Working Memory in Second Language Learning

Alan Baddeley

Introduction

The multi-component model of working memory (WM) was developed with the dual aims of providing a framework for the basic understanding of human memory, and at the same time providing a bridge to application beyond the laboratory. One of the earliest and, in my view, most successful of such applications was to language learning, principally to the acquisition and development of vocabulary in children (Baddeley et al., 1998), and also to second language learning (Atkins & Baddeley, 1998). As is clear from the current volume, research on WM and second language learning has flourished in the years since my own rather minimal involvement, and hence I was delighted to hear of the present enterprise, and to lend my support through writing this introductory chapter to express my speculative thoughts on the possible implications of the multi-component view of WM for second language learning.

While it would have been very nice to read the contributions and subsequently comment on them, the timing clashed with the process of revising and updating our memory text (Baddeley et al., 2015). My initial agreement was to write a Preface that envisaged little more than enthusiastic support of this particular application of the WM model. I was, however, tempted by the editor's suggestion that I might outline the current state of the multi-component model, and in doing so thought it would be interesting to think about second language learning and speculate as to how the various current components of the model might be involved. I should emphasise, however, that what follows is simply a brief account of the current model (a more extended discussion can be found in Baddeley [2012]), together with some initial thoughts that may well be proved misguided by the chapters that follow. If so, I shall simply fall back on the observation that progress is often made by discovering points where our predictions are clearly wrong.

Evolution of the Multi-Component Model

Our WM model has its roots in the proposal that short-term memory (STM), principally verbally based, played a useful role in cognition more generally, a view first presented by Broadbent (1958) and developed much more extensively by Atkinson and Shiffrin (1968). Our own work (Baddeley & Hitch, 1974) was prompted by difficulties encountered by this approach in dealing both with links to long-term memory (LTM) and with data from patients whose impaired verbal STM had surprisingly little impact on their broader cognition. Our own studies, relying heavily on dual task methods, led us to propose the model shown in Figure 1.1, which assumes an attentionally limited control system, the central executive, aided by two temporary storage systems, one specialised for acoustic and language stimuli, and the other counterpart the visuospatial sketchpad.

Further development was spurred by the challenge of finding an evolutionary function for the phonological loop that might prove more plausible than remembering telephone numbers. Our search was helped by access to a patient, PV, with a very pure phonological loop deficit. We first tested the hypothesis that the loop was necessary for language comprehension, but found little evidence of impairment, except for certain convoluted sentences explicitly designed to rely on the phonological loop (Vallar & Baddeley, 1987). We then tested the hypothesis that the phonological loop might have evolved for language acquisition, finding that our patient was greatly impaired in acquiring vocabulary in a foreign language, Russian (Baddeley et al., 1988). We went on to show that procedures interfering with the phonological loop function disrupt the acquisition of foreign language vocabulary but not memory for meaningful word pairs in healthy participants (Papagno et al., 1991), and that polyglots who have acquired several languages tend to show enhanced phonological loop capacity (Papagno & Vallar, 1995), while otherwise extremely able people with a reduced verbal STM also tend to have difficulty in acquiring foreign vocabulary (Baddeley, 1993). Furthermore, extensive research by Susan Gathercole and myself has shown a clear link between native language vocabulary learning in children and phonological memory (see Baddeley et al. [1998] for a more detailed account).

As a result of this line of work, we modified the basic model to that shown in Figure 1.2. The principal difference is a clear link between the phonological loop and phonological LTM, a link that operates in both directions; the phonological loop facilitates the acquisition of new words, and in due course the richer the available array of existing words, the easier it is to use these to help acquire new items. We speculate that a similar link will occur between the sketchpad and visuospatial semantics, although this has been little investigated so far.

The next version of the model was also driven by the study of language and, in particular, the very strong relationship between prose comprehension and the WM span measure originally developed by Daneman and Carpenter (1980). They presented their participants with sequences of sentences to read, subsequently asking them to recall the last word of each. This and other complex span measures, not necessarily involving language, have proved to be very powerful predictors not only of comprehension (Daneman & Merike, 1996), but also of cognitive processing more generally, including measures of intelligence (see Engle et al. [1999] for a review).

