231 64 2MB
English Pages [208] Year 2016
Working Memory and Second Language Learning
SECOND LANGUAGE ACQUISITION Series Editors: Professor David Singleton, University of Pannonia, Hungary and Fellow Emeritus, Trinity College, Dublin, Ireland and Dr Simone E. Pfenninger, University of Zurich, Switzerland This series brings together titles dealing with a variety of aspects of language acquisition and processing in situations where a language or languages other than the native language is involved. Second language is thus interpreted in its broadest possible sense. The volumes included in the series all offer in their different ways, on the one hand, exposition and discussion of empirical findings and, on the other, some degree of theoretical reflection. In this latter connection, no particular theoretical stance is privileged in the series; nor is any relevant perspective – sociolinguistic, psycholinguistic, neurolinguistic, etc. – deemed out of place. The intended readership of the series includes final-year undergraduates working on second language acquisition projects, postgraduate students involved in second language acquisition research, and researchers, teachers and policy-makers in general whose interests include a second language acquisition component. Full details of all the books in this series and of all our other publications can be found on http://www.multilingual-matters.com, or by writing to Multilingual Matters, St Nicholas House, 31-34 High Street, Bristol BS1 2AW, UK.
SECOND LANGUAGE ACQUISITION: 100
Working Memory and Second Language Learning Towards an Integrated Approach
Zhisheng (Edward) Wen
MULTILINGUAL MATTERS Bristol • Buffalo • Toronto
To My Mother (With all my love and all my memories therewith!) To Amanda, Carrie and Shanie (With all my love and all my loyalties therewith!)
Library of Congress Cataloging in Publication Data A catalog record for this book is available from the Library of Congress. Names: Wen, Zhisheng, author. Title: Working Memory and Second Language Learning: Towards an Integrated Approach / Zhisheng (Edward) Wen. Description: Bristol; Buffalo: Multilingual Matters, [2016] Series: Second Language Acquisition: 100 | Includes bibliographical references and index. Identifiers: LCCN 2016005125| ISBN 9781783095728 (hbk : alk. paper) | ISBN 9781783095711 (pbk : alk. paper) | ISBN 9781783095759 (kindle) Subjects: LCSH: Second language acquisition--Psychological aspects. | Memory. | Psycholinguistics. Classification: LCC P118.2 .W668 2015 | DDC 418.001/9--dc23 LC record available at http://lccn.loc.gov/2016005125 British Library Cataloguing in Publication Data A catalogue entry for this book is available from the British Library. ISBN-13: 978-1-78309-572-8 (hbk) ISBN-13: 978-1-78309-571-1 (pbk) Multilingual Matters UK: St Nicholas House, 31–34 High Street, Bristol BS1 2AW, UK. USA: UTP, 2250 Military Road, Tonawanda, NY 14150, USA. Canada: UTP, 5201 Dufferin Street, North York, Ontario M3H 5T8, Canada. Website: www.multilingual-matters.com Twitter: Multi_Ling_Mat Facebook: https://www.facebook.com/multilingualmatters Blog: www.channelviewpublications.wordpress.com Copyright © 2016 Zhisheng (Edward) Wen. All rights reserved. No part of this work may be reproduced in any form or by any means without permission in writing from the publisher. The policy of Multilingual Matters/Channel View Publications is to use papers that are natural, renewable and recyclable products, made from wood grown in sustainable forests. In the manufacturing process of our books, and to further support our policy, preference is given to printers that have FSC and PEFC Chain of Custody certification. The FSC and/or PEFC logos will appear on those books where full certification has been granted to the printer concerned. Typeset by R. J. Footring Ltd, Derby, UK Printed and bound in Great Britain by Short Run Press Ltd
Contents
Foreword viii Preface and Acknowledgements xi Abbreviationsxv 1
Introduction and Overview Research Scope, Themes and Issues Outline of the Book
1 1 5
Part 1: Theoretical and Methodological Foundations 2
Working Memory Theories and Models Evolution of WM and the Standard Model Current WM Models and Controversies Toward Unified Theories of WM An Integrated Model of WM
11 11 14 20 24
3
Working Memory Measures and Issues The Simple Memory Span Tasks The Complex Memory Span Tasks Theoretical Issues Surrounding WM Measures Methodological Issues Besetting WM Measures Summary
26 26 28 31 34 37
Part 2: Research Syntheses of Working Memory in L1 and L2 Learning 4
Working Memory in First Language Research WM in L1 Acquisition and Vocabulary Development WM in L1 Listening and Reading Comprehension v
43 44 48
vi Contents
WM in L1 Speech and Written Production WM in Linguistic Theories and Language Processing Models Summarizing the WM–L1 Association 5
Working Memory in Second Language Research Theoretical Perspectives on WM and SLA Empirical Studies of WM and SLA General Findings of the Current WM–SLA Studies Critique of the Current WM–SLA Studies Summarizing the WM–L2 Association
51 52 57 59 59 63 71 73 76
Part 3: Toward an Integrated Perspective on Working Memory and SLA 6
An Integrated Framework for Working Memory and SLA Research Reconceptualizing and Redefining WM in SLA Research Putting SLA Domains, Skills and Processes in Better Order Toward an Integrated Framework for WM in SLA Research Basic Tenets and Empirical Consequences of the Integrated Framework Summary
79 79 81 82 87 89
7
Working Memory in L2 Acquisition and Processing: The P/E Model WM in the L2 Acquisitional and Developmental Domains WM and L2 Sub-Skills WM in L2 Acquisition and Processing: The P/E Hypothesis General Principles for Applying the P/E Model Summary
8
Working Memory and Tasks in L2 Speech Performance Key Issues and Debates in Current L2 Task Performance Research Toward a WM Perspective on L2 Tasks and Performance Future Research on WM, Tasks and L2 Speech Performance Summary
118
Working Memory and Language Aptitude in L2 Development A Critical Review of the Current Language Aptitude Models Toward a WM Perspective on L2 Aptitude Future Research on Language Aptitude and WM in SLA Summary
131 131 135 140 144
9
91 91 99 107 113 117
118 122 128 129
Contents vii
10 Conclusions and Implications for Future Research Implications of the Integrated Approach Toward an Interdisciplinary Research Agenda for WM and SLA Additional Research: WM, Language Aptitude, SLA and Beyond Concluding Remarks
146 146
References Index
156 183
150 153 154
Foreword Peter Skehan
It is a great pleasure to me to write this Foreword to the book by Zhisheng (Edward) Wen. Edward completed his doctoral studies some years ago under my supervision at the Chinese University of Hong Kong (CUHK). I was blessed with a wonderful cohort of research students at CUHK before my retirement, and now I have the ultimate reward of one of them publishing a major monograph in the field of applied linguistics. The book has brought together a number of literatures, and thus can make important contributions to applied linguistics. Most obviously, the book fills an important niche in applied linguistics by providing a thorough analysis of the construct and the measurement of working memory. There have many journal articles in this area in recent years, and much progress in the neighbouring field of cognitive psychology. But we have needed a clear, up-to-date and comprehensive account of the area, as it applies to second language acquisition. This book fulfils that need. It reviews the general evidence and underpinning theory. This represents a real service for our field, where we tend to be more focused on applications than on basic issues. The updating here is timely, indeed. The book also examines the considerable literature that has been written specifically from the perspective of second language acquisition. In doing this, it uses wider reviews, for example of working memory theories and measures, to demonstrate some of the shortcomings of existing research. But then the book provides an extensive and extremely useful review of measurement and related issues. Edward argues that second language research using working memory measures has to become much more sophisticated than it has been in the past. Indeed, one of the problems with some current second language research is its dependence on partial measures of working memory. In the book, the argument is strongly made that from now on the different components of working memory, and in particular its phonological component (PWM) and its executive aspects (EWM), have to viii
Foreword ix
be conceptualized and assessed separately, due to their distinctive implications for specific SLA domains and processes. The book, though, is not simply a review. It also provides innovative theoretical insights and practical guidelines for conducting empirical studies in this increasingly important line of inquiry. In the book, working memory theories are integrated effectively with current research in second language acquisition, such as the role of task planning and the study of second language task-based performance. In doing this, Edward is rather neatly linking two of my own areas of interest, which I (on reflection, oddly) have kept separate – those of individual differences, in this case working memory, and task-based speech performance. I am extremely grateful to Edward for making such a link where I have not! So the focus is not simply on exploring working memory for its own sake. The overarching objective is to explore whether and how working memory interacts with task features/ conditions in the way they impact on second language performance. This Edward is able to argue, but he postulates that the effects can be subtle and therefore they should be approached by more carefully designed research. In so doing, he is bound to push the task performance research field into new and entirely desirable directions, which will have important consequences. One particularly distinctive and innovative aspect of this book is what Edward terms the ‘integrated perspective’ on research into working memory and second language acquisition. This conceptual framework portrays working memory as containing multiple components with embedded functions, and he uses this to demonstrate the importance of fractionated measurement if the construct of working memory is to be applied effectively in second language research. Previous research has been bedevilled by invalid composite measurements or partial measurement. The integrated perspective brings out the inadequacies of these approaches. But Edward also argues strongly for taking account of domain generality versus domain specificity in constructing working memory span tasks for second language studies, and this has important implications for future research. We have to concern ourselves with verbal material in second language acquisition, and verbal material which is closely linked to the dynamic L1–L2 scenarios as participants’ L2 proficiency grows. This too will have a strong impact on how working memory is to be conceptualized and effectively measured in second language research in the future. The integrated framework, though, is only the starting point for some major and original claims which are made in the book, and which go to the very heart of the linkage between working memory and second language learning and processing. Edward argues that the two key components, phonological working memory and executive working memory, relate to different aspects of L2 sub-domain acquisition and L2 sub-skills processing. To that effect, he posits a link between phonological working memory and the acquisitional and developmental aspects of SLA (such as L2 lexis,
x Foreword
formulaic sequences and morphosyntax), with these being based on a capacity to access chunk-based (or exemplar-based) language. This suggests that their complexity is underpinned by a sound-processing capacity. In contrast, he suggests that executive working memory is more likely to be associated with attention regulation and monitoring mechanisms that likely affect offline processing activities and real-time performance aspects in L2 interaction and production. With Edward’s so-called phonological/ executive (P/E) model in place, we now have a much finer-grained account of how L2 learners’ individual differences in working memory are related to different facets of second language acquisition, processing, performance and development, a claim which is uniquely Edward’s! Edward ends the book with the major claim that working memory occupies a central place in most contemporary language aptitude models, and therefore it should be included as a key aptitude component in explaining the different developmental stages of SLA. Most of the hypotheses as laid out in the P/E model have been about L2 acquisition and processing, but here he returns to the area of L2 development, and proposes that there is considerable evidence that working memory differences are dynamic ally implicated in longer-term development and change. His claim is that working memory is foreign language aptitude! This is the most effectively argued version of this claim that I have come across and this is a suitably strong way to end an excellent contribution to the interdisciplinary inquiry of second language acquisition. Peter Skehan Twickenham, London, 2015
Preface and Acknowledgements
This book represents a personal survey and interpretation of the role of working memory (WM) in first and second language acquisition (SLA). Through this book, I hope to bring together and integrate major areas of my current research interests, namely WM research in cognitive science and SLA research in applied linguistics. Before I embarked on this project 10 years ago, I intended to write an updated and extended version of Susan Gathercole and Alan Baddeley’s seminal book Working Memory and Language (1993). Greatly motivated by their pioneering work, I have always held the view that WM should play a greater role in second language learning than in first language acquisition, and that the very topic of working memory and second language learning should deserve an independent book-length volume in its own right. My long-held conviction on this WM–SLA enterprise was greatly reinforced by Professor Alan Baddeley’s opening speech (titled ‘Working memory at 40’) delivered at the 2014 International Conference of Working Memory held at Cambridge University, which marked the 40th anniversary of the seminal multi-component model of WM (i.e. Baddeley & Hitch, 1974). Most encouraging and relevant of all, Baddeley concluded his speech by predicting that ‘WM and second language learning’ would be one of the more promising areas of WM research in the next decade. I am very pleased and grateful indeed to Professor Baddeley for sharing his insights and vision for the WM–SLA enterprise with me during my short stint at Cambridge. Of course, I am even more grateful to Alan for contributing the introductory chapter to my recent co-edited volume (Wen et al., 2015) on this same topic (Baddeley, 2015), and for his endorsement written for this current volume as well (as printed on the back cover). Above all, I wish to thank Professor Baddeley for introducing the classical WM model, which has inspired most of the WM–SLA explorations in applied linguistics and also laid the theoretical foundation for the integrated perspective on WM and SLA that I propose in this book. xi
xii Preface and Acknowledgements
On the other side of the WM–SLA enterprise, many of the ideas developed in this book have been greatly influenced by the pioneering work on SLA of my advisor and mentor, Professor Peter Skehan. To begin with, Professor Skehan underscored the central role that WM plays in the major developmental stages of SLA in his ground-breaking monograph A Cognitive Approach to Language Learning (1998). Building on this assumption, in this book I aim to specify and delineate how WM can be conceptualized and operationalized in line with recent research insights from the cognitive sciences to make it more applicable to practical SLA research. In this sense, this book can be regarded as a supplement and expansion to Skehan’s major volume. Indeed, I owe a great personal debt to Peter with respect to my academic endeavours and career development. First, I wish to thank him for sharing his pioneering research on language aptitude and L2 task performance in his impeccable courses during his tenure at CUHK, which immediately lured me into this rather challenging yet extremely rewarding line of research. I am also grateful to Peter for his meticulous supervision and constant guidance during my academic pursuits from the early years until the current day. His never-failing endorsements of my WM–SLA ambition have meant so much to me. Above all, his insights and scholarship on memory and aptitude have been instrumental and fundamental in shaping my broad perspective on viewing ‘WM as language aptitude’ in SLA. I hereby extend my utmost gratitude to Peter for his never-failing mentorship over the years and a special ‘thank you’ for the Foreword he has written for this book. In addition to my great thanks to Alan Baddeley and Peter Skehan, I wish to place on record my sincere gratitude to two other long-term research collaborators, Professor Mailce Borges Mota (at the Universidade Federal de Santa Catarina in Brazil) and Dr Arthur McNeill (at the Hong Kong University of Science and Technology), for their enduring collegial encouragement and support. It has been my great pleasure and fortune to have known and collaborated with Mailce and Arthur on a number of rewarding projects on WM and SLA (e.g. Wen et al., 2013, 2014, 2015). To a great extent, Mailce and Arthur have reassured me that I never need to feel (too) lonely in pursuing this relatively unpopular research topic in Hong Kong (although I recently left and moved to Macao). Therefore, I also extend a heartfelt ‘thank you’ to Mailce and Arthur for accompanying me on this otherwise lonely journey. I also wish to extend my gratitude to the many other admirable cognitive psychologists cum SLA researchers who have helped me to reshape, revise and refine my integrated perspective on WM and SLA over the years. First, I would like to thank Professor Nick Ellis at the University of Michigan for the Morley Scholar Fellowship that he so generously offered me in 2011. Indeed, the summer I spent at the English Language Institute (ELI) in Ann Arbor was a most memorable and rewarding overseas experience for me. Most importantly, it allowed me to benefit substantially from Nick’s penetrating
Preface and Acknowledgements xiii
insights on the relationship between WM and the acquisition of linguistic sequences in first and second language acquisition, which have subsequently transformed me into an ardent follower of connectionism, construction grammar, and usage-based SLA approaches. As such, I attribute a great proportion (50%) of my P/E model (the ‘P’ part) to Nick! Of course, my sincere thanks also go to other faculty and staff members at the ELI (inter alia, Ute Römer, Diane Larsen-Freeman, John Swales and Veronica Moore) for making my stay there such an enjoyable and unforgettable experience (let’s go blue!). I also wish to acknowledge my gratitude to Professor Michael Ullman from Georgetown University for his sparkling insights into the neuro cognitive basis of memory systems and SLA. I want to thank Michael in particular for agreeing to present his keynote and a workshop at the Language Learning roundtable on ‘Memory and Second Language Acquisition’ that I convened and organized with Mailce and Arthur in 2012 in Hong Kong – see Wen et al. (2014) for more details. I have to admit that Michael’s multidisciplinary perspective on long-term memory and SLA (as encapsulated in his indomitable D/P model) is so ingenious it provided the inspiration for the name for my own WM–SLA model (i.e. the P/E model). Therefore, I am sincerely thankful to Michael for having greatly broadened my vision and research scope in pursuing the WM–SLA enterprise. I wish to also thank the many other respectable and knowledge able scholars and colleagues (in both the cognitive sciences and applied linguistics) for sharing their insights on relevant issues in WM, language aptitude and SLA over the years. These include, among many others, Brian MacWhinney, Nelson Cowan, Randall Engle, John Williams, Peter Robinson, Alison Mackey, Alan Juffs, Michael Harrington, Pauline Foster, Emma Marsden, Lourdes Ortega, Rod Ellis, Cem Aptekin, Mohammad Ahmadian, Clare Wright, Adriana Biedron, Melissa Baralt, Tan Lihai, Dong Yanping, Shen Jiaxuan, Zhao Lun, Wang Chuming, Lu Zhi, Li Shaofeng, Yang Jing and Yi Baoshu. Many thanks to you all for your insights, support and encouragement! Finally, the series editors, David Singleton and Simone E. Pfenninger, and Laura Longworth at Multilingual Matters and the several rounds of anonymous reviewers (at different stages of this project) certainly deserve my heartfelt gratitude for their incredible patience, meticulous comments, constructive suggestions and guidelines, which have significantly improved the quality of this book. The editors from Amstrong-Hilton Limited as well as the production editor, Ralph Footring, also deserve my heartfelt thanks for their professional and careful proofreading of the manuscript at its final stage. Of course, the usual proviso always applies: I am solely responsible for all of the remaining limitations and shortcomings. I wish to end my acknowledgements by extending both heartfelt thanks and apologies to the four most important women in my family. First of all,
xiv Preface and Acknowledgements
I have to thank my most beloved mother (Pan Lianying) for having been so supportive of my whole life. I do owe her an apology though for having not been able to publish this book in time before she had to leave us (though I am sure she is happy for me to get it done finally). Then, my thanks also go to my wife Amanda, for taking care of our family (amazingly efficiently) during my extended preoccupation with this book over the past 10 years. I simply cannot thank Amanda enough for her ever-increasing love to me over the last 25 years (so blessed), for her enormous understanding and support and, most important of all, for her great sacrifice for me and for the family (by quitting and changing her jobs so frequently to fit my academic pursuits that can sometimes go real wild and wide!). No doubt, I carried a strong sense of guilt whenever I needed to lock myself either in my office or at home on numerous weekends and holidays while writing and revising this book (‘such a relief to finally see it going into print’). There were some particularly difficult moments when I had to relentlessly force our two little angels, Carrie and Shanie, to move away (often amidst vulnerable tears in their begging eyes!) from my work desk right in the middle of their favourite cartoons (sorry ‘Peppa Pig’ and ‘Barbie Princess’!) and never-tiring songs (sorry Taylor Swift and the TFBOYS!). It is therefore to these four most important women in my life that I dedicate this book (and with it all my loyalties from both my heart and my every pocket!). Zhisheng (Edward) Wen Macao Polytechnic Institute, Macau, 2015
Abbreviations
CAF CANAL-F CI D/P DST EB EWM Hi-LAB ID(s) L1 L1A L2 LAA LCDH LST LTM LT-WM MGG MLAT NWR OST P/E model PFC PSTM PWM RST SAS SI SLA
complexity, accuracy, and fluency (framework) cognitive ability in acquisition of language – foreign consecutive interpreting declarative/procedural (model) digit span task episodic buffer executive (aspects of) working memory High-level Language Aptitude Battery individual difference(s) first language first language acquisition second language language analytical ability linguistic coding difficulties hypothesis listening span task long-term memory long-term working memory mainstream generative grammar Modern Language Aptitude Test non-word repetition span task operation span task phonological/executive (WM–SLA) model prefrontal cortex phonological short-term memory phonological working memory (phonological memory) reading span task supervisory attention system simultaneous interpreting second language acquisition xv
xvi Abbreviations
SST STM ST-WM TBLT VS-STM VS-WM WM WMC
speaking span task short-term memory short-term working memory task-based language teaching visuospatial short-term memory visuospatial working memory working memory working memory capacity
1 Introduction and Overview
Research Scope, Themes and Issues Language aptitude and working memory in SLA Even casual observations in our daily lives tell us that some people are able to learn a foreign or second language (L2) easier, faster and/or better than others (Grigorenko et al., 2000; Segalowitz, 1997). This common phenomenon is best captured by the concept of language aptitude in applied linguistics. By definition, L2 aptitude presupposes that ‘there is a specific talent for learning foreign languages which exhibits considerable variation between individual learners’ (Dörnyei & Skehan, 2003; Skehan, 1998). This underlying assumption of an L2 aptitude was put to considerable test as early as the 1950s and 1960s, mostly with respect to the research done by John Carroll on US military personnel (see Spolsky, 1995; Stansfield, 1989). From the 1970s, however, research on L2 aptitude languished, with ‘relatively little empirical work and little theorizing’ taking place during the next three decades (Skehan, 2002: 69). This lack of research interest stemmed partly from several major criticisms levelled against the very concept of language aptitude per se (for more detailed discussion see Dörnyei & Skehan, 2003; Skehan, 1998, 2002). The first accusation was related to the anti-egalitarian ‘labelling effect’ of a concept that assigns the label of ‘loser’ to anyone who gets a low aptitude score (e.g. from the Modern Language Aptitude Test, or MLAT; Carroll & Sapon, 1959). The second accusation targeted the ‘indecent origin’ of the outdated teaching methodology used during the heyday of early language aptitude research (i.e. the audiolingual method, which was dominant in the 1950s when Carroll conducted most of his aptitude research). As a result of these accusations and other criticisms, such as Krashen’s (1980) verdict that aptitude predicts only explicit learning, not language acquisition), there was little research on L2 aptitude from the 1970s until the 1990s (Wen et al., 2016). 1
2 Introduction and Overview
In recent years, however, research on L2 aptitude has managed to gain some renewed momentum (Granena & Long, 2013; Granena et al., 2016; Li et al., 2015; Robinson, 2002). Intriguingly, this new body of research has consistently contradicted the criticisms made of the concept of language aptitude (Skehan, 2015b). For example, instead of being ‘outdated’ and ‘ineffective’, the concept of L2 aptitude is now being viewed as being very relevant to L2 learning, even in today’s prevailing communicative L2 classrooms (Granena, 2013; Gregersen & MacIntyre, 2014; Skehan, 2015b; Vatz et al., 2013). More importantly, research on second language acquisition (SLA) has increasingly confirmed that L2 aptitude is not just confined to traditional instructional settings but is also relevant under different learning conditions (e.g. implicit versus explicit; Granena, 2015) and different learning contexts (Robinson, 2007). In a recent meta-analysis of the empirical research conducted over the past 50 years, Li (2015) provided compelling evidence of a positive link between L2 aptitude and L2 grammar learning. Nonetheless, this renewed wave of research interest has been accom panied by concerted calls among researchers to reconsider and reconceptualize the construct of L2 aptitude (Ganschow & Sparks, 2001; Granena, 2013; Kormos, 2013; Parry & Child, 1990; Robinson, 2002b, 2005, 2012; Skehan, 1998, 2002, 2012, 2015b, 2016; Wen & Skehan, 2011). The current research on L2 aptitude is being actively pursued by scholars from multiple discip lines – educational psychology, applied linguistics, cognitive science and neuroscience (Wen, 2012b; Wen et al., 2016). Through these research efforts, a multitude of L2 aptitude models have been proposed, such as the linguistic coding difficulties hypothesis (LCDH) by Sparks and colleagues, the cognitive ability for novelty in the acquisition of language (CANAL-F) model by Grigorenko, Sternberg and colleagues, Peter Skehan’s macro-SLA aptitude model and the ‘aptitude complexes’ model by Peter Robinson. Recently, cognitive scientists and neuroscientists have made significant contributions to L2 aptitude research by proposing innovative models from their own theoretical perspectives, including the impressive high-level aptitude battery (Hi-LAB) model and neurological and brain network-based aptitude models (Wen, 2012b; Wen et al., 2016). Most relevantly, a consistent theme that has emerged either directly or indirectly from this renewed research interest is the proposal to incorporate the cognitive construct of working memory (WM) as a central component of L2 aptitude (e.g. Aguado, 2012; Hummel, 2009; Kormos, 2013; Linck & Weiss, 2015; McLaughlin, 1995; Miyake & Friedman, 1998; Sawyer & Ranta, 2001; Skehan, 1998, 2002, 2012; Wen, 2007, 2012b; Wen & Skehan, 2011; Williams, 2012, 2015; Yoshimura, 2001). This proposal has garnered increasing attention in recent years from SLA scholars interested in language aptitude research (e.g. Dekeyser & Koeth, 2011; Juffs & Harrington, 2011; Kormos, 2013; see also Ellis & Shintani, 2013; Mitchell et al., 2013; Singleton, 2014). This innovative conception of ‘WM as L2 aptitude’ thus constitutes
Introduction and Overview 3
the overarching theme of this book, and is fully discussed in Chapter 9. In other words, the first and foremost motivation of this book is to review and evaluate the extent to which the cognitive construct of WM plays a central role in SLA as an aptitude component.
