the translation

A Yp2 A p i )

of a' back into 1 Q 0 as given in Example

4-2.

T h e o r e m 5.7 ( T r a n s l a t i o n of I K Y i n t o K ) There exists a faithful occurrences of Y.)

translation

a

> [ a ] from I K Y into K (i.e. eliminating

all

Proof Let a be any wff of I K Y . By the previous theorem a can be assumed to be in normal form a = y' f \ Y n ^ a i < 3 i j where a i j contains no occurrences of Y, i.e. they are pure K formulas. We shall rewrite each wff f\j YniJ Qjj' cLS a different formula without occurrences of Y. Consider a formula (3 of the above form: fc-i ¡3 = (3k A / \ Y m ' P j 3=0

where (3j for 0 < j < k, is a pure K formula. P can be rewritten as ¡3' where

Clearly ¡3' is equivalent to (3 where mk = 0 and m j = £¿>¿71*, for 1 < j < k. To motivate our syntactical translation, let tk be a point in a model with tk 1= (3Figure 1 below describes the state of affairs. We have

A> to

Pi £o,i

Xo,m~i t\

0k-i tk-1

Pk a^fc-i.i

Xk-i,nk-!-i

tk

Figure 1: to ~< xo,i -..,-< with tj t= f3j.

—l (BA • h A implies h BA.

BB)

3. Axioms for making Ji projection operators (a) Ji is a 'next' modal operator: • Ji{A A B) • JiA A JiB

tautologies.

Dov M. Gab bay

266 • Ji ~ A

JiA

• h A implies I- JlA (b) Ji are projections: • Jj((3 o Ji/3),i,j = l,...,k • JiJi/3 Jif3, i = 1,..., k (c) Interaction

with B, Q:

• Ai