Daneman and Carpenter explicitly based their test on the requirement to combine the storage and processing of information, as in our original model. We were therefore delighted that it proved so successful, but worried by the implications it had for our model, as it existed at the time. This concerned our assumptions regarding the central executive. Initially, it had been assumed to be capable of combining both storage and attentional processing, an assumption that was so broad and unconstrained that it effectively became a homunculus, the little man who could do everything that might create problems for the model. This presented a problem. Homunculi are good servants but bad masters; if our homunculus was able to do everything, how could it be tested and how could the theory be developed? Homunculi can, however, be useful if constrained by specifying exactly what tasks they are to perform, and then attempting to explain each of these in turn. When all are explained, the homunculus can be retired.

We began by removing the storage role of the executive, assuming that it is a purely attentional system (Baddeley & Logie, 1999). Unfortunately, this assumption created difficulty in trying to explain just how our model provided the substantial storage needed by the Daneman and Carpenter (1980) WM span task, since neither the phonological loop nor the sketchpad has enough storage capacity to hold entire sentences. Faced with this dilemma, I finally added a fourth component, the episodic buffer (Baddeley, 2000). This is assumed to be a multidimensional storage system, capable of combining information from the visuospatial and verbal subsystems and linking it with further information from perception and LTM. It was assumed to hold a limited number of episodes (cf. Cowan, 2005), with each episode binding together information from these various sources into unified chunks. Finally, the buffer was assumed to be accessible to conscious awareness, and indeed to be the basis of conscious awareness which, like Baars (2002), I proposed serves the specific function of binding together information into consciously experienced arrays or episodes. A problem with this ambitious speculation is the danger that it simply becomes another homunculus, all powerful but untestable. The next stage in collaboration with Graham Hitch and Richard Allen has been to demonstrate that this is not the case.

The 2000 model was explicitly designed with the assumption that the episodic buffer was under the strict control of the central executive, just a short but important step from our original all-purpose homunculus. We chose to focus on the process of binding features into the chunks stored in the buffer, using our dual task methods to block the various components of the model, and in particular the central executive. Our initial speculative model predicted that without the strong support of the executive, such binding would break down. We looked at both the binding of colours and shapes into visual objects, and also at the binding of words into the chunks that occur in remembering meaningful sentences. This is a very powerful effect, with memory span for unrelated words being around 5, and for meaningful sentences around 15 words. We argued that if the executive is needed in order to perform this binding, then a demanding attentional task should interfere with binding and hence reduce the sentence advantage. This did not happen; an attentional task like concurrently counting backwards reduced overall performance, but had no impact on the difference between recall of unrelated words and sentences. We concluded that the advantage from the semantic and syntactic constraints of sentences came from LTM and was relatively automatic (Baddeley et al., 2009b). An exactly equivalent story came from our study of the binding of features into objects (Allen et al., 2006). We concluded that the episodic buffer is an important but passive store, capable of holding bound episodes, but not itself performing a binding function. We suggest that binding is likely to happen in different ways for language and vision, using different systems at different brain locations. It appears to function as a passive store rather than an attentionally dependent component of WM.

Our current, now rather more complex model is shown in Figure 1.3. It still has the three original basic components, with the addition of a fourth, the episodic buffer. The phonological and visuospatial subsystems feed into the buffer, and are themselves envisaged as bringing together information from a range of contributory sources. In the case of the sketchpad, these involve spatial location, colour and shape together with information from the complex systems underpinning touch and kinesthesis. The phonological loop is seen as bringing together information not only from speech, but also from language-related sources such as lip read and signed information, together with access from non-verbal sounds. Note that the loop is atypical in having a specific and very effective process for rehearsal via vocal and subvocal articulation. Rehearsal more generally within the system is assumed to be equivalent to Johnson's concept of refreshing (Park et al., 2010), a process whereby focusing attention on an item within the episodic buffer allows continuous maintenance. This is more attentionally demanding than subvocal rehearsal, which is why we tend to use verbal rehearsal when possible.