L2 task-based planning and performance research in SLA Situated within the postulation of ‘WM as L2 aptitude’, this book is also partially motivated by the ongoing debate in current research on L2 task-based language learning and teaching (Robinson, 2011; Skehan, 2011, 2014, 2015a, 2015c). In the realm of L2 task-based planning research, for example, an early paper by Rod Ellis (1987) is generally regarded as seminal in that it triggered enormous enthusiasm among SLA scholars to examine the variegated effects of planning on L2 task-based performance (R. Ellis, 2005; Skehan & Foster, 2012). These SLA scholars include Rod Ellis, Michael Long, Graham Crookes, Peter Skehan, Peter Robinson and Martin Bygate, among many others (Bygate, 2015). These scholars have published a number of empirical studies on this topic, culminating in an edited volume (R. Ellis, 2005) and a special issue of the journal Applied Linguistics, led by a review article by Rod Ellis (2009). Overall, these studies have adopted various perspectives to investigate the different effects of planning, either independently or in combination with various task features/designs and implementation variables, and explore its effects on L2 task performance with respect to the three global dimensions of complexity, accuracy and fluency (i.e. the ‘CAF’ framework; Housen & Kuiken, 2009; Housen et al., 2012; cf. Lambert & Kormos, 2014). Indeed, most of the hypotheses regarding the effects of planning on L2 task performance have been borne out in empirical studies (R. Ellis, 2005; Skehan, 2014, 2015a). For example, giving L2 learners time to plan before executing a task normally results in the learners developing more fluent and complex speech (Foster & Skehan, 1996; Mehnert, 1998; Nielson, 2014). However, what is more intriguing and controversial is the inconsistent results observed with accuracy measures, which subsequently give way to two competing theoretical views on the cognitive mechanisms underlying these discrepancies in L2 task performance (Révész, 2014). For instance, Skehan (1998, 2009, 2014, 2015a, 2015c) postulated a ‘limited attention capacity’ theory epitomized by a ‘tradeoff hypothesis’ that assumes competition for cognitive resources with respect to complexity, accuracy and fluency (especially between the first two). In contrast, Robinson’s (2001, 2011, 2012, 2015) ‘cognition hypothesis’ advocates a ‘multiple-resources’ view of attention and processing, under which the learner is empowered with enhanced capacity to attend to more than one area of language performance (e.g. when prompted by manipulating the cognitive complexity of the task). In other words, if the task is complex enough, it is possible to
4 Introduction and Overview
expect improved performance in all three performance areas, that is, more complex, accurate and fluent speech. When this debate is examined from the broader perspective of cognitive psychology, the controversy can be interpreted as reflecting two rather different epistemological stances on the function of the cognitive construct of ‘attention’ and its consequential effects (which sometimes interplay with planning) on L2 speech performance. Given this unresolved issue, it is conceivable that a clearer understanding of the relevant cognitive functions supporting L2 speech planning and performance is necessary before such disputes can be resolved and a consensus reached (Nielson, 2014; Révész, 2014). Regarding the cognitive underpinnings of L2 task performance, Rod Ellis (2005) appeared to side with Skehan’s stance on the ‘limited attentional capacity’ hypothesis. In this respect, Ellis further highlighted three possible cognitive constructs that presumably influence L2 learners’ speech performance under planning conditions, namely the noticing/attention mechanism (Schmidt, 1990), the focus on form mechanism (Doughty, 2001) and limited WM capacity (Baddeley, 2003a). Among the three constructs, it can be argued that WM emerges as the most pivotal resource in regulating and modulating the effects of the other two mechanisms on speech planning and task performance. Indeed, some preliminary studies following this line of inquiry have demonstrated that WM plays an important role in mediating various L2 task features/designs and thus ultimately affecting L2 speech performance (e.g. Ahmadian, 2012, 2013; Fortkamp, 1999, 2003; Kim et al., 2015). Despite this initial evidence, the assumption that planning can compensate for WM limitations in L2 task performance remains largely inconclusive and merits further examination (Nielson, 2014). In this sense, a second motivation of this book is to elucidate the cognitive underpinnings of L2 task planning and speech performance by specifying the possible effects of the WM functions independently or in combination with the task features or designs. Hopefully, this WM perspective on L2 task performance will help shed light on the current ‘tradeoff–cognition’ debate in research on task-based language learning and teaching (Robinson, 2015; Skehan, 2015c). This second theme figures prominently in Chapter 8, which demonstrates how the integrated WM–SLA perspective can shed light on more focused research on the intricate relationships between WM, tasks and L2 speech performance.
Summary of the motivations for this book In recent years, an increasing number of SLA researchers have examined the role of WM in different areas of SLA (e.g. Juffs & Harrington, 2011; Sagarra, 2013; Wen, 2012a; Williams, 2012). An increasing number of
Introduction and Overview 5
empirical studies on SLA have also begun to converge on the pervasive effects of WM on L2 learning processes and outcomes – see for example Linck et al. (2014) for an updated comprehensive research synthesis and meta-analysis and Wen et al. (2013, 2015) for two collections of recent empirical studies. Nonetheless, given the myriad WM models (e.g. Miyake & Shah, 1999), complicated by the daunting number of currently available WM span tasks (e.g. the non-word repetition span, the reading span and the operation span task) from the feeder discipline of cognitive psychology, SLA researchers are likely to face and experience confusion in applying the WM construct in their research designs and methodologies. Indeed, the lack of consensus on the WM construct among SLA researchers and the inconsistent WM measures implemented in the current WM–SLA empirical studies are already imposing a considerable challenge for those SLA researchers who are seeking to replicate this research (such as L2 interaction studies; Gass & Valmori, 2015) and make it even more difficult to systematically compare their results across studies (Gass & Lee, 2011; Juffs, 2006; Juffs & Harrington, 2011; Linck et al., 2014). In the worstcase scenario, WM–SLA studies may suffer severe limitations and even pitfalls in their research designs and methodologies, which could result in potential fallacies and caveats in research practice (Linck et al., 2014; Wen, 2012a, 2014, 2015). To address some of these issues besetting current WM–SLA research and to further advance this interdisciplinary enterprise, in this book I aim to introduce a more principled approach to conceptualizing and operationalizing WM in SLA research. In pursuit of this goal, I first survey, then synthesize and finally integrate research insights from the cognitive sciences (in particular, WM research in cognitive psychology) and applied linguistics (in particular cognitively oriented SLA research) to advocate an integrated perspective on WM and SLA research. Then, expanding on this conceptual framework, I propose a theoretical model of WM and SLA (the phonological/executive model) that not only integrates and accommodates empirical evidence in the current WM–SLA research, but also provides an overarching framework for orienting future WM–SLA explorations.
Outline of the Book As mentioned above, the overall aim of this book is to offer a more principled approach to situating the cognitive construct of WM within SLA research. To that end, the book is organized as three parts. Following this introductory chapter, Part 1 comprises two chapters (Chapters 2 and 3) that lay the theoretical and methodological foundations of the book as a whole. More specifically, Chapter 2 first traces the evolution and development
6 Introduction and Overview
of the WM concept in multiple disciplines of the cognitive sciences that have provided sources for the WM theories used in most SLA research. The chapter provides a critical review of the current theoretical models and the controversies surrounding the multiple perspectives on the construct. The review culminates in consensual and nomothetic conceptions of the WM construct (as consisting of multiple components/functions) that integrate multidisciplinary perspectives from major fields of the cognitive sciences (e.g. psychology, biology, neuroscience, computer science, anthropology and philosophy). It is argued that such unifying characterizations of the WM construct are essential for providing a viable theoretical foundation for conceptualizing the WM construct for application in more practical areas (such as in general education and language learning). Following the discussion of the WM theories in Chapter 2, Chapter 3 discusses the methodological issues related to the WM measures and their assessment procedures. After describing the prevailing versions of the WM span tasks in the two major research paradigms in cognitive psychology, which include the simple memory span tasks (e.g. the digit span task and the non-word repetition span task) and complex memory tasks (e.g. the reading span task and the operation span task), the chapter also discusses some of the controversial issues besetting the wide array of WM measures with regard to their construction and implementation procedures. The chapter concludes by aligning the two versions of the WM span tasks (simple versus complex) with the putative underlying cognitive functions that are associated with the two distinct WM components (the phonological and executive components). After outlining the theoretical and methodological foundations of WM in Part 1, in Part 2 (comprising Chapters 4 and 5) I synthesize specific strands of research on the role of WM in first and second language learning. Chapter 4 reviews the more specific lines of cognitive psycholinguistic research on WM and first language acquisition (L1A). A review of this body of WM–L1 research reveals two distinct research traditions that have gradually emerged on the two sides of the Atlantic (the European and North American paradigms), with each adhering to its own research focus on a certain WM component or area (phonological versus executive) and distinct research methodologies for the WM assessment procedures. The analysis subsequently converges on the initial links between the two key language-relevant WM components, namely phonological working memory (PWM) and executive working memory (EWM). This dichotomy serves as the precursor for the integrated conceptual framework of WM and SLA proposed in this book. In Chapter 5, I first outline the major theoretical assumptions held by some SLA researchers that depict the potential links between the effects of WM and L2 acquisition and processing. Then, I further synthesize the current WM and SLA research to provide a state-of-the-art review of this
Introduction and Overview 7
area. The review not only points to the positive theoretical links between WM and the essential components of SLA, but also reveals the inherent shortcomings and caveats in the theoretical conceptualizations and research methodologies in current research practice. It is argued that a major factor in these shortcomings stems from the current disputes and controversies over the nature of the WM construct and the lack of a standardized assessment procedure for its measurement. Therefore, it is argued that a more principled approach to conceptualizing and measuring WM is needed in SLA research. To resolve the theoretical and methodological issues in the current WM– SLA research, in Part 3 of the book I present an integrated perspective on WM and SLA research. In Chapter 6, an integrated conceptual framework for WM and SLA research is presented. The chapter begins by redefining the WM construct in SLA research based on the unified theories (reviewed in Chapter 2) and proceeds to provide a detailed account of the integrated conceptual framework, including its structure and constitutive WM components. The chapter also highlights the theoretical and methodological implications of this integrated perspective for future WM–SLA research. Having defined the integrated conceptual framework, Chapter 7 begins by demonstrating how this integrated approach can help to reconceptualize and reframe specific areas of WM–SLA research. Drawing on the basic tenets of the integrated framework and emerging patterns with respect to the distinct effects of the two major WM components (phonological and executive), the chapter formulates an integrated theoretical model that serves to align each component and its associated functions with their likely affected SLA domains and processes. The resulting phonological/executive (P/E) model encapsulates hypotheses that capture these juxtapositions. Given that the research on L2 planning and L2 task-based performance now occupies a dominant position in SLA, Chapter 8 further explores how the proposed integrated WM perspective (i.e. the P/E model) can help illuminate the theoretical debate on the ‘tradeoff–cognition’ hypothesis postulated to underlie L2 task performance (e.g. as indexed by the three dimensions of complexity, accuracy and fluency; i.e. the CAF framework). Because most of the participants in L2 task studies are college students who have already obtained intermediate and post-intermediate L2 proficiency (to complete a task), it can be argued that effects of PWM are minimal, while EWM can be expected to exert more influence on the selective areas of L2 task performance that rely on attention-regulating and attention-monitoring mechanisms. Therefore, it can be claimed that most of the hypotheses regarding WM–L2 task performance can be restricted to the effects of EWM. In terms of L2 task performance, it is further postulated that EWM should be linked more closely to the fluency measures (that subsume lexical retrieval efficiency) and accuracy measures that are likely to draw on the monitoring and self-repair mechanisms. The chapter also calls for distinctions to be made between the
8 Introduction and Overview
main effects, interaction effects and threshold effects of WM in relation to the task characteristics and implementation procedures when designing future studies of WM–L2 task performance. In Chapter 9, the focus shifts from L2 acquisition, processing and performance to the relationship between WM and the longer-term development of L2 within the broader context of language aptitude research. After critically reviewing the current L2 aptitude models, the chapter proposes a reconceptualization of L2 aptitude from a WM perspective by elaborating the rationale, feasibility, perceived advantages and possible limitations of the concept of L2 aptitude. The chapter thus argues that WM should be incorporated as a central aptitude component and that the relationship between WM and language aptitude should be reconfigured within the SLA developmental stages. Chapter 10 concludes the book by first recapitulating the theoretical and methodological ramifications of the integrated perspective of WM for nuanced SLA research. As such, the nature and major characteristics of SLA, including its acquisitional and developmental domains and essential processing and performance components, are thus put into better order. These domains and components are then further aligned with the underlying and modulating cognitive mechanisms of the multiple WM components and functions, particularly those associated with PWM and EWM. Ultimately, the call is made for more concerted efforts from multiple disciplines, including the cognitive sciences and SLA, to explore the complexities and intricacies of WM and SLA. To facilitate these multidisciplinary efforts, an integrated research agenda is also proposed for the shared goals of arriving at a deeper understanding of human cognition and bilingualism. It is therefore my final hope that the integrated account of WM and SLA proposed in this book will shed light on the complex relationship between WM and SLA and ultimately inform L2 learning, training and classroom practice.
Part 1 Theoretical and Methodological Foundations
2 Working Memory Theories and Models
As an inextricable component of the ‘cognitive leap’ in the evolution of modern human thinking (Coolidge & Wynn, 2009), working memory (WM) has attracted enormous research interest from multiple disciplines in cognitive science since the inception of the seminal model of Baddeley and colleagues (Baddeley, 1986, 2007; Baddeley & Hitch, 1974). These waves of research have been accompanied by the propagation of a dozen WM models (Miyake & Shah, 1999) and inherent controversies and heated debates about the nature, structure and functions of the WM construct (Baddeley, 2007, 2012; Cowan, 2014). This chapter, which is divided into four sections, reviews these lines of interdisciplinary WM research in cognitive science. In the first section, an evolutionary perspective is used to trace the development of the WM concept. The following section summarizes the differences and commonalities of the major theoretical models of WM in cognitive psychology, culminating in nomothetic and consensual characterizations of the construct. The third section widens the interdisciplinary scope by incorporating neuropsychological evidence corroborating these unified theories of WM. The last section concludes the chapter by recapitulating the key characterizations of the WM construct that lend theoretical support to the integrated framework of WM proposed in this book.
Evolution of WM and the Standard Model The concept of WM can be traced back to as early as William James’s (1890) demarcation of ‘primary memory’ and ‘secondary memory’ (Logie et al., 2007). According to James, primary memory (synonymous with today’s concept of short-term memory, STM) temporarily stores a limited amount of information that is consciously accessible for use in higher-level cognitive 11
12 Part 1: Theoretical and Methodological Foundations
activities. This temporarily stored information is then either displaced by new information and forgotten, or transferred via (sub-vocal) rehearsal to a secondary or long-term memory (LTM) that presumably has unlimited storage capacity. In a similar vein, Hebb (1949) suggested that STM was the result of the temporary activation of neural connections between clusters of cells, while LTM resulted from permanent synaptic changes (also see Andrade, 2014: 93). These binary views on STM and LTM were revived in the 1950s and 1960s, riding the waves of the ‘cognitive revolution’ (Baddeley, 2012; Miller, 2003). Among all the events that marked the onset of this cognitive revolution, George Miller’s (1956) seminal quantification of the limited capacity of STM, that is, the ‘magical number seven plus or minus two (7±2)’ hypothesis, added much fuel to the research interest in STM. More recently, however, Nelson Cowan (2001, 2005, 2014) cast doubt on Miller’s original proposition and reinterpreted the magical number to be only around ‘four plus or minus one (4±1) chunks of information’ that we can actively hold in our trailing consciousness. The dual-store view of memory gained even more momentum in the late 1960s with the advent of Atkinson and Shiffrin’s (1968) influential ‘modal model’ of information processing. According to this model, information from multiple modalities (visual, audio and spatial, etc.) passes through a parallel series of sensory registers, before entering the limited-capacity short-term store/memory, from which information is soon lost, unless rehearsed in a timely manner. As conceived in this model, the short-term store (synonymous with STM) is responsible for not just encoding and elaborating information, but also for feeding it into and out of the long-term store (LTM), thus positing an intimate connection between STM and LTM. Despite its early appeal, Atkinson and Shiffrin’s model was not without its limitations and criticisms (as cited in Baddeley, 2000b, 2012). First, the underlying assumption that merely holding items in STM would guarantee learning failed to gain empirical support, which gave way to Craik and Lockhart’s (1972) alternative ‘levels of processing’ hypothesis. Second, the purported critical role of STM envisioned by Atkinson and Shiffrin was not reflected in the data from patients with STM deficits, who demonstrated apparently intact long-term learning and successful management of their everyday lives. To address these problems, Alan Baddeley and Graham Hitch (1974) proposed the abandonment of the unitary account of STM and postulated a multiple-component system (comprising a phonological loop and a visuospatial sketchpad, both regulated by a supervisory system, i.e. the central executive), which they labelled ‘working memory’. Thus, as conceptualized by Baddeley and colleagues, WM subsumed multiple components (the so-called M-WM model; Baddeley, 2012) and was governed by a supervisory attentional system (SAS) and two domain-specific slaves/buffers. The SAS, or in other words the central executive, was assumed to be responsible
Working Memory Theories and Models 13
for various executive functions, such as controlling and allocating attentional resources among the cognitive processes implicated in higher-level cognition. The phonological loop, one of the two domain-specific slaves/ buffers, was said to temporarily store and hold sound-based information via an articulatory rehearsal process. The other domain-specific slave/buffer, the visuospatial sketchpad, was considered to be responsible for handling visual and spatial information. Later, Baddeley (2000b) considered that a storage component should be attached to the central executive, and added a fourth component, the episodic buffer, to the original tripartite central executive model. The new component served to integrate various pieces of information from all kinds of sources (and multiple modalities) into limited episodes or chunks. Conceived this way, the episodic buffer was synonymous with its long-term namesake, ‘episodic memory’, postulated by Tulving (2002; as cited in Coolidge & Wynn, 2009: 221). Despite its apparent simplicity, the original tripartite model (Baddeley & Hitch, 1974) has proven to be an extremely powerful framework for addressing a range of questions on higher-level human cognitive activities, such as academic learning, speech and text comprehension, and prospection and future planning (Carruthers, 2013). Baddeley’s tripartite framework served as a standard WM model for the subsequent waves of research in multiple fields within the cognitive sciences, and effectively straddled the six constituent disciplines outlined by Miller (2003), namely, anthropology, linguistics, psychology, computer science, philosophy and neuroscience (Carruthers, 2013; Conway et al., 2007). Accordingly, it is not surprising that in a recent discipline-wide poll, the seminal model of Baddeley and colleagues was unequivocally ranked as one of the top ‘100 most influential works in cognitive science’ (Saito & Towse, 2007; Conway et al., 2007; also see Wynn & Coolidge, 2010a). Indeed, in recent years, cognitive anthropologists and philosophers have engaged in interesting discussions on the evolutionary dimensions of WM (e.g. Carruthers, 2013, 2015; Coolidge & Wynn, 2009; Coolidge et al., 2013; Wynn & Coolidge, 2010a, 2010b). It has even been postulated that human WM, or ‘an enhanced WM capacity’, constitutes a human species-unique feature that marks the second major leap in the cognitive revolution of modern human thinking (with the first major leap being the transformation from tree life to ground life). In this sense, the capacity for WM facilitated the evolution of humans from primitive Neanderthals to modern and intelligent Homo sapiens (Coolidge & Wynn, 2005, 2009; Russell, 1996), thus placing WM in a central position within human reasoning and cognition (Carruthers, 2015; Picciuto & Carruthers, 2013).
14 Part 1: Theoretical and Methodological Foundations
Current WM Models and Controversies As mentioned above, Baddeley and Hitch’s (1974) early work on WM was seminal in that it triggered considerable interest in WM among cognitive psychologists and other scientists who subsequently formulated a dozen other WM models based on their own research scope and perspectives (Carruthers, 2013, 2015; Conway et al., 2007; Dehn, 2008; Dehn et al., 2015; Miyake & Shah, 1999). Nonetheless, from the early 1970s to the mid-1990s, this line of research was conducted by two distinct camps on either side of the Atlantic: many Europe-based cognitive psychologists adopted Baddeley’s ‘structural’ view of WM, whereas their North American counterparts mostly embraced the more ‘functional’ views, such as the ‘embedded-processes model’ of Cowan (1999, 2005) and the ‘controlled attention model’ of Engle (2002; Engle et al., 1999a). Fortunately, these otherwise distinct theoretical conceptions of WM were brought together in two collected volumes edited by Miyake and Shah (1999) and Conway et al. (2007), respectively. Notably, in the volume edited by Miyake and Shah, almost all of the active WM research camps in all key areas of the cognitive sciences were commissioned to contribute a chapter explaining their own perspective on the WM concept and its theoretical model. Their answers were solicited from eight common sets of questions regarding the nature, structure and function of WM. In addition to a chapter by Baddeley and Logie (1999) on the classic multiple-component WM model, nine other WM camps (mostly based in the US) contributed their theoretical accounts by responding to the same sets of questions pre-designated by the editors of the book. These various models, together with the eight common questions, are summarized in Table 2.1. Table 2.1 WM models and their proponents (based on Miyake & Shah, 1999) Authors
Model
Cowan
The embedded-processes model
Engle, Kane & Tuholski
The ‘controlled attention’ framework
Lovett, Reder & Lebiere
The ACT-R model
Kieras, Meyer, Mueller & Seymour
The executive-process/interactive control (EPIC) model
Young & Lewis
The SOAR architecture
Ericsson & Delaney
The long-term working memory (LT-WM) framework
Barnard
The interactive cognitive sub-systems (ICS) model
Schneider
The controlled and automatic processing (CAP2) architecture
O’Reilly, Braver & Cohen
The biologically based computational model
Working Memory Theories and Models 15
As the editors, Miyake and Shah (1999), cogently argued, the underlying rationale and most likely the greatest advantage of such a ‘common question approach’ was that the commonalities and differences between these diverse WM models could now be more easily identified – indeed, Conway et al. (2007) used the same approach. However, a cursory reading of the individual chapters (Chapters 2–11) contributed by the WM teams suggests that the book as a whole leads nowhere, for the answers provided by the authors present diverse theoretical descriptions of the WM concept. For example, the conceptions of the construct appear to be particularly discrepant on issues concerning the unitary versus non-unitary nature of WM (Question 3), the underlying construct that accounts for the limitations of WM (Question 4) and on the utilization of complex tasks that WM purportedly plays a role in (Question 5). These diverse views are further illustrated in Table 2.2. Overall, there are a number of limitations with the current WM models that need to be overcome (Dehn, 2008; Dehn et al., 2015). According to Dehn (2008), the most apparent problem with the current WM models is the lack of a clear distinction between the three terms STM, WM and LTM. Although some models have succeeded in emphasizing the interactions between WM and STM or LTM, they have either ignored STM or LTM, or, in the worst-case scenario, broadened the concept of WM to such a degree that it is almost meaningless (e.g. by equating WM with almost all general executive functions). Thus, as Dehn cogently argued in his integrated model of WM, the construct of WM is best understood as ‘the management, manipulation, and transformation of information drawn from short-term and long-term memory’ (Dehn, 2008: 51). In this sense, WM is not about the (temporary or long-term) storage or memory of information, but rather a cognitive process whose primary function is to facilitate and enhance the capacity of ‘encoding, storage, and retrieval’ functions that are essential for learning and higher-level human cognitive activities (Dehn, 2008: 58). Another and more serious limitation of the current WM models is the lack of an account of the actual work conducted by WM, which should include not only executive functions (which are already delineated in most of the models) but also non-executive functions (Dehn, 2008: 56–57). These non-executive functions, which can be referred to as ‘WM operations’, encompass a whole list of manipulation and transformation functions, such as: encoding information into LTM (e.g. semantically); associating new information with existing LTM representations; transforming information, such as verbal recording of visually perceived material; and grouping related items into categories. More importantly, these WM operations are both conscious and unconscious. In other words, WM functions may be able to operate below the level of consciousness (cf. Soto & Silvanto, 2014). Baddeley’s postulation of the episodic buffer fits within these descriptions of the WM operations and Cowan’s ‘focus of attention’.
16 Part 1: Theoretical and Methodological Foundations
Table 2.2 Commonalities and differences among the current WM models, part I WM models
Unitary versus non-unitary system(s)
Sources causing limitations in WM
Multi-component model Domain-specific systems
Information decay; efficiency of controlled attention
Embedded-processes model
Domain-specific?
Information decay
Controlled attention framework
Domain-general aspects
Efficiency of controlled attention; lack of inhibitory control
ACT-R model
Domain-general aspects
Limits in availability of activation
EPIC model
Modality-specific codes
Information decay
SOAR architecture
Domain-specific knowledge or Lack of skill or knowledge skills for efficient encoding and retrieval
LT-WM framework
Domain-specific knowledge or Lack of skill or knowledge skills for efficient encoding and retrieval
ICS model
Modality-specific codes
Limitations in communication and interaction among different sub-systems or subcomponents
CAP2 architecture
Modality-specific codes
Limitations in communication and interaction among different sub-systems or subcomponents
Biological model
Domain-specific modules
Interference
Summary
Disagreed on (1) the number of different specific codes or systems; and (2) the sources of domain-specific effects
Disagreed on (1) the scope of limitations; (2) the inherent versus emergent nature of limitations; and (3) the modifiability of limitations
Despite these seemingly diverse descriptions of WM (e.g. as comprising multiple components or as a unitary executive control system), when the answers from each contributing team were pooled together and put into a broader perspective, in Chapter 12 (Kintsch et al., 1999), the scale of the disagreement detected earlier diminished significantly. Despite the apparent inconsistencies in the answers to the eight pre-designated questions, the 10 WM models were actually not that substantially different from each other in principle. In most cases, the so-called disagreements can be interpreted as indicating differences in research scope or focus (e.g. computational/ formal versus conceptual approaches) and different definitions (e.g. defining
Working Memory Theories and Models 17
Table 2.2 Commonalities and differences among the current WM models, part II WM models
Role of WM in cognitive Basic mechanisms and activities (examples of complex representations tasks)
Multi-component model Immediate serial recall; WM span tasks (reading span)
Encoding, maintenance, retrieval, representation
Embedded-processes model
Immediate serial recall; WM span tasks (reading span)
Encoding, maintenance, representation
Controlled attention framework
WM span tasks; WM span tasks (operation span)
Encoding, maintenance, representation
ACT-R model
Algebra problem solving
Encoding, maintenance, representation
EPIC model
Immediate serial recall
Encoding, maintenance, representation
SOAR architecture
Understanding and modifying Representation computer programs
LT-WM framework
Chess playing
Encoding, maintenance, representation
ICS model
Random number generation
Maintenance, representation
CAP2 architecture
(Not mentioned)
Representation
Biological model
AX-CPT and N-back tasks
Encoding, representation
Summary
Disagreed on the ‘complexity’ of tasks regarding whether they require (1) controlled processing; (2) coordination of multiple steps of processing; or (3) multiple components of the WM system
Although disagreeing on some basic mechanisms, such as encoding, maintenance and retrieval, they generally agreed on the representational issues
attention either as selective control or as mental energy) rather than absolute disagreements on fundamental issues. This is more clearly the case for the last five issues (commonalities and differences – column headings) listed in Table 2.2. Based on these comparisons, it can be argued that the current WM models converge more than diverge on fundamental issues regarding the nature, structure and function of WM (Shah & Miyake, 1999; Baddeley, 2012; Cowan, 2014). To sum up, this section has outlined the development of the theories and current models of WM in the cognitive sciences and some of the inherent controversies. Table 2.3 summarizes the characterizations of the three most representative perspectives on the WM construct as conceived by the three arguably most influential WM models applied in psychology and education, namely Baddeley’s multi-component model, Engle and Kane’s attentional control model and Cowan’s embedded process model (see Fenesi et al.,
18 Part 1: Theoretical and Methodological Foundations
Table 2.2 Commonalities and differences among the current WM models, part III WM models
Control and regulation of WM
Relationship between WM and attention
Multi-component model
Central control mechanism
Attention as subset of WM
Embedded-processes model
Central control mechanism
Selective attention as subset of WM
Controlled attention framework
Central control mechanism
WM = STM + controlled attention
ACT-R model
As emergent property
WM = attentional resources?
EPIC model
Strategies
Serving separate functions
SOAR architecture
As emergent property
Attention as emergent property
LT-WM framework
Skilled strategies
Selective attention
ICS model
Self-regulation
Awareness as subset of WM
CAP2 architecture
Regional and central control
Serving separate functions
Biological model
Auto-control
Attention as subset of WM
Summary
Most models agreed on this homunculus-like element explicitly (2,3,4) or implicitly (the rest)
Differences seemed to arise with respect to different areas of attention: selective attention vs. mental resources
Table 2.2 Commonalities and differences among the current WM models, part IV WM models
Links with LTM
Biological implementation
Multi-component model
Clear distinction, separate function
Prefrontal cortex (PFC): neuropsychology
Embedded-processes model
Continuous relationship
PFC: event-related potential (ERP) studies
Controlled attention framework
Continuous relationship
PFC
ACT-R model
Continuous relationship
PFC: neuropsychology
EPIC model
Structural distinction
Not mentioned
SOAR architecture
Structural distinction?
Not mentioned
LT-WM framework
Structural distinction?
Not mentioned
ICS model
Structural distinction
Not mentioned
CAP2 architecture
Clear distinction
Neuroscience findings
Biological model
Continuous relationship
Neuroscience findings
Summary
Differences in the way they explained the relationship; no model argued for a strict structural separation.
Unanimously agreed on the association between the PFC and executive function of WM
Multi-component model (Baddeley)
WM capacity reflects the limit of the domain-specific stores involved in processing. Episodic buffer has a capacity restriction of four chunks
Central executive, two sub-systems (visuospatial sketchpad and phonological loop), and episodic buffer
Less emphasis placed on attention. Central executive coordinates resource allocation between two sub-systems and integrates information
Visuospatial sketchpad and phonological loop are domain-specific stores that retain their respective information
LTM is a functionally distinct system, which integrates active representations via the episodic buffer
Key features
Capacity limitation
Structure
Domain generality (attention)
Domain specificity
LTM integration
Very close connection between WM and LTM
Acknowledgement of domain-specific stores (visuospatial, phonological), but emphasizes domain-generality as critical rather than specialized buffers
Attentional control drives active maintenance of goal-relevant information, especially during interference
No suggested hierarchy: LTM traces active above-the-threshold processes for achieving and maintaining activation, and limited-capacity controlled attention
WMC reflects the capacity for controlled, sustained attention during interference or distraction (not about storage)
Attentional control model (Engle & Kane)
WM is an activated subset of LTM (not a separate system)
Acknowledgement of domain-specific stores, but WM reflects attentional focus on activated sets of representations in STM or LTM
Focus of attention drives efficiency of WM processing, activation and maintenance of mental representations, attention switching and inhibition
Hierarchically embedded subsets of memory: activated portions of LTM in response to STM cues, and subset of STM in the focus of attention
WM capacity is limited but flexible (not specific to any processing domain). Capacity is approximately four chunk
Embedded process model (Cowan)
Table 2.3 Conceptions of the nature, structure and functions of WM in the three principal theoretical frameworks
Working Memory Theories and Models 19
20 Part 1: Theoretical and Methodological Foundations
2015). More importantly, the review of the seemingly disparate WM models culminates in nomothetic theories of WM that further give rise to unifying characterizations of the WM construct regarding its signature limited capacity in span and duration, its multiple components and executive functions, and its two-way interactions with contents from STM and LTM. Therefore, the next section elaborates these unified theories of the WM construct.
Toward Unified Theories of WM As Miyake and Shah (1999) state early in their book (p. 4), another important feature of the ‘common-question approach’ adopted in the edited volume is that it should help to achieve the ultimate goal of arriving at consensual theories of WM. By comparing and analysing the answers of the different WM camps, as in the concluding chapter by Kintsch et al. (1999), the editors were able to extract six common consensual themes from the seemingly diverse WM models. The six themes thus serve as essential pointers to nomothetic theories of WM that have implications for research on the effects of WM on any domain of human cognition (e.g. see Carruthers, 2013, 2015). (1) WM is not a structurally separate ‘box’ or ‘place’ in the mind or brain. All of the models denounce the purely structural point of view and consider WM from functional and/or control-oriented perspectives (Cowan, 2015). They all agree that WM should not be considered as a separate ‘box’ for shortterm storage that is structurally distinct from other memory systems (such as sensory STM and LTM; Baddeley, 2012; Cowan et al., 2012; Diamond, 2013). Rather, the WM system comprises multiple dynamic sub-systems that contribute to its multiple executive functions of sustaining, rehearsing, manipulating and inhibiting a limited amount of information that is readily accessible for serving high-order cognitive functions (Carruthers, 2013, 2015). (2) The maintenance function of WM serves complex cognition. All models agree that the maintenance function of WM is an active process that is closely related to cognitive processes that serve complex activities such as learning, language processing, prospection and future planning, and logical thinking and reasoning (Carruthers, 2013, 2015). That is, WM is not just for ‘memorizing’ per se and thus distinguishes itself from its previous incarnation (i.e. STM). (3) Executive control is integral to WM functions. As discussed above, all models agree that WM is not just about ‘memory’ per se but involves the ‘control’ and ‘regulation’ of cognitive actions, that is, the ‘working’ areas. According to Baddeley (1993b), WM could have been called ‘working
Working Memory Theories and Models 21
attention’ had he originally adopted the attention approach. In their original M-WM model, Baddeley and Hitch (1974) described executive functions such as information updating and multiple task switching as being taken up by the central executive component (the supervisory attentional system, SAS). In a similar vein, other WM theorists, particularly those based in North America such as Engle et al. (1999a), perceive WM as ‘controlled attention’. In the same manner, in his ‘embedded-processes model’ of WM, Cowan (1999) conceived WM as ‘the focus of attention’ that is accessible to consciousness within activated LTM. In other words, Baddeley’s (2012, 2015) concept of WM can be deemed to be synonymous with Cowan’s (2014, 2015) model of selective and focused attention. Moreover, other WM models that have embraced the computational perspective, although not explicitly postulating this executive function, also admit the existence of this homunculus-like (attentional control) element of WM (Diamond, 2013). (4) The limited capacity of WM reflects multiple factors and may even be an emergent property of human cognition. All of the theoretical models of WM acknowledge its limited capacity as a signature feature (Carruthers, 2013, 2015; Miller & Buschman, 2015; Miller & Cohen, 2001), although the underlying causes of such limitations are still heatedly debated (for more recent accounts of this specific issue see Conway et al., 2007; Coolidge & Wynn, 2009). This limited capacity manifests itself in two ways (Cowan, 2014). First, the amount of information that WM can actively manage in our immediate consciousness is rather limited, specifically around seven unrelated units of information (Miller, 1956) or four chunks of information (Cowan, 2001). Second, information is held actively in WM only very briefly, usually for just a few seconds, and its contents decay gradually if not rehearsed in a timely manner. Two explanations of the source of this limitation have been proposed. One is that the limited capacity of WM is a result of the fixed amount of available cognitive resources for executing cognitive tasks (Daneman & Carpenter, 1980; Daneman & Green, 1986). The other view is that factors outside the cognitive system (such as strategies and skills) are involved. In addition to these two views, some have speculated that WM is an emergent property of the dynamic interactions of the mind and brain (Postle, 2006). (5) A completely unitary, domain-general view of WM does not hold. WM is not completely unitary and any comprehensive theory of WM should be able to account for some domain-specific effects (Baddeley, 2012, 2015). Even some so-called unitary domain-general models (such as those of Engle and Cowan) only reflect the authors’ research focus/emphasis, as they also admit that WM is probably neither an entirely unitary system nor entirely separable into domain-specific systems (Cowan, 2015). The emerging consensus thus reached among contemporary cognitive psychologists is that WM is a cover term that subsumes multiple structures, including the domain-specific mechanisms (e.g. sound and visuospatial) and domain-general executive
22 Part 1: Theoretical and Methodological Foundations
functions (information updating, switching and inhibition) that underpin many areas of human cognition (Williams, 2012). (6) Long-term memory forms an integral part of the WM system. All of the models of WM, either enthusiastically or cautiously, endorse the role of LTM in WM performance. For example, Cowan (1999) described WM as simply the consciously accessible ‘focus of attention’ embedded within the activated portion of LTM. In a similar vein, Ericsson and Kintsch (1995) proposed a concept of long-term WM (LT-WM) in which they argued that previously developed structures in the LTM serve as a means of boosting WM performance (e.g. as manifested in the superior performance of chess experts). However, Baddeley (2012: 18) regarded the WM–LTM relationship as an interactive one (rather than a separate kind of WM) and admitted that WM involves the activation of many areas of the brain that involve contents from LTM. Again, these views can be reconciled with neuropsychological evidence suggesting that WM and LTM (especially episodic memory) are intimately related and that ‘LTMs are not stored in a separate region of the brain (although the hippocampus does play a special role in binding together targeted representations in other regions)’ (Squire, 1992, as cited in Carruthers, 2013; Baddeley, 2012). As Joaquin Fuster, the leading neuroscientist of the prefrontal cortex, wrote in his Foreword to the edited volume by Alloway and Alloway (2013): [It] is now becoming increasingly accepted that WM and LTM share the same cortical networks, and that WM consists in the temporary activation of a LTM network updated for reaching a goal or solving a problem. This prospective, purposive aspect of WM, which was already in Baddeley’s definition, is what distinguishes WM from all other forms of STM. (Fuster, 2013: xii) Based on these emerging consensus views of WM, Miyake and Shah (1999) were able to arrive at an all-encompassing answer to the most fundamental question that initially prompted their edited volume on WM models, ‘What is WM, anyway?’ Based on the six consensual themes, they provided this overarching definition of WM: Working memory is those mechanisms or processes that are involved in the control, regulation and active maintenance of task-relevant information in the service of complex cognition, including novel as well as familiar, skilled tasks. (Miyake & Shah, 1999: 450, emphasis added) This definition of WM is all-encompassing in the sense that it captures the essence of all six consensual themes emerging from the diverse WM perspectives described above. This integrated model is further augmented by evidence from neuropsychological research. For example, Coolidge and
Working Memory Theories and Models 23
Table 2.4 WM components and functions with their corresponding brain regions and cortical areas WM components
Associated WM functions
Phonological WM
Phonological STM Left
Posterior parietal; inferior parietal; Brodmann’s area 40; supramarginal gyrus
Rehearsal/verbal Left WM
Broca’s area; anterior temporal frontal; premotor cortex
Visual STM
Right
Occipital lobes
Spatial STM
Right
Parietal lobes
Episodic STM
Left and right
Left hippocampus; right middle temporal
Executive WM Executive (central executive) functions
Bilateral
Dorsolateral prefrontal; anterior cingulate
LTM
Declarative LTM
Bilateral
The hippocampus; middle temporal lobes; Brodmann’s areas 45 and 47 (within and near Broca’s area)
Procedural LTM
Bilateral
Premotor cortex and Brodmann’s area 44 in Broca’s area in frontal cortex
Visuospatial STM Episodic buffer
Hemisphere Activated cortical areas
Wynn (2009: 43) portrayed the multiple structure/function view of WM that is supported by corresponding brain areas and cortical networks and processes. They also made a clear distinction between the multiple components of STM (phonological STM, visuospatial STM and the episodic buffer), executive functions or processes (executive WM, visuospatial WM, verbal WM and long-term retrieval) and the components of LTM (subsuming declarative knowledge and procedural LTM). In this extended WM model, permanent knowledge saved in the LTM is usually unconscious (or ‘preconscious’ at the very best), while information held in the WM (the three STM components at the middle level and the central executive at the top) is normally conscious (or ‘currently activated’ in Cowan’s terms). However, other scholars have recently pointed out that WM should also govern non-executive processes or operations (i.e. encoding, retrieving, chunking, transforming, consolidating, etc.; Dehn, 2008) that are unconscious (Soto & Silvanto, 2014), implicit (Hassin et al., 2009) or procedural (Oberauer, 2010). More importantly, the multiple-structure/function view of WM has received increasing support not just from cognitive psychology, but also other disciplines in the cognitive sciences, particularly neuropsychology. For example, evidence from case studies of patients (Gathercole & Baddeley, 1993) and recent neuro-imaging research seem to concur with the conclusions about the neuroanatomy of WM (Smith & Jonides, 1999; also see
24 Part 1: Theoretical and Methodological Foundations
Dehn, 2008). Specifically, research in the area has postulated that: (1) the phonological short-term store is subserved by the temporal lobes of the left hemisphere; (2) visuospatial memory is related to the right hemisphere; and (3) the central executive and its episodic buffer are usually associated with the dorsolateral prefrontal cortex. To give a more comprehensive picture of these WM–brain relationships, Table 2.4 provides an overview of these multiple WM components and their associated cognitive functions that are aligned with the corresponding brain regions and cortical networks reported to be activated during WM processes (Smith & Jonides, 1999; Ullman, 2001, 2005; also see Conway et al., 2009; Dehn, 2008).
An Integrated Model of WM To sum up, increasing evidence from multiple disciplines in cognitive science (anthropology, developmental and cognitive psychology, cognitive neuroscience, computer modeling and philosophy) concurs with the multiple components/functions view of WM. In this regard, it is argued here that WM is best conceived as a primary memory system (as opposed to LTM as secondary; James, 1890) for learning that functions as an interface between STM components (e.g. the phonological short-term store/memory, the visuospatial sketchpad and the episodic buffer) and LTM components (e.g. declarative and procedural memory), which in turn affects real-world action/cognitive tasks/activities (cf. Baddeley, 2015). More importantly, conceived this way, WM consists of both a set of executive WM (EWM) processes that comprise executive functions, such as maintenance, rehearsal (verbal WM), manipulation and inhibitory control (Carruthers, 2015), as well as non-executive or implicit WM operations (encompassing encoding, chunking, consolidating, restructuring and transforming, retrieving, etc.) that act upon information from the STM and LTM (Dehn, 2008; Dehn et al., 2015). This unified understanding not only incorporates and integrates the multiple perspectives of the current WM models, but also provides a clear distinction between STM, WM and LTM. In other words, they should all be regarded as ‘distinct and independent types of memory’ (Dehn, 2008: 50). It is hoped that, this unified approach can present a viable theoretical foundation for applications of the WM construct in diverse areas of human cognition (Andrade, 2014; Carruthers, 2015), including cognitive development and education (e.g. Cowan, 2014; Fenesi et al., 2015), general academic learning (Dehn, 2008; Dehn et al., 2015) and general language acquisition and processing (Baddeley, 2003a; Gathercole & Baddeley, 1993). To reiterate, the integrated multiple components/functions view of WM advocated here gives rise to three unifying characterizations that are
Working Memory Theories and Models 25
fundamental to the construct: (1) WM possesses limited capacity with respect to information-holding span and its duration; (2) WM is a primary memory system comprising both executive processes and non-executive operations; and (3) WM interacts with information from the STM and LTM bi-directionally, thus rending it an interface between STM and LTM on the one hand, and between cognition and real-world action on the other. It is conceivable that this integrated perspective of the WM construct carries great promise, given that the applications of WM straddle a broad range of disciplines in the cognitive sciences (e.g. psychology, linguistics, neuroscience, artificial intelligence, anthropology and philosophy; Miller, 2003) and other more applied domains of human life (Andrade, 2014; Carruthers, 2013, 2015; Conway et al., 2007). As such, it can provide a prototype conceptual framework for research on the effects of WM on specific domains of cognitive activity, such as cognitive development and education (Cowan, 2014; Fenesi et al., 2015) and general academic learning (Dehn, 2008; Dehn et al., 2015). The three unifying characterizations, as I demonstrate in the coming chapters, thus lend theoretical support to the integrated approach to WM in first and second language learning advocated in this book (in Part 2). Having clearly outlined the integrated view of the WM construct, the next issue is to examine how the construct has been operationalized and assessed in practical research to shed light on more specific areas such as language learning and processing. As such, in the next chapter I review the various measures and procedures that are currently available in cognitive psychology and psycholinguistics to assess WM.
3 Working Memory Measures and Issues
Following the general review of the models and theories of WM in Chapter 2, in this chapter I discuss the measures and assessment procedures for WM currently available in cognitive psychology. The chapter is organized into five sections. The first describes the storage-only versions of WM span tasks that are often adopted by European-based cognitive psychologists. The second section then reviews the more complex storage-plus-processing versions of WM span tasks that are usually implemented by their North American counterparts. In the following section I describe the theoretical debates surrounding these WM measures. I then explore some of the methodological issues besetting these measures in the fourth section. The summary section then concludes the chapter by pointing out how the research in cognitive psychology can be expanded to shed light on implementing WM span tasks in more practical research areas such as academic learning.
The Simple Memory Span Tasks Among the different WM components (as conceived in Baddeley’s model reviewed in Chapter 2), the phonological component (hereafter referred to as PWM; or the phonological loop as originally conceived by Baddeley and colleagues) is the most widely researched by European-based cognitive psychologists (Baddeley, 2012, 2015), so much so that it sometimes appears synonymous with their conception of the whole WM construct. Given the domain-specific functions of WM in temporarily storing and rehearsing sound-based materials, most of the WM researchers adopting this paradigm have focused on the relationship between PWM and language acquisition, in particular the acquisition and development of vocabulary (Gathercole, 2006; Gathercole & Baddeley, 1993; Gupta, 2012). According to Baddeley and Hitch’s (1974) structural view of WM, its PWM sub-component 26
Working Memory Measures and Issues 27
encompasses a capacity-limited phonological short-term store (limited in terms of the amount of information it actively holds and the brief duration of such holdings and an articulatory rehearsal mechanism, analogous to sub-vocal/inner speech), which prevents the actively held information from decaying (Baddeley, 2003a, 2012, 2015; Gathercole, 2007; Gathercole & Baddeley, 1993). In other words, speech encoded in the phonological storage buffer quickly fades if it is not rehearsed in a timely fashion. Several assessment methods have been developed to measure PWM, ranging from tasks involving immediate serial recall of single digits or numbers (the digit span task) and letters or words (the letter span or word span task). The more recently introduced non-word repetition span task has, though, proven to be the most reliable and most popular method (Gathercole et al., 1994) among European-based WM researchers (for a detailed keynote article on this simple memory span task, followed by commentaries from other researchers, see Gathercole, 2006). In a standard version of the non-word repetition span task (e.g. Gathercole et al., 1994), the non-words are generally formed from a string of letters that do not exist in the given language but still conform to its phonotactic rules (e.g. acklar and veincort; Cheung, 1996). The task usually includes two implementation procedures, one based on recognition only (the so-called non-word recognition span task, e.g. O’Brien et al., 2006, 2007) while the other is based on recognition plus repetition (the so-called non-word repetition span task) (Gathercole, 2006; Gathercole et al., 1994). It is self-evident that the non-word repetition span task differs from its recognition variant in that the latter does not necessarily involve articulatory processes to the same extent as the former, or at least not with the same force (Gathercole, 2006; also see Gathercole et al., 1994). In essence, the non-word repetition span task measures the accuracy in encoding and grouping phonological representations/forms in WM so that they can be consolidated and transformed into the long-term knowledge base. This factor is believed to be instrumental in learning any new sound-based language material required for various areas of language learning, and is most evident in vocabulary acquisition and grammatical development (Baddeley et al., 1998; Gathercole & Baddeley, 1989). Its critical role is particularly relevant and obvious when the participants are young children acquiring vocabulary in their native language (Andrade & Baddeley, 2011; Baddeley, 2000a, 2003a; Baddeley et al., 1998) and an L2 (French, 2006; Service, 1992; Verhagen et al., 2015). Overall, the non-word repetition span task can be considered to be a standard procedure for assessing the phonological aspect of WM (PWM) in cognitive psychology and other more applied areas such as learning disability research or language research. Thus, when the non-word repetition span task is implemented, a participant’s WM (PWM in this case) can be expressed in terms of an index number, which is the highest number of non-words that he or she can repeat correctly.
28 Part 1: Theoretical and Methodological Foundations
The Complex Memory Span Tasks The executive component of WM (EWM; or the central executive as originally conceptualized by Baddeley and colleagues) is the most important but least understood component in Baddeley’s (2000b, 2012) early tripartite WM model. Baddeley and Logie (1999) posited that WM plays various executive functions, such as coordinating the two slave systems, focusing and switching attention, and activating representations within LTM, but claimed that it was not involved in the temporary storage of information. That function was more recently proposed to be a component of the episodic buffer (Baddeley, 2000b). As discussed in the previous chapter, the consensus between the various WM models is that the function of this homunculuslike element is to control and allocate/regulate attentional resources (Miyake & Shah, 1999) by subsuming executive functions such as information updating, switching and inhibitory control (Miyake & Friedman, 2012). In light of this, in this section I review some of the conventional procedures for measuring these executive functions of WM and then in the following two sections discuss some of the theoretical and methodological issues surrounding these EWM measures.
The reading span task The most seminal and widely adopted measure of EWM is the reading span task (Daneman & Carpenter, 1980). In the standard version, participants are requested to read aloud increasingly longer sets of sentences and to further state the plausibility of each sentence (e.g. in terms of its grammaticality). The participants are also required to recall the final words of each sentence in their presentation order (Waters & Caplan, 2003). For example, at the two-sentence level (level 2), participants may be asked to read aloud the following sentences (cited from Miyake, 2001) and decide on their plausibility: (1) Due to his gross inadequacies, his position as director was terminated abruptly. (2) It is possible, of course, that life did not arise on the earth at all. At the end of this trial, the participants are expected to recall the final words of the two sentences, ‘abruptly’ and ‘all’. The results from such reading span tasks correlate reasonably highly with language comprehension scores, usually with correlation coefficients of 0.4~0.6 (Daneman & Merikle, 1996). In essence, Daneman and Carpenter’s reading span tasks are different from the traditional simple memory span tasks (as reviewed in the previous section) in that they essentially impose dual-task demands on
Working Memory Measures and Issues 29
the participants (i.e. simultaneous storage plus processing), thus reflecting the contemporary models of WM that require more than a passive storage of target memory items (Miyake & Shah, 1999).
The speaking span task Daneman and Green (1986) also developed a spoken version of the reading span task that ‘models after the reading span test in that it taxes processing while simultaneously imposing a storage task’. The speaking span task consists of presenting the participants with increasingly longer sets of unrelated words, which they have to read silently. At the end of a set, the participants are required to read aloud a sentence for each individual word presented, in the original order and form of the presentation. For example, in Daneman (1991: 449), the participants silently read a set of words displayed on individual computer video screens for one second each and then, at the end of the test, use each word to generate a sentence containing that word which they read aloud. For example, when presented with the set ‘shelter, muscles, dangers’, a participant could generate the following three sentences: (1) Trees provide poor shelter during a thunderstorm. (2) Mr Universe has very big muscles. (3) There are dangers associated with every occupation. The speaking span task in Daneman (1991) is constructed with 100 unrelated words, each of seven letters. The words are arranged in five sets, each of two, three, four, five and six words. Each word is presented individually at the centre of a computer screen for one second, and the participants are required to read that word silently. Ten milliseconds after the word is removed, the second word in the set is displayed and the participants read that word. The procedure is repeated several times and then a blank screen is displayed with an accompanying tone which signals that the trial has ended. The participants are then asked to generate and read aloud a set of sentences for each of the words in the set. The participants are asked to ensure that each sentence is grammatically correct, both semantically and syntactically. However, there are no restrictions on the length of the sentence generated or the position of the target word. The participants are encouraged to generate as many sentences as they can. The speaking span score is based on the total number of words for which a grammatical and meaningful sentence is generated, the maximum being 100. Thus, it can be argued that the speaking span task and the reading span task impose very similar storage requirements, namely, storing increasingly longer sets of unrelated words, and both provide measures of the number of words the participants can retain. However, the speaking span task relies
30 Part 1: Theoretical and Methodological Foundations
on the number of words that can be retained while generating sentences, whereas for the reading span it is the number of words that can be retained while reading and comprehending sentences. That is, the fundamental difference between the reading and speaking span tests is that the former taxes comprehension processes while the latter taxes production processes. As discussed in the next section, the issue of domain-specificity (e.g. reading or speaking, as opposed to domain-generality) has significant implications for designing WM span tasks in language research (especially in SLA research).