An important issue remains however, that of how WM relates to LTM. Our current view is expressed in Figure 1.4 in which WM is seen as providing an interface between cognition and action. Our representation of cognition is intentionally kept general since the information entering WM can come from a range of different sources which themselves interact. A person in a digit span experiment for example will gain from his or her long-term knowledge of digit names; his or her span would be substantially less for digits in an unfamiliar language such as Finnish. In the case of vision, we tend to perceive the world in terms of meaningful objects, a process that itself depends on LTM. However, the function of all of this activity is to allow us to interact with the world, to take in information and to act upon it.

One final issue concerns the relationship between our version of WM and that proposed by Nelson Cowan whose chapter follows. Our theoretical views are frequently seen as in direct opposition; we focus on temporary storage and manipulation, while Nelson assumes a major role for activated LTM. However, as I think Nelson will agree, the differences are more apparent than real. His is a top-down approach, based initially on his interest in attention and its limited capacity. Our own approach has been essentially bottom-up, starting with memory span and only later worrying about the links with LTM. As Figure 1.4 shows, we are entirely comfortable with necessary links between WM and LTM, but regard the concept of activated LTM as a general term for a wide and complex range of processes. Such a concept acts as a place marker, allowing Nelson to focus on his principal interest, the role of attentional control. I see this as somewhat analogous to our use of the relatively broad and unspecified concept of the central executive, allowing us to focus more explicitly on the more tractable subsystems. As our models have developed, I feel they have come closer together, and would see Nelson's model within our own terms as focusing on the interface between the executive and the episodic buffer.

Do we differ? I think we do in that we assume that information is downloaded from LTM into the episodic buffer, whereas Nelson favours a system with pointers to the relevant LTM locations. We can neither of us currently think of a good way of resolving this, and whichever way it goes would not seriously challenge either of our broad theoretical frameworks. That does not mean, of course, that we will always agree on the best way of interpreting our data or choosing the next question to study, but it does mean that our approaches are sufficiently compatible to be mutually informative.

Working Memory and Second Language Learning

An important feature of the multi-component model is that it should be readily applicable to cognition beyond the bounds of the psychology laboratory. The learning of both first and second languages is an obvious area of application that has continued to interest me over the years. However, I cannot claim to have kept up with the large and expanding literature reflected in the chapters that follow. Hence, I am not able to provide an overview, but have instead agreed to offer some speculations as to how I might expect the current version of the model to apply to the learning of a second language.

It seems likely that different components of the model will be relevant to different aspects of the task of language learning, so I will go through the various subcomponents in turn and speculate as to where and how they might potentially be relevant.

The most obvious and I believe well established is the role of the phonological loop. We have argued that the loop plays a significant role in native language acquisition (Baddeley et al., 1998) with much of the evidence for this claim coming, as described earlier, from studies in which participants attempt to learn new items typically either foreign language vocabulary or non-words. Other supportive evidence has come from correlational studies based on second language learning in both children (Service, 1992) and adults (Atkins & Baddeley, 1998), with more recent evidence coming from the study of children participating in programmes involving an immersive language approach (Engel de Abreu & Gathercole, 2012; Nicolay & Poncelet, 2013).

All of this points to a link between the phonological loop and verbal LTM for language, a link that goes in both directions. Not only does the capacity of the phonological loop influence the rate of vocabulary acquisition, but also, 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).

In recent years, there has been considerable theoretical development in providing detailed models of how serial order is maintained in verbal STM (see Hurlstone et al. [2014] for a review). This in turn has encouraged theorists to explore ways in which information appears to be transferred from the phonological loop into LTM. Their tool of choice has been the Hebb Effect. This stems from the demonstration by Hebb (1961) of a long-term component in the standard serial verbal recall task. He presented participants with sequences of nine random digits for immediate recall. Each sequence appeared to be different, but in fact every third string was an exact repetition. Performance on these gradually improved, with no difference being found between people who noticed the repetition and those who did not. Melton (1963) interpreted this result as implying that it was only necessary to assume one type of memory, LTM; others disagreed, proposing that the task itself contained both STM and LTM components. Recent studies have supported this claim, demonstrating that factors influencing immediate verbal recall, such as articulatory suppression and phonological similarity, do not influence the Hebb Effect (Hebb, 1961). This is currently a very lively area that is likely to have long-term implications for implicit aspects of second language learning (see Page et al. [2013] for a recent survey of this area).