The operation span task The WM span as measured by the reading span task has been criticized as being language-related and a number of questions remain unanswered, for example whether the WM span is limited to only language processing or whether it has a more general capacity. Turner and Engle (1989) designed a study to address these issues. In so doing, they devised an operation span task in which one of the two competing processes is not language-based. The two processes of this operation span task are: (1) mathematical operations and (2) memorizing words. In the standard version, the participants first perform a simple mathematical operation while remembering a word that follows it for later recall. Each operation is presented with a word (although this can be a letter) and after each set of operations the participants are asked to recall the words presented within the same set (Engle, 2001). For example, the following are taken from a three-sentence set (from Engle, 2002): (1) Is (8/4) – 1 = 1? Bear (2) Is (6 × 2) – 2 = 10? Dad (3) Is (10 × 2) – 6 = 12? Beans It is clear that the reading span task and the operation span task differ in that the former comprises reading and word recall while the latter includes mathematical operations and word recall. However, Turner and Engle (1989) and Engle et al. (1992) found significant correlations between the two measures (0.4 ~0.6). Based on these results, the authors argued that WM capacity is a general ability used in the same way in language processing and other cognitive processes. That is, if language processing or learning is perceived as no different from any other human cognitive activity (as in N.C. Ellis, 2006), then this general-domain WM measure should be adopted (Sanchez et al., 2010). Otherwise, a domain-specific version of the reading span task should be implemented (cf. Dekeyser & Juffs, 2005).
Working Memory Measures and Issues 31
Theoretical Issues Surrounding WM Measures Despite their wide use in cognitive psychology, there are a number of important theoretical concerns relating to the above-mentioned WM span tasks (Dehn, 2008; Dehn et al., 2015; Miyake, 2001). In light of the controversies elucidated in the previous chapter on the theories of WM, two of these concerns deserve additional attention as they are also related to the WM measures. The first issue concerns the construct that underlies individual differences in WM and the second issue is related to the extent to which individual differences in WM are domain-specific or domain-general. Both issues are addressed in the present section.
The construct underlying individual variations in WM The first issue concerns the nature of the individual differences that are tapped by the WM measures. In other words, what do WM span tasks really measure and what makes them better predictors of people’s performance in complex cognitive tasks than the more passive, storage-oriented span tasks? Although these are fundamental questions that surround WM span research, there is no clear consensus among researchers regarding the underlying cause of these individual variations (see Miyake, 2001; Miyake & Shah, 1999; also see Conway et al., 2007). Until recently, the theoretical accounts of the reading span task were dominated by the notion of ‘resource sharing’ (Saito & Miyake, 2004). The model was originally proposed by Daneman and Carpenter (1980), who emphasized a tradeoff between the processing and storage demands within the single pool of cognitive resources. According to this resource-sharing model, a WM span task is a measure of the functional capacity of cognitive resources that can be flexibly allocated between the processing and storage activities of WM (see Saito & Miyake, 2004). For example, for an individual skilled at language processing, fulfilling the concurrent processing requirements of the reading span task (i.e. reading aloud or sentence verification) will consume a relatively small amount of resources, thus enabling him or her to allocate a relatively large amount of the leftover resources for maintenance of the target memory items. In contrast, if a person is not skilled at language processing, then performing a concurrent language processing task will consume a lot of resources, leaving only a small amount available to support the storage of target words, hence leading to a low span score. As this example illustrates, the resource-sharing model argues that an individual’s performance on WM span tasks reflects the amount of resources available to the individual after the processing requirements of the tasks are met. Although there are a number of variants on this resource-sharing view,
32 Part 1: Theoretical and Methodological Foundations
they all incorporate the notion of tradeoffs between processing and storage (Miyake, 2001). Although there have been some slight differences of opinion on this view in recent years, it is undeniable that this resource-sharing model has been dominant since its original conception and that it is still the most widely held among cognitive psychologists (see Miyake, 2001). It should also be noted that although a number of recent propositions seem to question or even challenge the validity of this view, they are actually not that distant. A particular case in point is the so-called ‘controlled attention’ view of Engle (2001, 2002). This view proposes that WM capacity as measured by WM span tasks is jointly determined by short-term storage plus what is termed ‘controlled attention’ ability, which is characterized as a domaingeneral, limited attentional capacity for performing controlled processing or sustaining focus on task-relevant information in the face of interfering or distracting stimuli (inhibiting irrelevant information). As discussed in Chapter 2, this seemingly conflicting characterization is consistent with the unified theories of WM outlined in which WM is conceptualized as a primary memory system comprising multiple components and multiple functions that encompass both executive processes and non-executive operations. That is, this view integrates the different research focuses (on the phonological or executive characteristics) rather than fundamental differences in the theoretical stance adopted by the two WM research traditions (the European versus North American traditions). By many measures, Engle’s proposition is analogous to the view that WM capacity equals the capacity of the slave system (e.g. the phonological loop) plus the efficiency of central executive functioning (Miyake, 2001). In other words, Engle’s conception of the attentional component is very much isomorphic to Baddeley’s (1986) notion of the ‘central executive’, that is, an attentional system involved in regulating information processing (see also Williams & Lovatt, 2003). Therefore, it is reasonable to adopt either version of these complex span tasks to assess the executive component of WM, although they may reflect different perspectives, with the reading span reflecting the domain-specificity view and the operation span reflecting the domain-generality view.
Domain-specificity versus domain-generality The second issue that needs to be addressed is whether WM is a general capacity system independent of task content (Turner & Engle, 1989). The foregoing discussion points to the possibility that slightly different constructs underlie the language-related (such as the reading span task and its variants; represented by Daneman and colleagues) and language-unrelated measures (such as the operation span task, as represented by Engle and colleagues) used by researchers in the field of cognitive psychology.
Working Memory Measures and Issues 33
First, because they are more language-related than the measure of the operation span tasks, domain-specific WM measures such as the reading span task (or its variants such as the speaking span task) are likely to be more sensitive to differences in the processes of language comprehension and production. Second, domain-general WM measures such as the operation span task involve fewer language components and are more likely to reveal the general capacity or ‘central executive capacity’ (Rosen & Engle, 1998: 418). Moreover, Engle and Oransky (1999: 542) emphasized that the differences in WM capacity are really individual differences in a single component of the WM system, namely controlled attention. That is, certain factors are likely to bias the results of a study in one direction or the other and thus researchers should take caution when designing or adapting these complex memory span tasks (Miyake, 2001). If this discussion holds, then the debate between the unitary functional view (such as that held by Cowan and Engle) and the multiple structural view (endorsed by Baddeley and colleagues) of the WM system has a bearing on the subtle differences in the outcomes of the measures of WM. However, it is also possible that rather than being mutually exclusive, these measures reflect ‘a common mechanism of information storage and processing’ (Rosen & Engle, 1997: 211). As clearly stated in the unified theories of WM discussed in Chapter 2, despite the many differences between these two interpretations, the emerging consensus (Miyake & Shah, 1999; Baddeley & Hitch, 2001) in the field is relatively clear; that is, ‘WM is not completely unitary, and any comprehensive theory of WM should be able to account for some domain-specific effects reported in the literature’ (Miyake & Shah, 1999: 448). In a similar vein, Engle et al. (1999b) proposed what is known as ‘the hierarchy view’ (also see Miyake, 2001), which incorporates the dual nature of WM by arguing that a variety of verbal and non-verbal WM span tasks are best explained by a model comprising domain-specific storage capacities plus a domain-general attentional resource (Williams, 2012). This view is echoed in the SLA field by Dekeyser and Juffs (2005), who cogently argued that if language learning is considered to be no different from other forms of human cognition, then complex span measures that reflect a general capacity (such as the operation span task) should be adopted (for a similar argument see Sanchez et al., 2010). In contrast, the complex span tasks that reflect domain-specificity should be adopted if language learning is considered to consist of specific processes or skills, such as listening, reading, writing and speaking. In other words, both versions of the complex memory span tasks can be adopted, although there may be hierarchical differences in the outcomes as a function of a number of linguistic and non-linguistic factors (e.g. Cai & Dong, 2012; Linck et al., 2014).
34 Part 1: Theoretical and Methodological Foundations
Methodological Issues Besetting WM Measures Besides the theoretical disputes over the WM measures, another issue related to the increasing popularity of the WM span tasks concerns their methodologies. In particular, although the methods with which the WM span tasks are administered vary widely, the ramifications of these different methods were not specified until very recently (Conway et al., 2005; Friedman & Miyake, 2004; Lewandowsky et al., 2010; Nutley et al., 2010). Fortunately, over the past few years, concerted efforts have been made in the cognitive psychology field to rectify this situation (for more details see Conway et al., 2005; Linck et al., 2014). In light of these encouraging developments, in this section I outline and clarify several of the methodological issues relating to the WM span measures, in particular those pertaining to the reading span task and the operation span task.
Reliability issues Despite the widespread use of the WM tasks, some of their basic psychometric properties remain inadequately characterized (Waters & Caplan, 2003). These include: (1) their internal consistency, that is, the homogeneity of the measures; (2) their test–retest reliability, that is, the extent to which the scores obtained in one testing session correlate with those obtained in another; and (3) the stability of the categorization of the participants, that is, whether the classification of the participants into discrete WM groups (e.g. high, medium or low) is stable over time and/or different WM tasks (also see Klein & Fiss, 1999). To address these issues, Waters and Caplan (2003) conducted a formal meta-analysis of the results of many of the previous studies on the psychometric properties of WM tasks. They found that the estimates of internal consistency (0.67–0.95, as measured by the split-half reliability) tended to be higher than the estimates of test–retest reliability (0.41–0.83), and that the stability of the participant classification into WM groups was poor in many studies. The authors continued to investigate seven frequently used measures of WM capacity (including the reading span task) in an elderly population, with a view to further characterize the basic psychometric properties of these tests, their correlations across the tests and their factor structures in the population. In summary, there are three main findings of the study by Waters and Caplan (2003). First, the results are consistent with many others in the literature in showing that older individuals perform worse than younger individuals in many WM tasks. Second, many of the WM tests applied to elderly participants have acceptable psychometric properties, although this needs to be qualified by the observation that test–retest reliability is
Working Memory Measures and Issues 35
acceptable only when performance is measured over the span size or items, not when the participants are classified into WM groups on the basis of either relative performance or cutoff scores. Third, the performances on different WM tests are only moderately correlated. Importantly, the authors recommended that a composite measure that combines the participants’ storage, recall accuracy and reaction time may capture commonalities among the WM measures and be more likely to result in a more reliable and stable characterization of the participants’ WM performance.
Validity issues Because WM span tasks are complex and impose considerable attention and memory demands on participants, it is perhaps inevitable that participants develop idiosyncratic strategies for balancing the processing and storage components of the WM measures (Friedman & Miyake, 2004). The types of strategies vary widely across individuals, and raise concerns that the nature of the measure may change according to the extent to which the participants are allowed to develop and use idiosyncratic strategies. Fortunately, a number of studies have examined how such strategies may influence the measures of WM capacity, and the results should shed light on the implementation of the WM measures. Engle et al. (1992) used the reading and operation span tasks and found that viewing times were longest for the first components of the sentences or equations (i.e. the sentence order effect), which they interpreted as being consistent with the rehearsal of the words to be recalled. When they controlled for individual differences in the viewing times for these components, the correlation between the operation span scores and participants’ verbal scores on the Scholastic Aptitude Test (SAT) varied from 0.36 to 0.41, although the increase was not significant in their sample of 70 participants (the correlation with reading span did not change when the viewing times were partialled out). They interpreted these non-significant results as evidence that strategy use did not mediate the relationship between operation span and comprehension. However, the observation that the correlation with operation span increased when the strategy times were partialled out suggests that allowing participants extra time for strategies may have reduced the criterion validity of the measure. Moreover, when Engle et al. (1992) allowed the participants unlimited rehearsal time in a word span task, the task no longer significantly predicted their SAT verbal scores. McNamara and Scott (2001) also examined the effect of strategy training on STM and WM performance. The participants were trained to use a chaining strategy in conjunction with a STM task. The participants read 15 words and were trained to create a story using these words. The WM span scores obtained using the reading span measure improved following
36 Part 1: Theoretical and Methodological Foundations
the training on the STM task. These researchers also found that participants who were more strategic before training displayed better WM performance. The authors attributed the changes in WM span scores to experience and learning, that is, ‘the knowledge-is-power’ hypothesis, rather than changes in the allocation of WM resources as suggested by the strategic allocation hypothesis. Turley-Ames and Whitfield (2003) conducted a series of strategy training experiments and found that when participants were trained to use a particular strategy to perform the operation span task, both the operation span scores and the correlations between operation span and reading comprehension ability increased. In Study 1, the participants’ WM span scores increased as a result of using a rehearsal strategy (i.e. saying aloud the words to be remembered). In Study 2, the correlation between the participants’ WM span and reading comprehension scores was enhanced as a result of using strategies, that is, low-span participants seemed to benefit the most from using the rehearsal strategy. In Study 3, when the time allowed for using the strategy was controlled, the WM span scores obtained by the participants while using the rehearsal strategy were, again, more predictive of their reading ability. The authors thus interpreted these results as supporting the hypothesis that variations in strategy use (idiosyncratic strategies) suppress the relationship between WM span and complex cognition. This is equivalent to stating that strategy use by participants when performing WM span tasks should be prohibited or controlled, as it may mask the ‘true’ correlations between the span scores and comprehension scores (for a similar argument see St Clair-Thompson, 2007). Friedman and Miyake (2004) tested the effects of the administration method on the criterion validity of the reading span task, and examined the relationship between processing and storage in this task. With respect to the first goal, although the experimenter- and participant-administered reading span tasks were equally reliable and induced similar types of strategies, the extra time taken to implement strategies in the participant-administered version reduced the correlations with the reading comprehension and SAT verbal scores. With respect to the second goal, although the sentence processing times were related to storage ability, they did not mediate the relationship between the reading span scores and comprehension measures, thus leading the authors to conclude that the theories focusing on processing as an explanation of what the reading span task really measures were incomplete. The overall patterns of these two results confirmed the concerns in the field, namely that the methodology used for administering WM span measures has important theoretical ramifications. Most relevantly, Friedman and Miyake (2004) cogently proposed three practical suggestions to eliminate (or at least to minimize) the effect of this validity threat. First, they proposed not allowing the participants to control the onset of each new stimulus or allowing them any time beyond
Working Memory Measures and Issues 37
that needed to process the stimulus. Second, they recommended obtaining measures of the processing time for the span task(s) that can help the experimenter evaluate whether the participants are implementing idio syncratic strategies. Third, they suggested incorporating a well established measure of criterion validity (such as reading comprehension for the reading span and operation span tasks) in the research design. Thus far, the most comprehensive and authoritative guidelines for WM measures in cognitive psychology have been provided by Conway et al. (2005), who suggested using vigorous measures to ensure the reliability and validity of the WM span tasks. For example, concerning the test materials used in WM measures, the authors argued that the words that serve as recall items should be common nouns, as abstract words have been found to increase the difficulty of word recall (Turner & Engle, 1989). In addition, none of the recall items in the same set should be semantically associated with each other, which can be triangulated and validated by pre-tests. That is, the stimuli for WM measures should be strictly controlled to ensure comparable levels of difficulty. Then, the instruction (rubric) of the task should be crystal clear (Miyake, 2001), so that the participants understand what they are supposed to do, and it is preferable to include a practice section before the real test begins. Last but not least, WM tests can be administered through more scientifically controlled means, preferably with the help of computer programs (such as E-prime) (MacWhinney, 2001; Schneider et al., 2002), to achieve higher validity and thus minimize the opportunities for participants to manipulate their idiosyncratic strategies. No doubt, these proposals provide invaluable resources and guidelines for implementing WM measures in practical areas.
Summary Thus far, it should be clear that the field of cognitive psychology does not just offer theoretical conceptualizations of the WM construct (Chapter 2), but also provides methodological insights with established assessment procedures for measuring the phonological and executive characteristics of the WM construct. To sum up, most of the currently available WM span tasks in cognitive psychology can be classified according to the complexity of the task demands (simple versus complex) and the nature of the span tasks (domain-specificity versus domain-generality) (as listed in Linck et al., 2014). An additional advantage for these WM span measures is that both simple and complex versions of the WM span tasks are not only available in print form, but also have readily and accessible computerized or automated versions (e.g. Alloway, 2007; Alloway et al., 2008; Unsworth et al., 2005). Another recent development is the equally reliable shortened version of the
38 Part 1: Theoretical and Methodological Foundations
operation span task (e.g. Foster et al., 2015; Swanson, 2015), which can be easily applied by researchers in more practical areas. Based on the reviews presented in this chapter, two other broad generalizations can be drawn regarding the two types of WM span tasks. First, with respect to the simple version of the memory span tasks (e.g. the non-word repetition span task), a considerable body of research (particularly that conducted by the European cognitive psychologists) suggests that they mainly tap into the two cognitive mechanisms associated with PWM, that is, the phonological short-term store and the articulatory rehearsal mechanism (Gathercole, 2006). Therefore, they provide reliable and valid means for measuring PWM. Second, with respect to the complex memory span tasks (e.g. the reading span task and its variants, and the operation span task), numerous studies (especially those conducted by the North American cognitive psychologists) have amply demonstrated that they tax the executive aspects of WM, such as information updating, task switching and inhibition. Therefore, these tasks could provide appropriate measures for assessing EWM. Indeed, previous analyses (e.g. Klein & Fiss, 1999; Waters & Caplan, 2003) have indicated that these complex memory span tasks have proven reliability and validity standards, provided they are administered cautiously (Conway et al., 2005; Lewandowsky et al., 2010; Sanchez et al., 2010). Despite these broad positive generalizations, numerous other theoretical and methodological issues regarding the WM measures and their assessment procedures need to be clarified in future research. One issue relates to the domain-specificity of the tasks versus their domain-generality. As pointed out by Dekeyser and Juffs (2005), the currently available WM measures reflect both the domain- and task-specificity (e.g. the reading span and speaking span tasks) and domain-generality (e.g. the operation span task). However, it appears that the current practice in cognitive psychology is that the use of either domain-general span tasks or domain-specific span tasks is largely a matter of choice and dependent on the researchers’ epistemological stance. In other words, there is a lack of a standard procedure for selecting or implementing WM measures and scoring the WM span tasks. Accordingly, it is likely that researchers who use these measures in more applied or practical areas (such as education and language learning) will encounter great challenges (e.g. Fenesi et al., 2015). Therefore, future research needs to further specify the consequences of using the two different types of WM measures (domain-specific versus domain-general). Most importantly, in line with the unified theories of WM (as outlined in Chapter 2), which distinguish between the components and functions of WM, future research on WM measures can take this distinction into account in two ways. First, instead of adopting the structural view (such as the non-word repetition span task, which taxes PWM, or the reading span task, which taps into EWM), the measures of WM can be functionally oriented
Working Memory Measures and Issues 39
by targeting specific executive functions. For example, the ‘running memory’ task paradigm measures the executive function of information updating of WM (e.g. Broadway & Engle, 2010; Bunting et al., 2006; Salthouse, 2014). Second, one implication of the integrated components/functions view is the adoption of a ‘WM profile’ that displays the results of all individual WM components or functions instead of one single component or function (Dehn, 2008; Dehn et al., 2015; Doughty et al., 2010; Politimou et al., 2015). Clearly, a perceivable advantage of adopting a multiple-span profile (rather than a single WM score) is that there may be strengths in some components/functions but weaknesses in others for individual participants (making the remedial interventions more effective) (Politimou et al., 2015). Together, these two research directions should have significant theoretical and methodo logical implications for future WM test construction and development. To end the chapter, the two existing formats of the WM span tasks provide readily applicable WM measures and assessment procedures that are central to pursuing the associations between WM and cognition. To a large extent, as Dehn (2008: 58) cogently put it, ‘[until] research and measurement tools allow us to further delineate WM processes, it might be safest to define WM as what simple and complex WM span tasks measure’. Given that the inherent controversies and debates relating to the theories and measurement of WM are likely to continue, these well established WM assessment measures are likely to continue to play an important role in the foreseeable future. Nonetheless, there is also much room to develop WM measures that go beyond these currently available formats by, for example, adopting a ‘profile’ perspective, with procedures and techniques that target putatively specific cognitive functions. Having outlined the concept of the WM construct in Chapter 2 and its assessment procedures in this chapter, the next step is to examine how WM has been conceptualized and operationalized in practical research on language learning and SLA, which is the focus of this book. Thus, in the next chapter, I endeavour to synthesize the research on the relationship between WM and first language acquisition and processing.
Part 2 Research Syntheses of Working Memory in L1 and L2 Learning
4 Working Memory in First Language Research
Concomitant with the growing research interest in WM in general cognitive psychology, a number of streams of research have investigated the specific relationship between WM and the learning and processing of a first language (L1). This should come as no surprise, given that language is arguably one of the most complex and intriguing elements of higher-level human cognition (Andrade, 2014; Gathercole, 2007; Jackendoff, 2007a, 2012; Van Dyke, 2012). These strands of WM–language research have gradually developed into two rather well defined research paradigms that are respectively influenced by the two distinctive theoretical research camps of WM – the European tradition versus the North American tradition as discussed in Chapter 2 (Andrade, 2001, 2014). As the discussion has unfolded, it has become clear that the European camp has mostly focused on the role of phonological WM in vocabulary acquisition and development, whereas the North American camp has explored the implications of the executive elements of WM for language processing, in particular, reading comprehension. In light of this broad distinction, in this chapter I provide a selective review of these well established strands of European and North American WM–L1 research, while excluding numerous other studies that have been conducted by scholars elsewhere. More specifically, the first section summarizes the WM and L1 research conducted by the European strand that focuses on the role of PWM in vocabulary acquisition and grammar development. The second and third sections examine studies on the relationship between EWM and L1 comprehension and production conducted by North American-based WM researchers. The following section discusses how the concept of WM has been incorporated into the theoretical frameworks of some of the general linguistic theories and sentence processing models. The fifth section summarizes these strands of WM–L1 research and concludes the chapter. 43
44 Part 2: Research Syntheses of Working Memory in L1 and L2 Learning
WM in L1 Acquisition and Vocabulary Development The European tradition of WM–language research was largely spearheaded by the pioneering empirical work conducted by the British cognitive and developmental psychologists Alan Baddeley and Susan Gathercole at the Cambridge University MRC Applied Psychology Unit. From the outset, their WM–language endeavours were guided by the early tripartite model of WM proposed by Baddeley and Hitch (1974; Baddeley, 1986). As discussed in Chapter 2, a key feature of this conception of WM is the modality-based demarcation of short-term storage functions, which usually subsume a sound-based phonological loop and a visuospatial sketchpad. The activities that take place in these two putative components are regulated and co ordinated by a third supervisory component, the central executive. Later, a fourth storage component, the episodic buffer, was divided from the central executive in the original model to integrate information coming from any possible sources or modalities into episodes that connect with long-term memory (Baddeley, 2000b; Baddeley et al., 2010). In line with this multi-component perspective on WM, most of the language-related research conducted by the British WM camp has aimed to demystify the possible connections between individual WM components and their relationships with the language learning domains and activities (Baddeley, 2000a, 2003a, 2015; Gathercole & Baddeley, 1993). Detailed reviews of these strands of WM–language research can be found in the works compiled by the two representative scholars (Baddeley, 2000a, 2003a; Gathercole, 2007; Gathercole & Baddeley, 1993). Gathercole and Baddeley (1993), for example, provide a comprehensive review of the detailed implications of the three key components of WM in the original tripartite model (with the phonological loop and the central executive being highlighted) for five central areas of the L1 learning sub-domains and activities (Table 4.1). As shown in Table 4.1, with respect to reading familiar words, the effects of both WM components are marginal or mostly absent, except when complex judgements about the phonological structure are required, in which case the phonological loop is drawn on. However, when it comes to vocabulary acquisition, both WM components are invovled, although differently: the phonological loop plays a critical role in coding and consolidating the phonological forms of new words into the long-term knowledge base (of the vocabulary item), while the central executive is involved in coordinating attentional resources for interpreting the semantic characteristics of these newly learned items. Indeed, most of the WM–language research conducted by the European camp has focused on the relationship between the phonological loop and vocabulary acquisition and development, whereas fewer empirical studies have examined the effects of other WM components (e.g. the visuospatial sketchpad, the central executive and the episodic buffer)
Working Memory in First Language Research 45
Table 4.1 WM components and L1 sub-domains and activities Language activities Phonological loop (PWM)
Central executive (EWM)
Reading familiar words
No involvement, except when complex judgements about the phonological structure are required
Not involved (as of 1993)
Vocabulary acquisition
Critical for long-term learning of the phonological form of new words
Involved in allocating attention to interpret the semantic characteristics of new words?
Syntax/grammar acquisition
Can predict native grammatical ability (including that of artificial grammar)
Not clear (as of 1993)
Learning to read
Contributes to the development of a phonological recoding strategy
Not clear (as of 1993)
Language comprehension
Used to maintain a phonological record that can be consulted during offline language processing
Involved in processing syntactic and semantic information and storing the resultant products for subsequent processing
Speech production
None (as of 1993)
Involved in planning the conceptual content of speech?
on L1 domains and activities. Clearly, the North American WM camp has explored a rather different research direction (as discussed in the next section). It is apparent that of the two putative WM components, the phonological loop, as conceived in the multi-component model, has been the most extensively studied in the European line of WM–language research (Baddeley, 2015). The phonological component of WM is further postulated to subsume two associated mechanisms: a passive phonologically based store and an active articulatory albeit sub-vocal rehearsal mechanism (Baddeley, 1986; Baddeley & Hitch, 1974). In terms of measurement, most of these early studies used storage-based recall tests (i.e. the simple memory span task reviewed in Chapter 3), such as the digit span task, the word span task and, more recently, the non-word repetition span task (Gathercole, 2006; Gathercole et al., 1994). Given its unique sound-related feature, the phonological loop became a major focus of most of the early WM–language research conducted by the European camp. Overall, the European cognitive and developmental psychologists made concerted efforts to reveal the possible implications of this particular WM component for essential language learning sub-domains such as vocabulary acquisition and grammar development, and to examine the role that the component plays in other language learning activities and skills (Baddeley, 2000a, 2003a).
46 Part 2: Research Syntheses of Working Memory in L1 and L2 Learning
The highlights of this European strand of WM–language research are the empirical studies contributing to the theoretical link between the phonological loop and vocabulary learning, particularly among young children (Gathercole & Baddeley, 1989, 1990, 1993). For example, Gathercole and Baddeley (1989) found that four-year-old children’s non-word repetition span (purportedly indexing the phonological loop) could accurately predict the vocabulary size of their native language one year later (even after statistically controlling for their prior vocabulary knowledge). In another training study, Gathercole and Baddeley (1990) demonstrated that children’s ability to repeat non-words was in direct proportion to their efficiency in learning new vocabulary items. Moreover, Gathercole et al. (1999) found moderately strong correlations between repeat non-words among children across age groups ranging from 4 to 13 years. In other words, the close link between the phonological loop and vocabulary acquisition seems to persist throughout childhood (Gathercole et al., 2004). Furthermore, other studies have indicated that the link between phonological WM and vocabulary learning among children is not uni-directional. Gathercole et al. (1997) investigated the intricate links between new word learning, the phonological loop and existing lexical knowledge among fiveyear-olds. They found that learning new words was influenced not only by the phonological loop, but also by children’s existing vocabulary knowledge, thus indicating a reciprocal relationship between the two. As the authors cogently pointed out, ‘with greater amounts of lexical knowledge, the degree to which knowledge about the structure of the language can be used to supplement and enhance fragile traces of new words in the phonological loop will also be increased’ (Gathercole et al., 1997: 977). A possible way to separate this complicated relationship is to look for ‘cross-lagged’ correlations, that is, correlations between relevant variables measured at two successive points (as in Baddeley, 2000a). This approach clearly indicates the interaction between the phonological loop (e.g. as measured by the non-word repetition span task) and vocabulary learning (Gathercole et al., 1992). In other words, not only does capacity of the phonological loop influence rate of vocabulary acquisition, but conversely a richer vocabulary is associated with increased verbal memory capacity, probably because the richer substrate of language habits allows elaborate and effective coding within the phonological loop. (Gathercole, 1995, as cited in Baddeley, 2015: 24) Notwithstanding the extensive research on the relationship between WM and vocabulary acquisition, considerably fewer studies have examined the role of WM in relation to syntax/grammar (Ellis & Sinclair, 1996: 236). Cognitive psychologists are more divided on this relationship, depending on their epistemological perspective (e.g. what is ‘grammar’?). On the one
Working Memory in First Language Research 47
hand, Adams and Willis (2001) suggested that WM might not be related to grammar development in L1. On the other hand, Ellis and Sinclair (1996) demonstrated the involvement of the phonological loop in the acquisition of a specific grammatical form within Welsh, and more recently Andrade and Baddeley (2011) corroborated its positive role in the acquisition of an artificial grammar. Even fewer European studies have explored the implications of the other components of WM (i.e. the central executive, the visuospatial sketchpad and the episodic buffer) for L1 learning tasks (Baddeley, 2015). Notably, most of these studies (especially those exploring the relationship between the central executive and language comprehension) were conducted in the American tradition (to be discussed in the next section). Also, the interactions between the putative WM components (such as the phonological loop and the central executive) remain unclear, although a training study conducted by Gathercole et al. (2006) indicated that a major problem in acquiring both vocabulary and reading skills occurs when a child has poor capacity with respect to both the loop and the central executive. Overall, this body of European research into WM–L1 acquisition demonstrated clearly that the phonological loop of the WM construct (i.e. phonological short-term memory), given its two associated mechanisms (the phonological short-store and the articulatory rehearsal mechanism), plays an instrumental role in the acquisition and development of vocabulary and possibly syntax/grammar (Verhagen & Leseman, 2016). According to recent advances in connectionism and corpus linguistics (e.g. Gries & Ellis, 2015), vocabulary and grammar are intermingled and inseparable in that both vocabulary (i.e. lexis) and grammatical structures (i.e. morphosyntactic constructions) can be considered linguistic sequences or chunks of different lengths (Bates & Goodman, 1997; Marchman & Bates, 1994; also see Römer, 2009). The rationale for this argument is that WM, in particular phonological WM (PWM), with its two cognitive mechanisms of phonological short-term store and articulatory rehearsal (albeit sub-vocally), allows new (incoming) sound sequences to be temporarily held in storage, subsequently encoded and consolidated into long-term representations, and thus ultimately to become long-term lexical knowledge (N.C. Ellis, 1996, 2002, 2012; Martin & Ellis, 2012; Verhagen & Leseman, 2016). To sum up this section, there appears to be considerable evidence for the importance of the phonological loop in first language acquisition, particularly so in the acquisition and development of vocabulary. This postulated link between the phonological component of WM (PWM) and language acquisition is clearly reflected in the later model of WM by Baddeley (2000a, 2000b, 2003a), which postulates that the evolutionary purpose of the phonological loop is to serve as a ‘language learning device’ (Baddeley et al., 1998: 158). Thus, it is safe to conclude that phonological memory plays an indispensable role in the acquisition and development of vocabulary
48 Part 2: Research Syntheses of Working Memory in L1 and L2 Learning
and grammar/syntax (Martin & Ellis, 2012; Verhagen & Leseman, 2016). Despite this claim, however, Baddeley (2015) also added that this does not necessarily demonstrate that phonological memory is essential rather than useful, as a graduate student (called SR) in an earlier case study (Baddeley, 1993b) seemed to have a good vocabulary despite developmental weakness in STM. In that sense, it is fair to claim that PWM is a very helpful but not essential tool in language acquisition (Baddeley, 2015: 24).
WM in L1 Listening and Reading Comprehension Many North American cognitive scientists have been motivated by Baddeley’s pioneering work on WM and become interested in the relationship between WM and language learning, for example Meredyth Daneman, Patricia Carpenter, Marcel Just, Gloria Waters and David Caplan. Notwithstanding, as discussed in Chapter 2, most of the North America-based WM researchers have distinguished themselves from Baddeley’s multi-component view of WM and opted to regard WM as the focus of attention, embedded within activated long-term memory (Cowan, 1999) or ‘controlled attention’ (Engle, 2002; Engle et al., 1999a). Many of the North American researchers who embraced the attentionand control-oriented perspectives on WM (similar to the central executive as conceived in Baddeley’s multi-component model) have subsequently focused more on investigating individual differences in these executive elements of WM and their effects on language processing, including language comprehension (e.g. listening and reading processes) and language production (e.g. speaking and writing processes). Over the years, a number of theoretical models of the relationship between WM and language comprehension have been proposed (for reviews, see Caplan & Waters, 2013; Cowan, 2013; Long et al., 2006; Van Dyke, 2012). In this section, I summarize some of these theoretical models and their research paradigms. The classical and most influential WM–language comprehension research paradigm in the North American tradition was initiated by Daneman, Carpenter, Just and colleagues (e.g. Daneman & Carpenter, 1980; Just & Carpenter, 1992; King & Just, 1991; MacDonald & Christiansen, 2002; cf. Williams, 2015), who postulated that WM was a single pool of cognitive resources implicated in the process of language comprehension. In terms of the measures of WM, these researchers have abandoned the European tradition of adopting the simple storage-based memory span task (such as the word span or the digit span task). Instead, they have constructed a more complex storage-plus-processing memory span task that involves sentence judgement and recall of the sentence final word. The basic assumption of this dual-task research paradigm is the postulation of competition (or
Working Memory in First Language Research 49
tradeoff) between the two essential cognitive functions associated with WM, namely the simultaneous processing of a sentence (e.g. evaluating its grammatical plausibility) and the storage of sentence final words. As discussed in Chapter 3, these dual-task assessment procedures include the most widely adopted benchmark version of the reading span task (RST; Daneman & Carpenter, 1980) and its variants of the listening span task (Daneman & Carpenter, 1986) and speaking span task (Daneman, 1991). The RST has almost become synonymous with the WM test (sometimes even as a WM model) in the North American tradition, as opposed to the simple memory span task of European WM scholars, and it is the most widely used task in language acquisition and sentence processing research (e.g. see Juffs, 2006, 2015; Juffs & Rodriguez, 2014; Rodriguez, 2008). Indeed, a meta-analysis of 77 studies conducted by Daneman and Merikle (1996) (involving 6179 participants) suggested that this type of storageplus-processing WM reading span task could outperform the storage-only versions of the WM measure adopted in the European tradition (e.g. the digit span and non-word repetition span) in predicting the participants’ reading comprehension scores. Furthermore, numerous studies have concurred that WM (as measured by these complex memory span tasks) plays a pivotal role in a wide range of specific language comprehension processes and activities, including resolving linguistic ambiguities, parsing syntactically complex structures, generating online inferences, comprehending and producing words in context, and processing relative clauses of differing levels of complexity (Miyake & Friedman, 1998: 343). Therefore, it is reasonable to suggest that to date, this version of the WM span task continues to enjoy widespread popularity both within and outside the fields of psychology and language research (Van Dyke, 2012). However, despite its far-reaching influence, Daneman and Carpenter’s version of the WM span task has been challenged by other cognitive psychologists within the North American camp. For example, Turner and Engle (1989) argued that to predict reading comprehension, the sentence processing part of the WM span task does not necessarily have to be reading-related and can be arithmetic-based. Extending this assumption, they devised the operation word span task (Ospan), which is based on some simple mathematical computations and final-word recall; Unsworth et al. (2005) provide a computerized version. In a series of experiments (e.g. Engle et al., 1992), Engle and colleagues demonstrated empirical evidence for the superior power of their domain-general version of the operation span task in predicting participants’ reading comprehension performance, thus lending support to their general capacity view of the WM–language comprehension relationship. Indeed, in several other methodological and comparative studies on a wide range of WM span tasks, the domain-general operation span task usually enjoyed a higher level of reliability and validity (e.g. Waters & Caplan, 2003; also see Conway et al., 2005).
50 Part 2: Research Syntheses of Working Memory in L1 and L2 Learning
Daneman and colleagues’ interpretation of the WM–language comprehension relationship has also been criticized by Gloria Waters, David Caplan and colleagues on two other fronts (e.g. Caplan & Waters, 1999, 2013; Waters & Caplan, 1996). The first criticism gives rise to further refinements of the assessment and scoring procedures of the reading span task (RST). Waters and Caplan argued that it was incomplete and even fallacious to represent WM capacity simply by a single recall score of the sentence final words on the grounds that it did not adequately tap into the major processing function of WM. To rectify this, they suggested that the participants’ WM capacity should be represented by a standardized ‘composite score’ that takes into full account (1) the recall span, (2) the reaction times and (3) the sentence judgement accuracy data (also see Conway et al., 2005; Waters & Caplan, 2003). Another criticism from Waters and Caplan targeted the purported ‘single-resource’ view of the relationship between WM and language comprehension as depicted by Daneman and colleagues. To that effect, Caplan and Waters (1999, 2013) argued for a two-stage account of the involvement of WM in language comprehension. More specifically, they postulated that WM is not implicated in the first stage, which is normally characterized by online and automatic processing, and proposed that its resources are likely to be drawn upon in the second stage, by certain offline ‘post-interpretive’ processes (such as anaphor resolution and pronoun detection). The results of a series of studies conducted by the group (e.g. Caplan et al., 2007; Caplan & Waters, 2013) provided compelling evidence for these separate and specialized WM resources (for syntactic processing). The general findings suggest that the syntactic analysis of people with a smaller WM is not impaired compared with people with a larger WM, and that syntactic memory load and other types of memory load result in different patterns of activation in brain imaging studies (also see Cowan, 2013). Caplan and Waters (2013) further expanded their WM–language comprehension model into an integrated short-term WM (ST-WM; i.e. Baddeley’s conception) and long-term WM (LT-WM; i.e. Ericsson & Kintsch’s (1995) conception). Based on a comparative review of the distinctive contributions of these two types of WM models (ST-WM and LT-WM), Caplan and Waters described the limitations of Baddeley’s ST-WM model, and embraced the LT-WM model as a more viable theoretical framework for future research on assigning syntactic structure (parsing) and sentencemeaning interpretation. From a broader perspective, Waters and Caplan’s assumption of a domain-specific ‘separate syntactic WM’ for holding the type of grammatical information required for sentence formation is largely in line with the tenets of the ‘language acquisition device’ (i.e. universal grammar) in Noam Chomsky’s generative theoretical framework (Cowan, 2013).
Working Memory in First Language Research 51
WM in L1 Speech and Written Production Compared with the enormous number of studies that have explored the relationship between WM and language comprehension, relatively little research has investigated the connection between WM and language production (Acheson & McDonald, 2009). In the specific domain of language production, Levelt’s (1989, 1999) speech production model is probably the most notable and well established in the fields of psycholinguistics and applied linguistics (R. Ellis, 2005; Skehan, 1998). In this model, speech production subsumes three processing stages: a conceptualization stage, which gives rise to a pre-verbal message; a formulation stage, in which the lemmas or lexical items are selected (lexical encoding) and the syntactic structures are planned (grammatical encoding); and, finally, an articulation stage, in which the resultant product of speech is produced. Levelt (1983) also proposed three regulatory self-monitoring loops to accompany both the processes and the products of these three stages (also see Gilabert, 2007; Kormos, 2000a, 2006). However, with respect to the effects of WM, Levelt postulated that only the first stage, of conceptualization, needed to draw on WM resources to some extent, while the remaining two stages (formulation and articulation) proceeded in a parallel processing manner that was ‘largely automatic’ (Levelt, 1989: 21) without resorting to WM. Given this marginal contribution of WM in speech production, relatively few studies have examined this topic. However, it should be noted that the postulation of a minimal role for WM in speech production has been challenged in a number of empirical studies (e.g. Ferreira & Pashler, 2002; Hartsuiker & Barkhuysen, 2006), which have demonstrated that even the formulation stage of L1 speech production can be constrained by WM effects. For instance, Hartsuiker and Barkhuysen (2006) clearly demonstrated that participants with a smaller WM (as measured by the speaking span task) are more susceptible to subject–verb agreement errors as opposed to their high-span counterparts (particularly under the secondary task condition). In this sense, the reduced role of WM in speech production may need to be further verified, with more empirical studies and finer-grained scoring procedures. Therefore, it is fair to state that the jury is still out with respect to the relationship between WM and L1 production. The final domain of language production to be discussed is writing, which is generally viewed as a complex activity that presumably involves many simultaneous sub-goals and interacting processes, and whose overall quality is therefore limited by the writer’s WM resources (Olive et al., 2008; Swanson & Berninger, 1996). In the cognitive research on writing, a modest exercise to explicitly align the effects of WM with the cognitive demands of the writing process was proposed by Torrance and Galbraith (2005).
52 Part 2: Research Syntheses of Working Memory in L1 and L2 Learning
Different models have also incorporated WM in the writing process (for a review, see Chanquoy & Denis, 2002), the most notable and influential being the one proposed by Kellogg (1996, 1999). Juxtaposed with Levelt’s (1989) speech production model, Kellogg postulated three sub-stages of written production: formulation, execution and monitoring. Each of these sub-systems in turn subsumes some basic processes within their own associated processes (see Kellogg, 1996: 59). More relevantly, Kellogg hypothesized that five of the six basic processes of writing place specific demands on the central executive of WM (i.e. EWM), the verbal or PWM and the visual/spatial component of WM (Kellogg, 1999: 46), with only the execution of programmed muscle movements proceeding fully automatically, without any demands on WM. Kellogg has corroborated these claims in a number of empirical studies by demonstrating that WM is significantly correlated with a number of writing measures, particularly those related to text generation (Kellogg, 2001, 2004; McCutchen, 2000; Olive, 2012; Olive et al., 2008; Swanson & Berninger, 1996).
WM in Linguistic Theories and Language Processing Models Another interesting development regarding the WM–L1 relationship is that a number of general theoretical accounts of language acquisition and sentence processing models have incorporated WM as a key mediating factor in their frameworks. The first is the so-called ‘emergentist account’ of language acquisition and processing proposed by William O’Grady (1997, 2005, 2012a, 2012b, 2013, 2015), which is anathema to the mainstream generative grammar (MGG) advocated by Chomsky, and to the usage-based approaches (Ambridge & Lieven, 2015; Tomasello, 2003). Chomsky (1965, 2005; also see Hauser et al., 2002) proposed that three key factors are crucial for the design and acquisition of (first) language: (1) the pre-wired or innate language faculty that is specific to the human species (i.e. the universal grammar); (2) specific language experience or exposure, which leads to minor variations among languages; and (3) certain general-domain principles or mechanisms (i.e. computational efficiency or memory limitations), which are not specific to the language faculty. Among these three factors, the majority of the generative grammar studies (following the Chomskyan tradition) have taken the first factor (the human-unique language faculty) as a primary factor for (first) language acquisition and have tasked linguists with revealing the universal elements that govern all human languages. From this perspective, the remaining two factors in language acquisition (i.e. language experience and computational efficiency) play only a secondary (peripheral) role in language acquisition.
Working Memory in First Language Research 53
For example, as Chomsky (1965: 3; as cited in Lu, 2011) unequivocally pointed out: Linguistic theory is concerned primarily with an ideal speaker-listener, in a completely homogeneous speech-community, who knows its language perfectly and is unaffected by such grammatically irrelevant conditions as memory limitations, distractions, shifts of attention and interest, and errors (random or characteristic) in applying his knowledge of the language in actual performance. In contrast to this view of MGG, in his ‘processing amelioration’ or ‘processing determinism’ hypothesis, O’Grady (2012b: 116) argued that it is not the first factor, but rather the third, namely, the non-grammatical, domain-general ‘processor’, that lies ‘at the heart of the human language faculty’ (O’Grady, 2012a: 496). Citing cross-linguistic examples from English, Korean and Russian, O’Grady illustrated how some core grammatical structures (e.g. the filler-gap dependency phenomenon between the two typologically distinct languages of Russian and English) emerge as a function of the operation of an efficiency-driven processor whose primary goal is simply to reduce the burden on our limited WM resources (O’Grady, 2013, 2015). In other words, the limited capacity of WM constitutes the ultimate constraint or processing pressures on universal human language processing. In a similar vein, Lu (2011) argued that the restriction of linguistic structure in WM should be considered the most basic and universal phenomenon of human language, and that it should be the starting point of linguistic structural analysis. Following Miller’s (1956) ‘magical number seven’, Lu speculated that any sequence of human speech at any moment should comprise no more than about seven chunks. Similarly, any syntactic construction, be it a phrase or a sentence, cannot contain more than about seven chunks, because all syntactic constructions are a sequence of speech. Take the sentence shown in Figure 4.1 as an example, which is composed
John Sub. NP
carefully Mann. AP
read Head Verb
a dictionary Object NP
in the library for hours Location Duration PP PP
Figure 4.1 WM and chunks in sentence parsing (Lu, 2011)
yesterday Time NP
54 Part 2: Research Syntheses of Working Memory in L1 and L2 Learning
of a head noun plus six phrases (not exceeding seven chunks). In this sense, human language is not only restricted by the span of WM, but also fully utilizes it in action. Moreover, Lu argued that WM actually underlies and restricts the chunking process of linguistic structures in most language families or typologies, thus crucially affecting language parsing and the linguistic structure (as demonstrated in Figure 4.2). Therefore, WM should be internalized as part of a language (processing) device for syntax – similar to the claim by Baddeley et al. (1998) for the ‘phonological loop’ as a ‘language learning device’. In this sense, ‘Grammar must code the universal and omnipresent cognitive constraint’, as opposed to Du Bois’s (1987) claim that ‘Grammars code best what speakers do most’. Another language processing model that incorporates the WM construct is the ‘parallel architecture’ framework proposed by Ray Jackendoff (2002, 2007a). Jackendoff ’s perspective on language processing aims to preserve all of the mental and biological characteristics of the MGG as advocated by Chomsky. However, it also differs from MGG by treating phonology, syntax and semantics as equally independent generative components, whose structures are linked by interface rules (Jackendoff, 2007a). More relevantly, Jackendoff argued that these three core grammatical structures (e.g. words and rules alike) are stored in the LTM but are integrated or assembled in WM during real-time processing (although these structures can be in competition). In other words, the limited capacity of human WM underpins and constrains both the processing and the construction or integration of various phonological, syntactic and semantic structures. In this sense, linguistic WM as conceived by Jackendoff (2007b) has three key variables, corresponding to the processing, construction and integration of these three core grammatical structures: (1) a phonological WM component for processing phonological structures; (2) a syntactic WM component for syntactic structures; (3) a semantic WM for semantic structures For a similar view for SLA, see Truscott (2015, 2016). Despite the prominent position that O’Grady and Jackendoff ascribe to WM in their theoretical linguistic models, they do not provide details for operationalizing and measuring WM and, as such (given the above-mentioned multiple conceptions and perplexing assessment procedures), many of their claims regarding the constraints of WM on language acquisition and processing remain at the speculative level and therefore cannot be falsified (for the time being). With respect to the scientific spirit of falsifiability, a third WM-based approach to language processing, the ‘dependency locality theory’ proposed
D L I M V V Japanese, Korean, Basque
V V
T D L I M
T L I M
T D M
T D
T
1:
2:
3:
4:
5:
6:
M I L D T English, Portuguese, Vietnamese, Yoruba
M I L D Latvian
M I L Hebrew
Figure 4.2 WM and word-order universals (Lu, 2011). Note: (T)ime location, (D)uration, spatial (L)ocation, (I)nstrument and (M)anner, plus the head (V)erb
V
V I L Tagalog
V D Chinese
T
Working Memory in First Language Research 55
56 Part 2: Research Syntheses of Working Memory in L1 and L2 Learning
by Edward Gibson (1998, 2000), may ‘warrant a revision’ of all of the previous theoretical accounts of language processing (Cowan, 2013: 92). In his model, Gibson postulates that our limited WM capacity constitutes the underlying cognitive resources required for the two essential procedures in sentence processing: structural integration (connecting words to the syntactic structures) and storage of the resultant structure. More relevantly and importantly, Gibson claims that the complexity of a sentence’s structural integration depends on the locality and distance of the new referents and the number of events that intervene between a head element and its dependent structure (Gibson, 2000: 102). It should now be clear that the most striking feature of this WM-based approach to linguistic processing is that its two major components can be quantified and, further, calculated. The ‘integration cost’ is indexed by the number of energy units ‘consumed’ in the comprehension process. The ‘storage cost’ can be calculated as the number of memory units that are maintained in WM. In this sense, the complexity level of reading and understanding a sentence becomes a mathematical problem that can be worked out (Cowan, 2013: 92). For instance, a simple sentence such as ‘The reporter disliked the editor’ involves only two memory units (one noun phrase and one verb phrase) that need to be stored for comprehending. However, a more complex sentence, such as ‘The reporter who the senator attacked admitted the error’, which contains relative clauses and has four memory units, will be much more costly to store and process. Indeed, the most compelling evidence for dependency locality theory comes from its prediction of the asymmetric processing of subject-extracted relative clauses (SRCs; e.g. ‘The senator who attacked the reporter admitted the error’) versus those that contain object-extracted clauses (ORCs; e.g. ‘The senator who the reporter attacked admitted the error’). Although a number of models have been proposed to explain why SRCs are usually easier to process than ORCs (Fedorenko et al., 2006, 2007), Gibson’s dependency locality theory is highly promising as a framework for explaining a wide range of syntactically complex phenomena, including resolving relative clause ambiguities and attachment preferences (Cowan, 2013; Kim & Christiansen, 2012). Nonetheless, there is still much room for further refinement of the theory. In this regard, more cross-linguistic evidence is needed to support the model, particularly from languages that are typologically different from English; for example, Gibson and Wu (2013) suggested that Chinese ORCs may be easier to process than SRCs. In addition, further research is needed on conceptualizing and operationalizing the WM construct (in terms of WM measures) and incorporating these into the model, for example with respect to the relationships between the components/mechanisms of WM and the integration cost or storage cost.
Working Memory in First Language Research 57
Summarizing the WM–L1 Association Overall, the research on the WM–L1 association is clearly substantial and diverse, and has rendered the task of synthesizing these factors extremely challenging. However, I have reviewed only the relevant studies conducted within the two well established European and North American research traditions. While acknowledging the narrow scope and possible bias of such a selective review (by ignoring the huge amount of similar research conducted elsewhere), I believe that these studies represent a reasonable departure point for capturing the complex relationship between WM and first language acquisition and are able to shed light on the major concern of this book, namely the role of WM in the learning and processing of a second language. In summary, the review of the European and the North American studies presented in this chapter, along with their respective research focuses and paradigms, has revealed that both camps have made complementary contributions to portraying the WM–language connections. For example, both camps have contributed to a better understanding of the separate and distinctive roles of the two key WM components (PWM and EWM) as they relate to specific L1 learning sub-domains and processes. The European tradition’s substantial body of research on the phonological nature of WM (in this sense, PWM, or the phonological loop as conceived in Baddeley’s multi-component model) corroborates the hypothetical link between PWM and vocabulary acquisition and development (and possibly for grammar acquisition and development, if grammar is interpreted as longer sequences of morphosyntactic constructions). Alternatively, the North American tradition has helped to reinforce the pivotal role that the attentional or executive elements of WM play in many facets of language processing (subsuming many of the sub-level cognitive processes in language comprehension and production). Aside from the conceptual differences and research focuses in their respective research paradigms and programmes on the WM construct, both traditions have made complementary yet distinctive contributions to developing a clearer picture of the multiple facets of the WM–language connection. Extending these emerging patterns, the strands of WM–L1 research reviewed in this chapter afford two valuable insights into integrating the components/functions of WM with specific L1 sub-domains and activities. Specifically, PWM, as articulated by the European tradition, has been shown to play a significant role in the major acquisitional and developmental facets of L1 acquisition, especially in the acquisition and development of vocabulary, and possibly for morphosyntactic constructions (syntax/ grammar), thus making it a ‘language learning device’ (Baddeley et al., 1998). In contrast, the North American WM research paradigms demonstrate that
58 Part 2: Research Syntheses of Working Memory in L1 and L2 Learning
Table 4.2 A summary of the WM–L1 associations L1 domains and activities
PWM
VS-STM
Verbal WM (rehearsal)
Vocabulary
Yes
Yes
Morphosyntax/grammar
Yes?
Yes?
Listening Reading
Yes Yes
Speaking Writing
EWM
Yes
Yes
Yes
Yes
Yes Yes
PWM = phonological working memory; VS-STM = visuospatial short-term memory; WM = working memory; EWM = executive (aspects of) working memory.
the effects of EWM are more evident in the processing-oriented L1 activities (Cowan, 2013; Just & Carpenter, 1992; Miyake & Friedman, 1998; cf. Van Dyke et al., 2014), and are particularly obvious in certain secondary, offline or ‘post-interpretive’ comprehension processes, such as pronoun detection, ambiguity resolution and the processing of relative clauses (e.g. Caplan & Waters, 1999, 2013; Waters & Caplan, 1996), thus rendering this executive component of WM (EWM) a ‘language processing device’ or simply a ‘language processor’ (Jackendoff, 2007a; Lu, 2011; O’Grady, 2005, 2015). Putting these two views together, Table 4.2 summarizes the various associations between WM and L1 activities. Indeed, these emerging patterns regarding the PWM–EWM distinction between the WM components and their embedded executive functions in cognitive activities such as language learning (be it a first or second language) are being increasingly recognized by more and more researchers in cognitive psychology and psycholinguistics (e.g. Aben et al., 2012; Engel de Abreu & Gathercole, 2012; Szmalec et al., 2013). This dichotomy also serves as the precursor for the integrated framework of WM and SLA that I advocate in Part 3 of this book. Before that, however, I will first synthesize the WMfocused studies in the current SLA research.
5 Working Memory in Second Language Research
In recent years, an increasing number of SLA researchers have been motivated by the WM–L1 association (as reviewed in Chapter 4) and become increasingly interested in exploring the relationships between WM and SLA (e.g. see Juffs & Harrington, 2011; Sagarra, 2013; Wen, 2012a, 2015; Williams, 2012). Due to the fundamental differences between L1 acquisition (L1A) and SLA, researchers have hypothesized that WM may play a greater if not an equal role in SLA than in L1A. In this chapter, I aim to provide an updated review of these theoretical perspectives on WM and SLA, and synthesize the empirical studies on the effects of WM on various areas of L2 learning and processing. In the first section, I summarize the major theoretical perspectives adopted by cognitive-oriented SLA researchers on the role of WM in L2 learning and processing. I mainly discuss the role that WM plays in Skehan’s classical ‘information processing’ perspective and in Nick Ellis’s recent ‘connectionist’ account of SLA. The section following this theoretical discussion tabulates and synthesizes the current strands of empirical research on the effects of WM on various areas of L2 learning and processing. In the third section, I further reveal that these WM–SLA studies have not only produced positive findings on the relationship between the components/functions of WM and the L2 domains and skills, but also revealed the inherent limitations and pitfalls in the current research designs and methodologies. The last section then summarizes the various WM–L2 associations that have emerged from the current empirical research.
Theoretical Perspectives on WM and SLA WM in Skehan’s ‘information processing’ perspective on SLA Embracing the classical information processing approaches, some SLA researchers (e.g. Harrington, McLaughlin, Miyake and Skehan) postulate that L1A is unequivocally dominated by implicit and mostly automatic 59
60 Part 2: Research Syntheses of Working Memory in L1 and L2 Learning
processing (a typical example being the seemingly effortless language acquisition by children in their mother tongue), while SLA usually takes place much later in life (e.g. among adults) and is generally characterized by explicit and more controlled processing (Harrington, 1992; McLaughlin, 1995). In particular, given the perceivable differences in mental lexicon and grammatical competence between L1 and L2 (more restricted), it is reasonable to assume that the SLA process naturally demands more cognitive resources (e.g. during comprehension and production) and is thus presumably more reliant on WM (Skehan, 2014, 2015a; also Wen, 2012a, 2015). Indeed, in recent years there has been a surge of research interest in the WM–SLA relationship, which has been further propelled by the renewed interest in foreign or second language (L2) aptitude research in SLA (Singleton, 2014; Wen, 2012b; Wen et al., 2016). To that effect, some scholars have posited that WM, given its role in so many key areas of SLA, is likely to emerge as the most apt candidate for modifying John Carroll’s L2 aptitude theory (Miyake & Friedman, 1998; Sawyer & Ranta, 2001; Wen & Skehan, 2011; Wen et al., 2016). Of the various contemporary models of L2 aptitude, I examine only Skehan’s macro-SLA aptitude model here, given its pioneering role in this line of research subsequent to Carroll’s early work, while other aptitude models are discussed in Chapter 9 (for recent reviews see Skehan, 2016; Wen et al., 2016). In line with the assumption of the central role of WM stemming from the fundamental differences between L1 and L2, Skehan (1998, 2002, 2012, 2015a, 2015b, 2016), for example, postulates that WM (together with other aptitude constructs, such as phonetic coding ability, language analytical ability and memory retrieval; see column 3 in Table 5.1) permeates different macro-stages of SLA (column 1 in Table 5.1) and their embedded cognitive processes (column 2 in Table 5.1). These SLA stages include, inter alia, the language input stage (e.g. with its embedded processes such as input processing and noticing), the central processing stage (e.g. subsuming pattern recognition, complexification, handling feedback and error avoidance) and the language output stage (e.g. automatization, repertoire creation and lexicalization). Most relevantly, WM is portrayed as a central aptitude construct that affects most of the L2 cognitive processes that are embedded within all three ‘macro-stages’ of SLA (input processing, central processing and output). Taking ‘input processing’ as an example, Skehan (2016) hypothesizes that, all other things being equal, a larger WM confers an advantage on L2 learners as it enables deeper analysis of the incoming materials. The same should hold for ‘noticing’ and ‘pattern recognition’, in which a larger WM is likely to enable the person to notice more material and hold the material longer for internal pattern recognition or identification. In terms of ‘handling feedback’, a growing body of interaction-driven studies in SLA (e.g. Goo, 2012; Granena et al., 2016; Kim et al., 2015; Mackey et al., 2002,
Working Memory in Second Language Research 61
Table 5.1 WM in the stages and cognitive processes of SLA (based on Skehan, 2016) SLA stages
L2 cognitive processes
Aptitude constructs
Language input
Input processing (segmentation)
Attentional control Working memory
Noticing
Phonemic coding ability Working memory
Central processing
Pattern recognition
Phonemic coding ability Working memory Language analysis ability
Complexification (extending, restructuring, integrating)
Language analysis ability Working memory
Handling feedback
Language analysis ability Working memory
Error avoidance
Working memory Retrieval memory
Language output
Automatization
Retrieval memory
Repertoire creation
Retrieval memory Chunking
Lexicalization
Chunking
2010; Révész, 2014) has demonstrated that a larger WM enables L2 learners to notice more corrective feedback, which facilitates L2 development. As Skehan has acknowledged, thus far, only the first five cognitive processes (from ‘input processing’ to ‘handling feedback’; Table 5.1) or the so-called ‘knowledge acquisition’ processes (with a focus on development and reorganization of knowledge) have received preliminary support and are known to be linked particularly more with such aptitude constructs as WM. The remaining four cognitive processes (starting with ‘error avoidance’), which represent areas of increasing ‘control’ over this acquired knowledge, are less known and they should merit much further research in the future (Skehan, 2016). The relevant issues relating to WM and L2 aptitude are further discussed in Chapter 9.
WM in Nick Ellis’s ‘connectionist account’ of SLA In recent years, SLA has increasingly been depicted from the connectionist, usage-based and construction-oriented perspectives (in particular, N.C. Ellis, 1996, 2012, 2013; N.C. Ellis et al., 2013, 2015; Gries & Ellis, 2015). Unlike
62 Part 2: Research Syntheses of Working Memory in L1 and L2 Learning
the mainstream generative grammar (MGG) approaches, the connectionist and construction-based accounts view language learning (be it L1A or SLA) as a process of acquiring linguistic sequences (N.C. Ellis, 1996, 2012) or constructions (N.C. Ellis, 2013; Gries & Ellis, 2015) that is subserved by general cognitive mechanisms such as frequency, saliency and joint attention (N.C. Ellis, 2002). Viewed from this connectionist perspective, for example, the traditional concepts of vocabulary and grammar (from MGG) are no longer considered separable or even necessary dichotomies in language learning (also see Bates & Goodman, 1997; Marchman & Bates, 1994; Römer, 2009) and are deemed to be instances of linguistic sequences or morphosyntactic constructions of different lengths. According to this logic, these linguistic sequences or formulaic chunks can range from lexis (i.e. sequences of individual sounds or phonemes) and formulaic sequences (i.e. fixed routine phrases or collocations) to morphosyntactic constructions (i.e. sequences of morphological and syntactic structures such as verb–argument constructions). Interpreted this way, the main task of language learning is to acquire an adequate number of linguistic sequences or chunks (i.e. the contingency of form–function pairings) that can be drawn upon during language use and communication (e.g. comprehension and production). More relevantly, in a series of laboratory experiments, N.C. Ellis and colleagues (Ellis, 1996, 2002; Ellis & Schmidt, 1997; Ellis & Sinclair, 1996) have demonstrated that WM and, in particular, its phonological component (PWM; or the ‘phonological loop’ in Baddeley’s multi-component model), which consists of the phonological short-term store and the articulatory rehearsal mechanism, plays an instrumental role in ‘consolidating, entrenching and automatizing activation of stable, long-term mental representations of novel phonological material such as individual words, morphemes and lexical sequences’ (Ellis, 2012: 19). Here, however, a distinctive role is also postulated for the two sub-level cognitive functions of phonological shortterm store and the articulatory rehearsal mechanism subsumed under PWM (Ellis, 1996; Verhagen & Leseman, 2016). Indeed, an increasing number of empirical studies have converged on this assumption, generating positive results for constructing a close link between the phonological elements of WM (PWM; including its phono logical short-term store and its articulatory rehearsal mechanism) and L2 lexis/vocabulary (e.g. Cheung, 1996; Ellis & Sinclair, 1996; French, 2006; French & O’Brien, 2008; Service, 1992), and L2 formulaic sequences and collocations (e.g. Bolibaugh & Foster, 2013; Foster et al., 2014; Skrzypek, 2009) and morphosyntactic constructions or grammar acquisition and development (e.g. Martin & Ellis, 2012; O’Brien et al., 2006, 2007; Sanz et al., 2014; Verhagen & Leseman, 2016; Williams & Lovatt, 2003). Thus, these findings suggest that a strengthened connection can be forged between PWM and the acquisition of lexis, formulaic sequences and morphosyntactic constructions (N.C. Ellis, 1996, 2012, 2013; Martin & Ellis, 2012).
Working Memory in Second Language Research 63
Summary Based on the above two major proposals of the WM–SLA association initiated by SLA researchers such as Skehan and Ellis, some preliminary links can thus be forged between WM and SLA, which are similar to the WM–L1 connection (as reviewed in Chapter 4). Indeed, these assumptions have been increasingly recognized and accepted by more and more SLA researchers, who have conducted a considerable number of empirical studies on the possible effects of WM on various aspects of L2 learning and processing (for a comprehensive meta-analysis, see Linck et al., 2014; also see Wen et al., 2013, 2015). These studies have largely emulated the research paradigms of the two WM–language traditions in cognitive psychology and explored the relationships between the various WM components (in particular, PWM and EWM) and different domains of L2 acquisition and development (e.g. vocabulary, formulaic sequences and grammar) and skills learning and processing (e.g. listening, speaking, reading, writing and interpreting). To develop a comprehensive view of the status quo in current WM– SLA research, in the next section I synthesize all of the major strands of the recent WM–SLA research. The expectation is that the major findings from these empirical studies support the established relationship between the WM components and various areas of L2 learning and processing as proposed by Skehan and Nick Ellis.
Empirical Studies of WM and SLA To review the body of WM–SLA research, it is necessary to adopt some of the latest techniques that have recently emerged in the SLA literature, such as the research synthesis and meta-analysis procedures (Norris & Ortega, 2000, 2006; Plonsky & Oswald, 2012). Previously, Watanabe and Bergsleithner (2006) and Linck et al. (2014) conducted meta-analytical studies of WM–SLA research. Building on these two meta-analyses, in this section I tabulate representative current WM–SLA studies and highlight the key results of the studies that have investigated the various relationships between WM and specific areas of SLA. The first step in the research synthesis normally involves identifying the relevant literature (i.e. the L2 domains) to be covered. Following the two meta-analyses, I manually track and locate the empirical studies that have investigated the effects of WM (components) on different areas of SLA. ‘Fugitive literature’ such as in-house journals, unpublished papers and unpublished/uncirculated proceedings are not included in the analysis (Norris & Ortega, 2000, 2006). However, similar to Linck et al. (2014; cf. Watanabe & Bergsleithner, 2006), unpublished doctoral dissertations are
Participants (age and proficiency)
Adults/intermediate
Children
Child
Children
Children
Adults (Italian polyglots)
Children
Children
Adults/low
Adults/low
Adults/low
Adults/intermediate
Adults/intermediate
Adults/intermediate
Authors
Harrington & Sawyer (1992)
Service (1992)
Speidel (1993)
Geva & Ryan (1993)
Service & Craik (1993)
Papagno & Vallar (1995)
Service & Kohonen (1995)
Cheung (1996)
N. Ellis (1996)
Ellis & Sinclair (1996)
Ellis & Schmidt (1997)
Berquist (1997)
Miyake & Friedman (1998)
Atkins & Baddeley (1998)
PWM (NWR)
EWM (RST)
PWM (DS); EWM (RST)
PWM (NWR)
PWM (NWR)
PWM (rehearsal suppression)
PWM (NWR)
PWM (NWR)
PWM (NWR)
PWM (NWR)
PWM (NWR)
PWM (NWR)
PWM (NWR)
PWM (DS); EWM (RST)
Targeted WM components and WM measures
Table 5.2 Empirical studies investigating WM–SLA connections (from 1992 to 2015)
Correlational
Path analysis
Correlational
Experimental: correlational
Experimental: correlational
Experimental: correlational
Correlational
Longitudinal: correlational
Experimental: correlational
Longitudinal: correlational
Correlational
Longitudinal: case study
Longitudinal: correlational
Correlational
Research design and methodology
L2 vocabulary learning
L2 reading comprehension
L2 reading comprehension
L2 grammar (local agreement rule)
L2 vocabulary & grammar
L2 vocabulary & grammar
L2 vocabulary
L2 vocabulary development
L2 (Russian) words
L2 vocabulary development
L2 sub-skills
L2 oral development
L2 vocabulary development
L2 reading comprehension
Results and findings (affected SLA areas)
64 Part 2: Research Syntheses of Working Memory in L1 and L2 Learning
Adults/ intermediate
Adults/intermediate
Adults/ intermediate
Adults/intermediate
Adults/intermediate
Adults/intermediate
Adults/intermediate
Adults/low
Adults/intermediate
Adults/intermediate
Adults/low
Adults/intermediate
Children
Adults/intermediate
Fortkamp (1999)
Service et al. (2002)
Robinson (2002b)
Mackey et al. (2002)
Payne & Whitney (2002)
Fortkamp (2003)
Abu-Rabia (2003)
Williams & Lovatt (2003)
Chun & Payne (2004)
Juffs (2004)
Speciale et al. (2004)
Walter (2004)
Masoura & Gathercole (2005)
Payne & Ross (2005)
PWM (NWR); EWM (RST)
PWM (NWR in L1 and L2)
EWM (RST)
PWM (NWR)
EWM (RST)
PWM (NWR); EWM (RST)
PWM (NWR)
EWM (RST)
PWM (NWR) + EWM (SST)
PWM (NWR); EWM (RS)
Composite WM score (NWR + LS)
EWM (RST)
PWM (NWR)
PWM (NWR); EWM(RST)
Correlational
Experimental: ANOVA
Correlational
Study 2: longitudinal
Study 1: experimental; correlational
Correlational
Correlational
Experimental: correlational
Correlational
Correlational
Correlational
Correlational
Experimental: correlational
Correlational
Correlational
L2 oral development in CMC context
L2 vocabulary
L2 reading comprehension
L2 vocabulary
L2 syntactic processing
Frequency of click-up in dictionary
L2 grammar (rule-learning)
L2 reading & writing
L2 task-based speech performance
L2 oral speech (CMC context)
L2 feedback noticing
L2 incidental learning
L2 vocabulary
L2 speech
Working Memory in Second Language Research 65
Adults
Adults /intermediate
Children
Adult
Adults/intermediate
Adults/intermediate
Children + adults
Adults/intermediate
Children
Adults/intermediate
Adults/intermediate
Winke (2005)
Van den Noort et al. (2006)
French (2006)
O’Brien et al. (2006)
Bergsleithner (2007)
Gu & Wang (2007)
Kurvers & van de Craats (2007)
Leeser (2007)
O’Brien et al. (2007)
Trofimovich et al. (2007)
Sagarra (2007)
Composite WM (NWR + RST)
Composite WM (NWR + RST)
PWM (NWR)
EWM (RST)
PWM (DS + NWR)
EWM (RST)
EWM (SST)
PWM (NWR)
PWM (NWR)
PWM (DS + letter number order) + EWM (RST)
Composite WM score (NWR + LST)
Noticing and L2 development
L2 speech
L2 speech
L1, L2, and L3 proficiency
L2 reading
Experimental: correlational
Experimental: correlational
Partial correlational; hierarchical regression
Experimental: correlational
L2 noticing/recasts
L2 recasts
L2 speech
L2 reading comprehension
Correlational analyses L2 vocabulary and reading skills
Correlational analyses Listening comprehension
Experimental; correlational
Partial correlational; hierarchical regression
Partial correlational; hierarchical regression
Correlational
Experimental: correlational
66 Part 2: Research Syntheses of Working Memory in L1 and L2 Learning
Adults/intermediate
Adults/intermediate
Adults/highly proficient EWM (RST)
Adults/intermediate
Adults/non-novice
Adults/intermediate
Adults/intermediate
Adults/intermediate
Adults/intermediate
Adults/intermediate
Guará-Tavares (2008)
Kormos & Sáfár (2008)
Rodriguez (2008)
Fontanini & Tomitch (2009)
Hummel (2009)
Payne et al. (2009)
Sunderman & Kroll (2009)
Wright (2009)
Weissheimer & Fortkamp (2009)
Alptekin & Erçetin (2010)
Experimental: correlational
Correlational
Partial correlational; hierarchical regression
EWM (L1 RST + L2 RST)
EWM (L2 SST)
Composite WM score (backward digit + story recall + combined word span and sentence span
EWM (RS)
Counting span task
PWM
EWM (RST)
Correlational
Experimental: correlational
Experimental: correlational
Experimental: multiplicative regression analysis
Correlational
Regression
Correlational
Experimental: correlational
PWM (NWR); EWM (N-Back) Cross-sectional
EWM (SST)
EWM (SST)
Adults/intermediate
Finardi & Weissheimer (2008)
PWM (NWR)
Children
French & O’Brien (2008)
L2 reading comprehension (literal and inferential comprehension)
L2 speech performance
L2 grammatical proficiency (immersion setting)
L2 lexical comprehension & production (study-abroad context)
L2 reading comprehension
L2 proficiency (vocabulary, grammar)
L2 reading comprehension (linear texts and hypertext)
L2 sentence processing/ parsing
L2 sub-skills (speaking, grammar, etc.)
L2 speech performance
L2 speech development
L2 speech
Working Memory in Second Language Research 67
Children
Adults/intermediate
Adults/intermediate
Adults/intermediate
Adults/intermediate
Adults/low
Adults/low
Adults/slightly above intermediate
Adults/low
Adults (elderly)
Adults/intermediate
Andersson (2010)
Bergsleithner (2010)
Dussias & Piñar (2010)
Goo (2010)
Mackey et al. (2010)
Skrzypek (2010)
Gass & Lee (2011)
Kormos & Trebits (2011)
Linck & Weiss (2011)
Mackey & Sachs (2012)
Rai et al. (2011)
EWM (OST)
PWM + EWM (LST)
EWM (OST)
L2 reading comprehension and written production
L2 syntactic process (longdistance ‘wh-’ questions)
L2 written performance
L2 reading comprehension (listening, reading, story reading)
ANOVA and bivariate regression analyses
Cross-lagged
Regression analysis and ANOVA
Correlational
Longitudinal (6 weeks): regression analysis
L2 comprehension (inferences)
L2 interactional feedbacks
L2 explicit knowledge (vocabulary and grammar)
L2 narrative tasks speech performance
L2 speech
L2 vocabulary and collocational knowledge
Experimental: general L2 production of modified linear model (GLM) output
Experimental: regressional; ANCOVA
Experimental: ANOVA
Experimental: correlational
Longitudinal (1 to 2 years): correlational; multiple regression analyses
EWM (backward digit span) ANOVA and MANOVA
EWM (RST in L1 & L2)
PWM (NWR)
EWM (LST)
EWM (LST + OST)
EWM (RST)
EWM (OST)
PWM (DS + WS + phonological fluency) + EWM (animal dual task + counting span + trail making span)
68 Part 2: Research Syntheses of Working Memory in L1 and L2 Learning
Adults/novice
Adults/intermediate
Children
Adults/intermediate
Adults/intermediate
Adults/intermediate
Adults/intermediate
Adults/intermediate
Adults/intermediate
Adults/low (ab inito)
Children
Martin & Ellis (2012)
Révész (2012)
Miettinen (2012)
Li (2013)
Yilmaz (2013)
Zhao (2013)
Alptekin et al. (2014)
Foster et al. (2014)
Hopp (2014)
Sanz et al. (2014)
Tsimpli et al. (2014)
PWM (BDS) & VSWM
EWM (LST)
EWM (RST in L2)
PWM (serial recall)
Four RSTs (two in L1; two in L2)
PWM (DS & NWR) + EWM (LST)
EWM (OST) + LAA
EWM (OST) + LAA
PWM (NWR in L1 and L2)
PWM (DS & NWR) + EWM (RST)
PWM + EWM (LST)
L2 corrective feedback
L2 knowledge (vocabulary, sentence, and story)
L2 recasts
L2 vocabulary and grammar learning
L2 corrective recasts
Correlations
Longitudinal: correlations & repeated-measures ANOVAs
Mixed linear regression analyses
Regression analysis
L2 narrative retelling
L2 morphosyntax
L2 relative clause attachment preferences
L2 native-like selection in immersion
Correlations; principal L2 reading comprehension components factor analysis
Experimental: 3×3 ANOVA
Experimental: mixed- L2 explicit correction and factor ANOVA recasts
Experimental: structural equation modelling
Correlation and cluster analysis
Many-faceted Rasch model and regression analyses
Hierarchical regression analysis and structural equation modelling
Working Memory in Second Language Research 69
Adults/low
Adults/beginning, PWM (DS); EWM (OST) intermediate, advanced
Adults/intermediate
Children
Children
Linck & Weiss (2015)
Serafini & Sanz (2015)
Suda (2015)
Swanson (2015)
Verhagen et al. (2015)
Correlations and regressions
Longitudinal: regression analysis
Experimental: 4-way ANOVAs
Longitudinal (3.5 months): correlation & regression analysis
Longitudinal (6 weeks): correlation & regression analysis
Correlations and regressions
Correlations and regressions
L2 grammar skills (subject– verb agreement, auxiliaries, and verb placement, and a Dutch vocabulary test)
L2 reading/literacy
L2 processing (relative clauses)
L2 morphosyntax (only among low-level L2 learners)
L2 proficiency (self-rating in reading, writing, speaking, listening)
L2 grammar (OST with explicit; RST with incidental)
L2 overall competence (vocabulary, semantics, syntax) except for morphology.
WM = PWM + EWM; PWM = phonological working memory; EWM = executive working memory; DS = digit span; FDS = forward digit span; BDS = backward digit span; WS = word span; NWS = non-word span; LAA = language analytical ability; RST = reading span task; NWR = non-word repetition/recognition span task; SST = speaking span task; LST = listening span task; OST = operation span task; VSWM = visuospatial working memory.
PWM (serial recall, highprobability non-words, and low-probability nonwords)
PWM (FDS; BDS; WS; NWS) & EWM (a conceptual span, LST, rhyming span, updating task)
RST
EWM (OST) + inhibitory control (Simon Task)
EWM (OST & RST)
Adults/novice
Denhovska et al. (2015)
PWM (FDS)
Children
Vulchanova et al. (2014)
70 Part 2: Research Syntheses of Working Memory in L1 and L2 Learning
Working Memory in Second Language Research 71
included in the research synthesis (e.g. Bergsleithner, 2007; Guará-Tavares, 2008; Mizera, 2006; Skrzypek, 2010; Winke, 2005). In addition to the inclusion of unpublished doctoral dissertations, there are two other notable differences between the research synthesis method implemented here and that executed by Watanabe and Bergsleithner (2006). Similar to Linck et al. (2014), the analysis also includes more updated studies. Whereas Watanabe and Bergsleithner (2006) examined papers published between 1972 and 2006, I include more relevant WM–SLA studies published between 1992 (beginning with the seminal paper by Harrington and Sawyer) and 2015; also see Wen (2014) for a research time-line of the developments in cognitive psychology in terms of theorizing and measuring WM in language research that dates back to much earlier periods. Moreover, Watanabe and Bergsleithner only covered studies involving participants aged 18–60 years in their analysis, while this synthesis includes studies with participants aged well below 18 and those aged over 60 (e.g. Mackey & Sachs, 2012). Both younger learners and senior participants are included because age (and L2 proficiency) is an important factor in the analysis of the relationship between WM and L2 development (Li, 2015). Moreover, in line with the key concerns of this chapter, which is to delineate the state-of-the-art WM–SLA research, four key content areas are coded and extracted from the literature (as clearly indicated by the column headings in Table 5.1). In Table 5.2, the publication details of the studies are kept to a minimum, that is, only the family names of the authors and the year of publication are included, while the other details are provided in full in the list of references. The coding procedures are based on four categories, namely, the age of the participants (i.e. children versus adults, but more importantly, their L2 proficiency as well), the research methodology (WM components and measures) and design (correlation, experimental, longitudinal etc.), and finally the research objectives or research questions and their results and findings (i.e. targeted L2 domains that are found to be related to the variable of WM). Based on these inclusion and exclusion criteria, the final pool comprises 80 studies, compared with 20 studies in Watanabe and Bergsleithner (2006) and 77 studies in Linck et al. (2014). All of the extracted information is displayed in Table 5.2, with the studies presented in chronological order. It is hoped that this sample of empirical studies represents a broad overview of the diverse streams of current WM–L2 research.
General Findings of the Current WM–SLA Studies As Table 5.2 demonstrates, four focus points are apparent among the current WM–SLA studies (as reflected in the four column headings): (1) the
72 Part 2: Research Syntheses of Working Memory in L1 and L2 Learning
age of participants and their L2 proficiency levels; (2) the specific WM measures adopted; (3) the research design and methodology; and (4) the major results and key findings regarding the targeted L2 domains and processing activities as a function of the effects of WM. Each of these factors is briefly discussed in this section. The studies clearly cover two major age groups: children and college-aged adults, with the exception of Mackey and Sachs (2012), who specifically targeted elderly participants. However, in terms of the WM components and assessment procedures, the studies are more diverse and far from consistent. First, although most of the authors claimed that they aimed to investigate the relationship between WM and L2 learning, they vary substantially with respect to the manner in which they conceptualized the WM construct, and the WM measures and assessment procedures they adopted. This is the case even among the studies that targeted the same L2 areas and studies that tested the same component of WM. When discussing the possibility of replicating the work of Révész (2012) and Goo (2012), Gass and Valmori (2015) expressed a similar concern with respect to the L2 interaction-oriented studies on the relationship between WM and corrective feedback, and they called for more consistency in WM measures and procedures (more details are provided in the next chapter). In line with the conception of the WM construct (as discussed in Chapter 2), these WM–SLA studies can be further categorized into two groups according to the specific WM span tasks they implemented. That is, studies that adopted any version of the simple span tasks (such as the digit span task or the non-word repetition span task) are classified and coded as investigating the effects of PWM. Alternatively, the studies that adopted any version of the complex span tasks (such as the reading span task or the operation span task) are coded as investigating the effects of the executive function of WM. Studies that implemented both the simple and complex span tasks are coded as investigating the overall effects of WM (subsuming both the PWM and EWM components). In terms of research design, most of the studies adopted a correlational (and regressional) design, and some incorporated an experimental design (such as the interaction-driven L2 studies). However, few studies adopted a longitudinal design to provide more insights into the co-adaptive developments of WM and SLA (Denhovska et al., 2015; Skehan, 2016; Wright, 2015). Finally, regarding the major results and key findings, the targeted SLA areas mostly comprise the domains followed by the European tradition of WM–language research (i.e. by focusing on L2 vocabulary or grammar) or any of the four L2 sub-skills (i.e. listening, speaking, reading and writing) domains followed in the North American tradition. In addition to the areas of focus that are similar to those in the WM– L1A research (as reviewed in Chapter 4), there are a number of additional L2-specific learning domains that are not present in the L1A-based research
Working Memory in Second Language Research 73
paradigms. For example, (1) studies that explored the relationship between WM and L2 formulaic sequences and collocations (e.g. Foster et al., 2014; Skrzypek, 2010), and (2) the large number of studies that investigated the relationship between WM and L2 interaction, particularly regarding its role in noticing or the intake of L2 corrective feedback or recasts (e.g. Goo, 2010; Mackey et al., 2002; Révész, 2012; for a comprehensive review see Mackey, 2012). These studies, together with the strands that are not included here (such as those that investigated the relationship between WM and bilingual processing and interpreting, which are discussed in the next chapter), have largely broadened the research scope of the existing WM–language research paradigms initiated in cognitive psychology and psycholinguistics. Thus, these studies are likely to make significant contributions to our understanding of the crucial human cognitive constructs of WM (cf. Dekeyser & Juffs, 2005; Wen, 2015; Wen & Yi, 2015). These contributions are further discussed in later chapters.
Critique of the Current WM–SLA Studies Despite the generally close association between WM and SLA, a closer look at the current WM–SLA studies (Table 5.2) reveals some inherent limitations in research practice that deserve more caution on the part of SLA researchers (including cognitive psychologists) pursuing this line of inquiry. These caveats, and their potential threats to the SLA field, are reviewed below (also see Wen et al., 2015). The first and most imperative issue relates to the confusing use of the term ‘working memory’ (or its short form WM) in the SLA literature. In particular, some of the WM–SLA studies used the same broad term ‘WM’ to denote several different things or, to be more exact, to implicate different factors or components of the same construct (e.g. its phonological component or its executive component). This approach can create formid able difficulties when the results and findings of the studies are synthesized in meta-analyses of WM effects (e.g. Watanabe & Bergsleithner, 2006; Linck et al., 2014). A more detrimental consequence of this inconsistent use of the term WM (and its related terms) is that it can lead to methodological pitfalls. A possible caveat is that some WM–SLA studies (e.g. Mackey et al., 2002; Sagarra, 2007; Winke, 2005) combine L2 learners’ PWM and EWM scores in a so-called standardized ‘composite Z score’ of WM for statistical analysis. This practice also merits some caution. First, such composite scores are anathema to the multiple components/functions view of the WM construct (as discussed in Chapter 2). Second, such composite scores may obscure or confound the separate and distinctive contributions (or constraints) of
74 Part 2: Research Syntheses of Working Memory in L1 and L2 Learning
the two WM components (i.e. PWM and EWM) to specific SLA processes. Intuitively, a large PWM does not necessarily imply a large EWM (Skehan, 2012: 385). Last but not least, this approach (standardizing and averaging the simple and complex memory scores to obtain a composite Z score of WM) is not common practice in cognitive psychology research (which actually explains the divide between the European and the North American WM research camps). For these and other reasons, it is imperative that a more clear and finegrained distinction be made between the multiple components/functions of WM (particularly the phonological and executive components) implicated in SLA. Resolving this caveat is the first key tenet of the integrated framework of WM and SLA discussed in the next section. If this caveat implicates the issue of the theoretical taxonomy at the conceptual level, the second caveat in the current WM–SLA studies relates to the methodological considerations. An issue here concerns the domainspecificity versus domain-generality of the WM measures (Chan et al., 2011; Skehan, 2016), as most of the WM–SLA studies do not address this dichotomy directly when deciding on which WM measures to adopt (for similar arguments, see Dekeyser & Juffs, 2005; Juffs, 2006; Juffs & Harrington, 2011; Wen et al., 2016). This becomes a particularly acute issue when SLA researchers need to choose, for example, among the daunting number of WM span tasks offered in cognitive psychology. As Table 5.2 demonstrates, SLA researchers have used a variety of simple versions of the memory span tasks, such as the digit span, word span, non-word recognition span and non-word repetition span tasks, to investigate the effects of PWM. Other SLA studies have also adopted an array of complex memory span tasks to assess EWM, such as the reading span, operation span and N-back span tasks. Moreover, other studies have used a (backward) digit span task to assess EWM (e.g. Brunfaut & Révész, 2015; Kormos & Sáfár, 2008; Révész, 2012). That is, there is much inconsistency in the WM measures and scoring procedures used in the current WM–SLA studies, which often creates difficulties and limitations in interpreting and comparing the results (Gass & Lee, 2011; Gass & Valmori, 2015; Juffs, 2006; Linck et al., 2014; Wen, 2012b). Adding to this second methodological issue with the WM measures is the complicated scoring procedures of the WM span tasks. At least two calculation procedures are available in the cognitive psychology literature (Conway et al., 2005; Miyake, 2001). First, the so-called ‘maximum set size’ score gives the participants a credit for a set if they repeat all of the items in the set correctly. Second, the so-called ‘total performance score’ gives the participants a credit for all of the items they successfully recall. However, the advantages and disadvantages of these two scoring measures are not clearly stated. Overall, most WM–SLA studies have not sufficiently addressed these issues when deciding on the WM measures and scoring procedures to be
Working Memory in Second Language Research 75
implemented (for a similar argument, see Dekeyser & Juffs, 2005; Juffs, 2006; Juffs & Harrington, 2011). In view of these difficulties, it is essential to develop a more standardized procedure for SLA researchers to select the appropriate WM measures and to score the WM span tasks. This is a second key concern in developing the integrated framework for WM and SLA. The third caveat in the extant WM–SLA studies relates to the research objectives and research design. The synthesis of the WM–SLA studies (e.g. Table 5.2) shows that most of them mainly focused on the direct association between WM and SLA (i.e. by targeting the main effects of WM). In this sense, they may have ignored other possible and even more illuminating effects of WM that should have been included in the research questions or research design (cf. Linck & Cunnings, 2015). An obvious candidate is the interaction effects of WM on other internal (e.g. L2 proficiency; Gass & Lee, 2011) and external factors (e.g. planning time in task completion; Guará-Tavares, 2008). Admittedly, targeting the main effects of WM is a necessary and important first step, but it is far from sufficient. Other effects (such as the interaction effects) that may play a role also merit investigation. In addition, most of the current WM–SLA studies adopted a ‘correlational design’ to correspond with their research objective (targeting the main effects of WM). Obviously, correlation is not equal to causation (McLaughlin, 1995; Miyake & Friedman, 1998). Moreover, although a correlational design is helpful in establishing the initial association between WM and SLA, it falls short of offering a more satisfactory and powerful explanation of the nature and strength of the associations that contribute to developing a deeper understanding of the SLA process (from a WM perspective). However, an encouraging sign in the WM–SLA research is that a number of recent studies adopted either an experimental design (e.g. Guará-Tavares, 2008; Mackey et al., 2002, 2010) or a longitudinal design (e.g. Andersson, 2010; French, 2006; Linck & Weiss, 2011, 2015). Another commendable practice is the use of sophisticated statistical methods such as structural equation modelling (SEM; e.g. Li, 2013; Miyake & Friedman, 1998; Verhagen & Leseman, 2016) and meta-analysis (Linck et al., 2014; Watanabe & Bergsleithner, 2006). Nonetheless, given that the WM–SLA enterprise is still in its infancy (Bunting & Engle, 2015; Williams, 2015), more fruitful and effective links between the WM components and SLA domains/activities (that can reflect the complexities of the relationships) need to be forged. This is the overarching and ultimate motivation for the integrated WM–SLA framework proposed in this book. Above all, as many of the WM–SLA studies relied on existing and established monolingual-oriented and comprehension-oriented WM research paradigms in cognitive psychology, they may have failed to adequately consider the specific and unique features of SLA research (Wen, 2015). Unlike L1 language learners, second language learners have two languages
76 Part 2: Research Syntheses of Working Memory in L1 and L2 Learning
in their (bilingual) brain. Studies ignoring this key difference do not reflect the current situation (i.e. there are now more bilingual learners than monolingual speakers; Grostean & Li, 2013). Therefore, it is critically important to include both the WM and SLA perspectives so that researchers can more effectively address research questions that relate to both fields.
Summarizing the WM–L2 Association The above review of the WM–L2 theories and empirical studies demonstrates that the WM construct is not only related to L1 acquisition and processing (Chapter 4), but also has significant implications for different areas of L2 learning and processing, including domains such as L2 vocabulary, formulaic sequences, grammar and L2 sub-skills and processes (Juffs & Harrington, 2011; Williams, 2012). Echoing the theoretical assumptions held by some SLA researchers (e.g. Skehan and Ellis), this body of WM–SLA research largely corroborates their assumptions with affirmative evidence of a positive and strong link between WM and SLA. For this reason, the critical role of WM in L2 learning can now be considered firmly established. Furthermore, these strands of WM–SLA research (Alptekin & Erçetin, 2010; Andersson, 2010; Ando et al., 1992; Gu & Wang, 2007; Hasegawa et al., 2002; Miyake & Friedman, 1998) indicate that, compared with its effects on L1 learning, WM seems to exercise greater if not equal effects on L2 learning and processing (Skehan, 2015a; Wen, 2012b). Furthermore, due to the many controversies and debates surrounding the WM construct and the daunting number of WM measures and assessment procedures in the source discipline of cognitive psychology, some of the WM–SLA studies are fraught with limitations and caveats that have led to severe pitfalls in their research designs and methodologies. In view of this, it is critical to develop a more principled approach to conceptualizing and operationalizing the WM construct in SLA so that future research can be based on a more solid theoretical foundation. To address these issues, in Part 3 of the book, I propose an integrated perspective on WM and SLA.
Part 3 Toward an Integrated Perspective on Working Memory and SLA
6 An Integrated Framework for Working Memory and SLA Research
In the last chapter, I identified some potential limitations and caveats in the current WM–SLA research, which naturally led to a call for a more principled approach to conceptualizing and measuring WM in SLA research. To this end, I propose a conceptual framework for theorizing and measuring WM in SLA research that integrates current research insights from both the cognitive sciences (Chapters 2 and 3) and the emerging patterns in the findings on WM in L1 (Chapter 4) and L2 learning (Chapter 5). As such, in the first two sections of this chapter, I reconceptualize and redefine the two key constructs of WM and SLA, and the new versions of these will be adopted in the integrated framework. Based on these working definitions, in the third section I elaborate the integrated conceptual framework, including its overall structure and key constituents. The following section highlights the basic tenets of the framework and further outlines some general principles for applying the integrated framework in practical research. The last section summarizes the possible empirical consequences for future research and concludes the chapter.
Reconceptualizing and Redefining WM in SLA Research The review of the theoretical literature on WM in the cognitive sciences in Chapter 2 showed that despite the theoretical debates and controversies surrounding the concept, most cognitive psychologists conceptualize WM as a multiple components/functions system that comprises both domainspecific storage mechanisms and domain-general executive functions (Conway et al., 2007; Coolidge & Wynn, 2009; Fenesi et al., 2015; Miyake & Shah, 1999; Williams, 2012, 2015). In line with this fractionated view of WM in cognitive psychology (Baddeley, 1996, 2003a, 2012, 2015), a working 79
80 Part 3: Toward an Integrated Perspective on Working Memory and SLA
definition of the WM construct in the integrated framework presented here is as follows: those multiple cognitive mechanisms and functions implicated in the temporary maintenance, access and control of a limited number of pieces of linguistic information to facilitate the acquisition, representation, processing and development of various domains and activities in learning a second language. This definition should concur with the unifying characterizations of the WM construct discussed in Chapter 2. First, the key and signature feature of the WM construct is its limited capacity (Carruthers, 2013, 2015). Therefore, it is better to conceptualize WM in SLA as possessing limited capacity (in terms of the amount of linguistic information it can hold or focus on, and the duration of such holding), and thus as likely to incur some kind of central bottleneck effect on complex cognitive activities such as L2 learning and processing, which normally involve more than one sub-task (Ferreira & Pashler, 2002; Pashler, 1984, 1992). In terms of SLA, this particular characterization of the WM construct translates into the so-called ‘tradeoff ’ effect evident in many first language learning processes (e.g. Just & Carpenter, 1992) and second language performance areas, such as L2 task-based speech planning and performance (more details are provided in the Chapter 8; also see Skehan, 1998, 2009, 2014, 2015a). A second important feature of WM as defined here are the multiple components/functions that facilitate the execution of complex human cognitive activities/tasks (in this case SLA). Based on this characterization, it is better to conceptualize WM in SLA as a memory sub-system (e.g. alongside sensory-based short-term memory and storage-based long-term memory) that comprises multiple components (e.g. the phonological, visuospatial and executive components, and the activated long-term memory component). Each component is associated with embedded cognitive mechanisms or executive functions, such as phonological short-term store and articulatory rehearsal, for the phonological component (PWM), and executive functions, such as information updating, switching and inhibition, for the executive component (EWM). A third characterization of WM derived from the unified theories presented in Chapter 2 is that long-term memory (LTM) plays an integral role in WM (in terms of both content and function). In this case, WM can be regarded as a gateway to LTM, thus indicating a bi-directional transfer of information between WM and LTM. For example, in language comprehension processes (such as listening and reading), linguistic material temporarily processed in WM may be marshalled into LTM, thus permanently changing its contents (information is thus consolidated and consequently learning takes place). In contrast, in the language production processes of speaking and writing, linguistic materials are either activated in or retrieved from the LTM before they are produced as output (in spoken or written form). In this sense, the third characterization reflects the interaction between what is
An Integrated Framework for Working Memory and SLA Research 81
being processed or manipulated in the active WM and the more passive LTM warehouse. In Cowan’s terms, WM is the ‘focus or spotlight of attention’ (Cowan, 1999, 2012, 2014, 2015).
Putting SLA Domains, Skills and Processes in Better Order To define the SLA side of the ‘WM–SLA nexus’ (Wen, 2012a, 2012b), the framework incorporates the latest trends and approaches to SLA, in which L2 acquisition is no longer reduced to a monolithic and static concept, but is, rather, increasingly recognized as a multi-faceted, complex, dynamic, emergent and adaptive process (e.g. de Bot, 2015; Ellis & Larsen-Freeman, 2009; Larsen-Freeman, 1997, 2011, 2015a, 2015b). More relevantly, this definition specifically describes the nature, structure and contents of SLA as consisting of sub-domains and sub-skills that span acquisition/ learning, processing/performance and development. On the one hand, the L2 sub-domains include acquisitional and developmental domains such as L2 phonemes, lexis, formulae and morphosyntactic constructions (as conceived in Nick Ellis’s connectionist account). On the other hand, the L2 processing and performance areas straddle L2 sub-skills such as listening, speaking, reading, writing and translation/interpreting. Finally, L2 development is conceptualized mainly in terms of L2 proficiency, which includes the beginning (ab initio or novice learners), intermediate, post-intermediate and very advanced/native-like stages. This assumption is also largely compatible with VanPatten’s (2010, 2013) recent argument for a two-facet view of SLA. In this sense, the domain areas of SLA proposed here can be interpreted as equivalent to VanPatten’s concept of the ‘mental representation’ domain, while the processing part is more or less the same as his concept of ‘skills’ (except that the skill of ‘translation/interpreting’ is more conspicuous here as a separate and independent skill). Most importantly, it should be clear that this working definition of SLA does not embrace the mainstream generative grammar (MGG) approach to first language acquisition (e.g. that of Chomsky), which potentially prioritizes grammar or grammatical rules as the core of language learning, and puts everything else (such as memory) in the periphery (Hauser et al., 2002). Instead, the SLA domains are interpreted through the lens of the increasingly prevalent approaches of dynamic systems theory and complexity theory (Larsen-Freeman, 1997, 2015a, 2015b), in which vocabulary (lexis) and grammar (syntax) are no longer regarded as separate entities (cf. Marchman & Bates, 1994; also Römer, 2009; Vulchanova et al., 2014), but rather as intermingled linguistic sequences or formulaic chunks or morphosyntactic constructions that co-adapt (N.C. Ellis, 1996, 2012, 2013). Given the elusive
82 Part 3: Toward an Integrated Perspective on Working Memory and SLA
nature and problematic concept of the generative term ‘grammar’ (Cook & Singleton, 2014; also see de Bot, 2015), the integrated framework follows these latest trends in SLA and thus opts for the terms ‘lexis’, ‘formulaic sequences’ and ‘morphosyntactic constructions’ (or ‘morphosyntax’) to represent the variegated units of linguistic sequences/constructions in the SLA acquisitional and developmental domains. Overall, it can be argued from the above discussion that the working definitions of WM and SLA adopted here not only draw on the unifying theories of the WM construct in multiple disciplines in the cognitive sciences, but also put the SLA domains and processing activities in better order by incorporating the latest accounts of the connectionist, emergentist and construction-oriented views of language acquisition and development (Ellis & Larsen-Freeman, 2009; N.C. Ellis, 2013). It is also hoped that such definitions not only echo but significantly expand the ideas initiated by Nick Ellis regarding the complex relationships between the components/ functions of WM (in particular, PWM) and the SLA domains and activities (N.C. Ellis, 1996, 2012, 2013).
Toward an Integrated Framework for WM in SLA Research Having adequately defined both WM and SLA, in this section I elaborate the structural design of the integrated framework and then discuss its essential constituents. As such, I develop a conceptual framework that integrates the research findings from both cognitive psychology (on WM) and applied linguistics (on SLA) to guide future interdisciplinary WM–SLA studies, with the ultimate goal of elucidating the complex and dynamic relation ships between the components/functions of WM and the SLA domains and processing activities. I now describe the design and structure of the integrated framework for conceptualizing and measuring WM in SLA. To provide a schematic view, the proposed conceptual framework of WM and SLA is presented in Figure 6.1. As shown, the architecture consists of four levels or layers that integrate the components and functions of WM. From bottom to top, the layers comprise: (1) long-term memory (LTM); (2) the WM components (PWM, EWM and VWM) that constitute the overall WM construct; (3) the putative mechanisms/functions associated with each WM component fractionated from the WM construct; and (4) the specification of the WM measures for assessing each putative WM component. Each of these four components is discussed in the following sections. At the bottom level (the shaded area in Figure 6.1), long-term memory (LTM) serves as a permanent storehouse for various forms of long-term knowledge regarding the two languages in the bilingual brain. Supposedly,
Figure 6.1 An integrated conceptual framework for WM in SLA research
An Integrated Framework for Working Memory and SLA Research 83
84 Part 3: Toward an Integrated Perspective on Working Memory and SLA
LTM includes, inter alia, the bilingual speaker’s native language or L1 competence, which subsumes, for example, the mental lexicon and grammatical competence (e.g. phonology, morphology, semantics and syntax). Alongside this L1 competence, LTM in the bilingual brain also stores the language learners’ L2 knowledge and/or L2 proficiency. In slight contrast to the L1 mental lexicon and grammatical competence, which are more implicit in nature (i.e. more automatic in terms of access or retrieval), bilingual speakers’ L2 knowledge or proficiency tends to be more explicit in nature, that is, usually constrained by more controlled access or retrieval, depending on their developmental stage of L2 proficiency (e.g. see Skehan, 2015a). In addition, organization of the repertoire of L2 knowledge and L2 pro ficiency is postulated to encompass various domains (N.C. Ellis, 1996, 2012), including, inter alia, phonemes (speech sounds and pronunciations in the L2), lexis (words or vocabulary items), formulaic sequences (fixed phrases or other linguistic chunks), morphosyntactic constructions (morphological and syntactic structures), semantics (meaning mappings of word forms and sentences), grammatical and meta-linguistic knowledge (e.g. explicit rules of grammar) as well as pragmatic knowledge (e.g. Bardovi-Harlig, 2008, 2013). Of most relevance and importance, LTM – as conceptualized in the integrated framework – also contains activated items of bilingual speakers’ L1 and L2 knowledge in WM (the overlapping part of EWM and LTM), which can be considered as LT-WM. This idea draws heavily on the LTM perspective advocated and pursued by Ericsson and Kintsch (1995; for a recent update of this concept see Caplan & Waters, 2013). Then, in terms of its overall organization, LTM in the integrated framework embraces and incorporates the declarative and procedural dichotomy (e.g. the declarative/ procedural model of Ullman, 2001, 2005, 2012, 2013; also Morgan-Short et al., 2014; cf. Coolidge & Wynn, 2009). Despite the controversies about the exact nature and structure of the bilingual mind/brain (e.g. Abutelabi, 2008), increasing evidence from multiple disciplines and sources (e.g. behavioural, neurological, genetic and developmental) concurs with this declarative/ procedural distinction of LTM knowledge in the bilingual brain (cf. Green & Abutalebi, 2013; MacWhinney, 2008; Paradis, 2009). More specifically, for example, as conceived in the declarative/procedural (D/P) model of Ullman (2005, 2013, 2015), the declarative memory system mainly subserves idiosyncratic knowledge of the mental lexicon, which includes lexical items and word-specific information such as simple words, irregulars and complements. In contrast, the procedural memory system is postulated to underlie the rule-governed linguistic knowledge that corresponds more to the grammatical elements, which likely include the rule-governed hierarchical and sequential composition of complex forms. Of course, it should be acknowledged that it remains to be seen whether bilingual learners’ L1 and L2 knowledge systems reside in the same area of the brain (Abutalebi, 2008; Tan & Li, 2015).
An Integrated Framework for Working Memory and SLA Research 85
The next level up from the bottom of the integrated architecture consists of the multiple ST-WM components (as opposed to the LT-WM component residing in LTM; also see Caplan & Waters, 2013) that have been fractionated from the general WM construct, as conceived in, for example, Baddeley’s (1986, 2003a, 2012) multi-component model. Supposedly, these ST-WM components should include a phonological component (PWM), an executive component (EWM), the visuospatial WM, or sketchpad (VWM), and the episodic buffer (EB). That is to say, they belong to the activated portion of LTM, i.e. LT-WM (Caplan & Waters, 2013); though they are deemed part of WM, they are not ST-WM in the usual sense. However, among these multiple WM components (ST and LT), the integrated framework focuses on only two ST-WM components, namely PWM and EWM, thus excluding VWM (as indicated by the dotted box outline for this component in Figure 6.1), EB (which does not appear in the framework) and LT-WM. Notably, the exclusion of the VWM, EB and LT-WM does not indicate their absence in language acquisition and use, but is mainly because their relationships with SLA have not been steadily forged (as opposed to the other two components, PWM and EWM; Baddeley, 2015; Wen, 2015). Although the role of VWM has received significant attention in neuroscience research (which unfortunately is beyond the scope and focus of the current volume), relatively few studies have examined its specific role in SLA (cf. Kim et al., 2015; Tan & Li, 2015). Juffs and Harrington, for example, mentioned that future WM–SLA research indeed should focus on ‘replicable measures of WM in new areas in writing in non-alphabetic scripts’ (Juffs & Harrington, 2011: 160), such as Chinese (Tong & McBride-Chang, 2010). This trend, as the authors claimed, should be ‘promising’. This obviously may involve VWM, and more empirical research needs to be conducted in this area before it can become a focus of WM–SLA research. Similar to VWM, the EB has also received relatively little attention among SLA researchers (Baddeley, 2015). However, as discussed in Chapters 4 and 5, PWM and EWM have been demonstrated to be most directly implicated and engaged in various L1 and L2 learning activities. These emerging patterns thus lend support to the theoretical links between PWM and EWM and SLA, to be further elaborated in later sections of the chapter and subsequent chapters of the book. Moreover, as clearly indicated in Figure 6.1 (by the double-ended arrows), PWM and EWM both interact bi-directionally with LTM, including an overlapped section between EWM and LTM that represents the activated portion of LTM (i.e. LT-WM; see Caplan & Waters, 2013; cf. Cowan, 1999, 2005). That is, it is possible that linguistic information can be channelled from WM to LTM through the encoding and consolidating processes (as in the case of the comprehension processes in listening and reading) or vice versa (as in the case of the production processes in speaking and writing). Interpreted this way, the language processes involved in (bilingual) interpreting
86 Part 3: Toward an Integrated Perspective on Working Memory and SLA
can be considered a special sub-skill of language processing that combines the processes of comprehension (of the source text) and production (of the target text) (Cai et al., 2015; Dong & Lin, 2013; Dong & Wang, 2013). This interpretation should have implications for the theories and methodologies of WM–SLA research (as discussed in later chapters). Level 3 of the integrated framework incorporates the putative cognitive mechanisms or functions that are postulated to be associated with the two ST-WM components focused upon (i.e. PWM and EWM). First, a distinction needs to be made between the two concepts of ST-WM components and ST-WM mechanisms/functions. In this sense, the ST-WM components are more relevant for conceptualizing (i.e. understanding) and therefore are mostly used for more abstract and theoretical purposes (as in the case of Baddeley’s multi-component WM model). In contrast, the ST-WM mechan isms or functions are the underlying cognitive mechanisms or functions that are to be operationalized in actual research as they are tapped by the WM span tasks (to be considered in the discussion of the fourth level in the framework). For convenience, the prefix of ‘ST-’ will be dropped from now on and both of these two ST-WM components will be simply called PWM and EWM. Based on the research on WM in cognitive psychology, as discussed in Chapter 2, the mechanisms associated with PWM can be further fractionated into a phonological short-term store and an articulatory rehearsal mechanism (Baddeley, 1986, 1996, 2012, 2015; Baddeley et al., 1998). However, the executive functions of WM are purported to subsume the attention monitoring/regulation or control processes, such as information updating, switching and inhibition (e.g. Miyake & Friedman, 2012). Similar to the ST-WM components, the putative mechanisms associated with the other three possible WM components (VWM, EB and LT-WM) are not the focus of the present framework. Again, their exclusion has more to do with the lack of a thorough understanding and sufficient evidence on their potential effect on L2 learning as afforded by the current SLA research (Chapter 5). Clearly, this exclusion should not pre-empt studies on their possible involvement in L2 learning and processing. As evidence on their effects on language learning and processing accumulates – for instance, Rudner and Rönnberg (2008) have outlined the possible implications of the EB on language comprehension – they can be expected to contribute to a fuller understanding of the overall effects of WM on SLA in the future. Having identified the two most language-relevant WM components and their associated mechanisms and functions, the next issue is how to operationalize (i.e. measure) them in practical SLA research. As shown at the top level in Figure 6.1, the integrated framework proposes to implement separate memory span tasks for assessing PWM and EWM, respectively. These specifications are again derived from the extensive literature on the WM measures in cognitive psychology, as discussed in Chapter 3. As such,
An Integrated Framework for Working Memory and SLA Research 87
the framework uses versions of the ‘simple memory span’ tasks (e.g. the digit span task and the non-word repetition span task) to measure PWM, whereas versions of the ‘complex memory span’ tasks (e.g. the reading span task and its variants, the operation span task, the running memory span task and the N-back span task) are used to assess EWM. More specifically, the rationale for advocating separate measures for assessing PWM and EWM draws on previous discussions in the cognitive sciences on the nature of these two distinctive sets of WM span tasks (Conway et al., 2005; Linck et al., 2014). For example, research conducted by the European WM camp has demonstrated that the non-word repetition span task approximates the two associated mechanisms of PWM – for L1, see Gathercole (2006) and the accompanying commentaries; for L1 and L2, see N.C. Ellis (1996, 2012). However, numerous studies by the North American WM camp have demonstrated the close relationships between complex memory span tasks and the attention-regulating or executive areas of WM in language learning (e.g. Daneman, 1991; Daneman & Carpenter, 1980). Conway et al. (2005) reviewed the major WM measures and general guidelines for the assessment procedures in cognitive psychology, while Linck et al. (2014) offered insights on the WM measures that have been adopted by SLA researchers. To sum up this section, the conceptual framework described above aims to integrate insights from the WM research in the cognitive sciences (in terms the conceptualization and measurement procedures, as discussed in Chapters 2 and 3) with emerging knowledge of first and second language acquisition, especially the two components PWM and EWM (for a more detailed historical account of how the WM construct has been theorized and measured, see Wen, 2014). For instance, as Caplan and Waters (2013) pointed out, ST-WM mechanisms (e.g. storage, rehearsal and updating) and LT-WM mechanisms (e.g. retrieval, underlying L2 proficiency) are essential for the specific domains of language learning and processing. It is hoped that this integrated perspective on WM and SLA will lead to a deeper understanding and a more principled and balanced approach to theorizing and measuring WM in SLA research, thus allowing the two contributing fields of cognitive psychology and SLA to achieve synergies in future WM–SLA explorations, which can in turn benefit both fields.
Basic Tenets and Empirical Consequences of the Integrated Framework It should now be evident that the most important underlying assumption of the integrated framework is that the constituent components of WM, along with their distinctive cognitive mechanisms and functions,
88 Part 3: Toward an Integrated Perspective on Working Memory and SLA
should be treated separately in WM–SLA studies, as they are purported to exert distinct effects on the various SLA domains and activities. That is, the two (ST-)WM constructs of PWM and EWM as identified in the integrated framework (Figure 6.1) should be operationalized and measured by separate WM span tasks in practical WM–SLA studies. This is the key tenet of the integrated framework for WM and SLA. This key tenet should has implications for future WM–SLA research and practice. First and foremost, it indicates that a clear demarcation between the WM components and their associated mechanisms/functions is imperative in order for confusion in taxonomy to be avoided. In this sense, the term ‘(ST-) WM components’ refer to the putative and artificial (i.e. human-imposed) fragmentations of the cognitive construct (for conceptualizing or understanding purposes), while their associated mechanisms or functions are those that should be operationalized and directly (or indirectly) measured by the WM span tasks. By the same token, they are the underlying mechanisms that subserve the cognitive resources for real-world (language) learning activities and tasks (e.g. the SLA domains and processing activities in this case). Moreover, couched within this integrated framework, the closely related components of WM (including PWM and EWM) can be distinguished and further demarcated in future WM–SLA studies (also see Wen, 2012a, 2014, 2015). For that matter, it is proposed that the term PWM (also known as phonological memory, which is equivalent to Baddeley’s phonological loop) should refer only to the sound-based phonological component of WM. However, EWM (equivalent to the WM concept used by most North American research groups) refers to the executive and control mechanisms or functions of WM (but including non-executive processes such as encoding, consolidating and long-term retrieval). Finally, the term ‘working memory’ (‘WM’) should be preserved as an umbrella term (particularly for lay purposes) that subsumes all possible components and all of their associated mechanisms and functions (e.g. maintenance, rehearsal, updating, switching and inhibition) that constitute the whole cognitive construct of WM (Conway et al., 2007; Miyake & Shah, 1999). Furthermore, given that these two key components of (ST-)WM (PWM and EWM) purportedly exert distinctive influences/effects on the different SLA domains and processing activities, the integrated conceptual framework proposes to adopt separate WM measures to assess these two WM components. Specifically, the framework measures PWM by adopting a simple memory span task (e.g. the digit span, or the non-word repetition span task) and posits that EWM should be assessed by a complex memory span task (e.g. the reading span task or the operation span task). Following this basic tenet of the framework (that PWM and EWM are two separate WM components with respect to the SLA domains and processes), future studies should not confound the different versions of each component by obtaining a composite WM score (as discussed in the next chapter).
An Integrated Framework for Working Memory and SLA Research 89
Obviously, a direct consequence and the most significant implication of this key tenet are the concerns that it highlights about the extant WM–SLA studies. That is, some studies have indeed obtained a standardized ‘composite Z score’ by combining and averaging the non-word repetition span scores and the listening span scores (e.g. Mackey et al., 2002; Mizera, 2006). As mentioned, this practice likely confounds and conceals the distinctive contributions or effects of the two separate (ST-)WM components (PWM and EWM), which are in fact associated with rather different mechanisms and functions (see Figure 6.1 and Table 4.1). Given these considerations, the practice of combining the simple memory span tasks (generally indexing PWM) and the complex memory span tasks (purportedly tapping into EWM) to obtain a ‘composite Z score’ of WM should be avoided in future WM–SLA studies. Alternatively, considering the distinctiveness of SLA (which implies that a bilingual learner will have two languages at his or her disposal) and first language acquisition (L1A), the practice of obtaining a composite score may still be viable or feasible provided that it is implemented in a different way (e.g. Wen, 2009, 2016a). That is, the WM measures adopted should purportedly tax the same WM components in the two languages in question (i.e. the participants’ L1 and L2). Obtained in this way, the composite score is likely to make more sense than that obtained by combining PWM and EWM, in that it taps into the same types of mechanisms associated with particular WM components (despite the different languages implemented in the span tasks) that are likely to be implicated in language learning (for a similar argument see Cai et al., 2015; Linck et al., 2014). For example, it is clearly more reasonable to combine the scores of the reading span tasks conducted in participants’ L1 and L2 to arrive at a more balanced score that indexes their ‘bilingual’ WM (i.e. an average of L1 WM and L2 WM). Although a number of empirical studies have assessed the participants’ WM capacity by adopting the WM measures constructed in both L1 and L2 (e.g. Cai et al., 2015), they have not, though, combined the scores by following this procedure (averaging the two scores representing the participants’ WM components in L1 and L2). Nonetheless, this innovative (or unusual?) procedure (as opposed to the normal WM procedures currently implemented in cognitive psychology) needs to be supported by more empirical research before it can be put into practice. This issue should be considered in future WM–SLA research.
Summary To sum up, the key tenet of the integrated framework of WM for SLA is the postulation that the construct of WM consists of multiple components
90 Part 3: Toward an Integrated Perspective on Working Memory and SLA
that are associated with multiple mechanisms/functions distinctively implicated in complex L2 cognitive tasks. Furthermore, it is also argued that, according to the integrated conceptual framework, the two principal components of WM involved in language acquisition and processing, namely PWM and EWM, are the most relevant for SLA research and should be the focus of such research for the time being at least (other possible components, such as the visuospatial sketchpad, are either shown in dotted lines in Figure 6.1 or completely excluded from the framework to indicate that they are not yet ready to be subject to SLA research). To end the chapter, it can be safely argued that the proposed integrated conceptual framework for WM in SLA research has its theoretical foundation in, and derives strong empirical support from, the broad theoretical conceptualizations of the WM construct in cognitive sciences and the more specific WM–language explorations in applied linguistics. It is hoped that this principled approach to theorizing and measuring the WM construct in SLA research will have significant theoretical and methodological implications for future research and practice. As should be apparent from the description of the conceptual framework, the motivation for developing this integrated perspective is to capitalize on research insights from the cognitive sciences and applied linguistics to acknowledge the unique features and common themes in both fields (WM and SLA) and thus achieve synergy effects in future research. Constructed in this balanced manner, the integrated framework can be expected to provide perceivable theoretical advantages that neither cognitive psychology nor SLA alone can afford (thus making this WM–SLA enterprise a genuinely interdisciplinary endeavour). Building on these key tenets of the integrated perspective on WM–SLA, in the remainder of this book I reconceptualize the concept of WM in nuanced SLA research areas. To achieve this, I explore how the integrated perspective on WM and SLA can inform specific strands of the WM–SLA connection. To begin with, in the next chapter I focus on the intricate relationships between the various domains of WM and L2 acquisition and development, and the learning and processing of various L2 sub-skills (listening, reading, speaking, writing).
7 Working Memory in L2 Acquisition and Processing: The P/E Model
Starting with this chapter, I demonstrate that the integrated perspective on theorizing and measuring the WM construct can facilitate a reconceptualization of WM in specific SLA research areas. To achieve this goal, in this chapter I explore the possible links between WM and the L2 acquisition and development domains and sub-skills learning and processing. In the first three sections, I recapitulate the findings of current WM–SLA studies and single out the distinct and separate roles of PWM and EWM in specific domains of L2 acquisition and development. Then, based on these findings, in the fourth section I propose an integrated model, namely the phonological/executive (P/E) model, to accommodate these patterns. The summary section outlines the theoretical and methodological implications of this integrated P/E model for future WM–SLA research and concludes the chapter.
WM in the L2 Acquisitional and Developmental Domains WM in the acquisition of L2 lexis/vocabulary To recall, although previous reviews of the European WM–language studies (e.g. Baddeley, Gathercole and colleagues) have focused on the relationship between WM and L1 learning (especially the relationship between the phonological loop and L1 vocabulary learning, as reviewed in Chapter 4), researchers have also investigated the relationship between WM and L2 learning (e.g. Atkins & Baddeley, 1998; see Baddeley, 2015). For example, Gathercole and colleagues (Gathercole & Thorn, 1998; Masoura & Gathercole, 1999) postulated that WM, in particular PWM, may play an 91
92 Part 3: Toward an Integrated Perspective on Working Memory and SLA
even more important role in L2 vocabulary learning than in L1 vocabulary learning. To that effect, Gathercole and Thorn (1998) showed that PWM was strongly correlated with children’s vocabulary knowledge, with correlation coefficients ranging between 0.4 and 0.6, and the correlation seemed to be independent of non-verbal intelligence. In another study, Masoura and Gathercole (1999) tested children’s PWM (as measured by the non-word repetition span task) and L1 (Greek) and L2 (English) vocabulary knowledge. They found that PWM was correlated with both L1 and L2 vocabulary knowledge and the correlations were independent of general factors (non-verbal ability, age). However, a more interesting finding was that although PWM was correlated with L2 vocabulary knowledge independent of native vocabulary knowledge, it did not correlate with L1 vocabulary knowledge when the participants’ L2 vocabulary scores were partialled out. This study, to some extent, also lends support to previous assumptions of language specificity in the memory of non-words, which is the language independence hypothesis (Osaka & Osaka, 1992; Osaka et al., 1993; also see Van den Noort et al., 2006). Indeed, the link between PWM and new word learning has been naturally extended to L2 learning. For example, in studies conducted by Service and colleagues (Service & Craik, 1993; Service & Kohonen, 1995; Speidel, 1993), Finnish primary school pupils’ PWM, as measured by the digit span and non-word repetition span tasks, was found to strongly predict English language learning success as represented by the children’s grades in English. In another study, Cheung (1996) pointed to the positive relationship between PWM (measured by a non-word repetition span task) and L2 vocabulary learning (as expressed by the number of trials the participants needed to learn some difficult English words) among 12-year-old Chinese (Cantonese) high school students. The author found that PWM uniquely predicted the number of L2 vocabulary-learning trials, thus signalling its overall involvement in L2 vocabulary learning among the 84 participants. More interestingly, the effect of PWM was significant for learners with a small vocabulary, suggesting a shift, with increasing proficiency, from dependency on PWM to dependency on long-term knowledge for vocabulary acquisition. That is, long-term knowledge mediates the learning of new words for high-proficiency L2 learners. Most of these studies involved children learning L2 vocabulary; however, other studies suggest that this relationship may also exist among adults. For example, Atkins and Baddeley (1998) found that the participants’ phonological WM (as measured by the non-word repetition span) was significantly correlated with the rate of their vocabulary learning. Papagno and Vallar (1995) compared Italian university students who were ‘polyglots’ (i.e. spoke at least two foreign languages) and students who were ‘non-polyglots’ and both groups were controlled for other variables, such as academic achievement and non-verbal intelligence. They found that polyglots outperformed
Working Memory in L2 Acquisition and Processing: The P/E Model 93
non-polyglots both in tests of PWM and in learning Russian–Italian word pairs. However, when learning Italian word pairs, these two groups performed at similar levels. In a later study, Speciale et al. (2004) explored the relationship between PWM (coupled with sequence learning) and vocabulary learning among undergraduate university students by conducting two experiments. Their results suggested that both variables (PWM and phonological sequence learning) contributed to vocabulary learning. Williams and Lovatt (2003) confirmed this significant correlation between PWM and vocabulary learning among university undergraduates learning a semi-artificial microlanguage (resembling Italian).
WM in the acquisition of L2 formulaic sequences and collocations As briefly discussed in Part 1 of this book, within the connectionist and construction-based views of SLA, an important role has been attributed to WM in relation to its functions in the acquisition of formulaic sequences (N.C. Ellis, 1996, 2012). Most notably, Nick Ellis argued that language users make use of ‘ready-made’ linguistic units or chunks (i.e. formulaic sequences), that is, frequently used multi-word units retrieved from the long-term knowledge base, for greater efficiency and speed in language comprehension and production. Therefore, it can be hypothesized that access to such prefabricated units in PWM frees L2 users to attend to other demands during communication, particularly under conditions where the WM is overloaded and there are few if any attentional resources available for linguistic analysis (Skehan, 2014, 2015a, 2016). Indeed, several recent studies have provided preliminary results that reinforce the tentative link between PWM and formulaic sequences in terms of collocational proficiency (e.g. Skrzypek, 2009) and native-like selection (Bolibaugh & Foster, 2013; Foster et al., 2014). Skrzypek (2009), for example, reported that the participants’ PWM was related to their collocational accuracy and development. Recently, Foster et al. (2014) reported positive effects of PWM in predicting both the rate of acquiring and the ultimate attainment of formulaic sequences, although these differences were related to the learning context and the age of the bilingual learner. Given these encouraging results, more empirical studies can be expected to explore other facets of the close connection between WM and the use of formulaic sequences as conceived by Nick Ellis (1996, 2012).
WM in acquisition of L2 morphosyntax/grammar An increasing number of WM–SLA studies have suggested that a close relationship exists between PWM and L2 morphosyntactic constructions (N.C. Ellis, 1996, 2012; French & O’Brien, 2008; Martin & Ellis, 2012; O’Brien et al., 2006; Sanz et al., 2014; Serafini & Sanz, 2015; Verhagen & Leseman,
94 Part 3: Toward an Integrated Perspective on Working Memory and SLA
2016; Williams, 1999; Williams & Lovatt, 2003; Wright, 2009). Previously, with respect to WM and L1 grammar, Adams and Willis (2001) believed that PWM was not implicated in native grammar learning, and postulated that L1 grammar learning was mainly implicit (particularly among children), and therefore may not involve explicit learning mechanisms such as WM. Recent views on language acquisition, such as the connectionist, emergentist and construction-oriented perspectives, have taken a different perspective on grammar that is anathema to the traditional view (e.g. Chomsky’s mainstream generative grammar, or MGG). More relevantly, Nick Ellis (1996, 2012, 2013) suggested that PWM is responsible for storing the sequential information that underlies the acquisition of linguistic sequences/chunks. According to this view, PWM is assumed to play an instrumental role in the acquisition of all sorts of lin guistic sequences, including lexis, formulaic sequences and morphosyntactic constructions. This connectionist approach blurs the distinctive concepts of vocabulary and grammar as adopted in MGG. In a laboratory study, N.C. Ellis and Sinclair (1996) presented experimental evidence showing that the articulatory rehearsal mechanism of PWM (sometimes referred to as verbal WM) is beneficial for English-speaking learners in acquiring meta-linguistic knowledge of Welsh grammar (a morphological mutation rule) and for producing more accurate Welsh morphological inflections. In a later study, Ellis and Schmidt (1997) found that learning a local agreement rule is also mediated by the participants’ PWM. The assumption of a close link between PWM and morphosyntactic constructions was also borne out in a semi-artificial grammar study by Williams and Lovatt (2003). In this regard, the authors managed to demonstrate that PWM was related not only to vocabulary learning, but also to the ability to remember sequences of familiar morphemes that are necessary for supporting subsequent distribution-based generalizations (of determiner–noun combinations) in grammar learning (of a semi-artificial micro-language resembling Italian). This finding is particularly important, as it suggests that the involvement of PWM in grammar learning is not qualitatively different from that reported in vocabulary learning, at least in situations where the grammatical material is unfamiliar to the learner (cited from French, 2006: 129). These results were replicated in a longitudinal study by French and O’Brien (2008), who found that PWM (as measured by Arabic and English non-word repetition tasks) significantly predicted the L2 grammar development of morphosyntactic structures (accounting for 28% of the variance) in addition to the contribution made by vocabulary knowledge (9.5%). This finding also suggests that PWM makes a unique contribution to L2 grammar learning and development that is unmediated by L2 vocabulary knowledge. In a recent study, Verhagen and Leseman (2016) addressed these issues more directly and produced very interesting results. The authors investigated
Working Memory in L2 Acquisition and Processing: The P/E Model 95
whether the phonological short-term storage component of WM (PWM – or ‘verbal short-term memory’ in their terms) is related to vocabulary learning, and whether the articulatory rehearsal of PWM is related to grammar learning among five-year-old children L2 learners (n = 63 Turkish children learning Dutch) as opposed to monolingual children (n = 45 Dutch L1) in a naturalistic learning environment. The participants completed a series of PWM tasks that measured their phonological short-term store and rehearsal mechanism, respectively a Dutch vocabulary task and a Dutch grammar task (in which two types of L2 grammar skills, morphology and syntax, were examined). Their structural equation modelling (SEM) results showed that the phonological short-term store significantly predicted both vocabulary and grammar, while the rehearsal mechanism predicted only grammar. Moreover, both functions of PWM (store and rehearsal) were found to be significantly related to the acquisition of morphology and syntax, irrespective of the bilingual or monolingual group. These results suggest that both aspects of PWM play a distinctive role in language learning and further corroborate Nick Ellis’s view (1996, 2012) that it plays an instrumental role for children learning L1 and L2 vocabulary and grammar (in a naturalistic setting).
WM in L2 development and L2 proficiency Although most WM–SLA studies adopted a cross-sectional design to investigate the effects of the two WM constructs (PWM and EWM) on specific areas of L2 learning, some SLA researchers have recently extended the pioneering research in educational psychology (Service, 1992; Service & Kohonen, 1995; Speidel, 1993) to explore the role of WM in the longitudinal development of L2 sub-skills. These researchers have targeted the relatively longer-term (five months to two years) effects of WM on L2 activities such as oral production, listening comprehension, reading comprehension and written production. French (2006), for example, investigated the developmental relationship between PWM (measured by English and Arabic non-word repetition span tasks) and L2 proficiency (expressed as composite scores on reading and listening comprehension, receptive and productive vocabulary and grammar items) among grade-6 Francophone children (n = 54) enrolled in a five-month intensive English programme in the Saguenay region of Quebec, Canada. The results of the partial correlations (controlling for non-verbal ability scores) indicated that PWM not only significantly predicted the overall L2 proficiency at both periods (time 1, the first week of the first month; and time 2, the last week of the fifth month), but the magnitude of the relationship also remained similar over time for the whole group (n = 54) as well as when the group was divided according to level of L2 proficiency (28 high-proficiency participants and 26 low-proficiency participants).
96 Part 3: Toward an Integrated Perspective on Working Memory and SLA
Another interesting finding concerned the predictive value of PWM on the absolute increase in L2 proficiency between the two periods. Specifically, a significant correlation was detected only for the low-proficiency group (r = 0.76, p