Modularity,working memory,and second language acquisition: A research program

1 The Baddeley and Hitch model

The Baddeley and Hitch model initially hypothesized three components: a visuospatial sketchpad, a phonological loop, and a central executive directing attention and providing storage of a nebulous sort. Baddeley subsequently reduced the executive to its attention function and added an episodic buffer as a place where everything comes together to form a coherent conscious episode. My concern is with the sketchpad and the loop.

The sketchpad is not a genuine entity. At one point, Baddeley and colleagues described ‘visuospatial’ as a ‘blanket term’ covering any short-term memory that is not verbal (Della Sala et al., 1999). He concluded long ago that visual and spatial WM are distinct in terms of both storage and the processes that manipulate the stored representations (see Baddeley, 2007); in other words, two distinct modules are involved. More recently, he tentatively incorporated haptic (including kinesthetic and tactile) information in the ‘visuospatial sketchpad’, but research indicates that this is in fact a distinct WM domain (e.g. Katus et al., 2015; Pasternak and Greenlee, 2005).

The notion of a unified ‘phonological loop’ is also problematic. Considerable evidence exists that we must posit both a phonological store and an auditory store to explain the phenomena for which the loop was hypothesized (see especially Cowan, 2005). Penney’s (1989) review emphasized the observed importance of modality effects in short-term recall of verbal stimuli: Visually and auditorily presented information are both encoded phonologically, but recall is strongly influenced by the original modality.

Many domains are left out of this picture. Recently, Baddeley (2007, 2012; Baddeley et al., 2011) tentatively incorporated music, lip-reading, sign language, and ‘sound’ in the loop (in addition to haptic in the sketchpad), but he was very tentative and offered little or no evidence that a shared WM exists.

These considerations all point to a more fractionated WM, including at least visual, auditory, haptic, spatial, and phonological, and possibly a great many more. In other words, they point in the direction of a genuinely modular conception of WM.

2 State-based theories of WM

State-based theories reject the idea of WM as a location or a store, in favor of an activation approach: Working memory is the set of items in long-term memory (LTM) that are currently active. The central theme for research is thus the nature of the activation, with attention taken as its primary source. These theories abandon the distinctions among the various components of the Baddeley and Hitch model and are thus in a sense even less modular. But they are actually very amenable to a modular interpretation. Cowan et al. (2014) concluded from their experiments that, contrary to their previous view, most of working memory capacity (WMC) is peripheral; i.e. domain-specific, available only for processing within a given domain. This is a large step in the direction of a modular view of WM.

1 In general

Modularity can mean different things, but the central idea is functional specificity (Barrett and Kurzban, 2006; Coltheart, 1999; Pinker 1997). If distinct mental functions are carried out by distinct units, the mind is modular. The logic of a modular design is that a system composed of specialist sub-systems can operate more accurately and efficiently than a general-purpose system, and that damage or change to one function has limited effect on others. Carruthers (2006) argued, for these and other reasons, that a modular design makes more sense from an evolutionary perspective than does a more homogeneous system.

Some clarifications are needed. First, modularity does not imply absence of interaction, just constrained interaction. Second, neural plasticity is not problematic. The innate portion of a module includes only the basic processor and the primitives it works with. The overwhelming majority of a module’s neural range is likely to be connections among these primitives, constituting new representations formed through experience, guided by the processor. Finally, modularity does not assume a one-to-one mapping between genes and features of the module; the brain’s development is far more complex and interesting (Marcus, 2004). One question that I will leave open for the general case (see below for language) is how many modules exist and how fine-grained the modularity is.

The modular approach is closely related to the popular dual-processes idea (Evans, 2008; Kahneman, 2011). On this view, the mind includes two kinds of processes: specialist (S1) and non-specialist (S2), the latter able to perform much the same functions as the former but typically less efficiently and successfully. Workings of the non-specialists, but not the specialists, are strongly associated with consciousness. The specialists can be identified with modules while the non-specialists can be understood either as a general, non-modular system or as interactions among modules.

Much of the motivation for a modular approach comes from cognitive neuroscience, where modularity represents standard thinking (e.g. Bertolero et al., 2015; Dehaene, 2011; Gazzaniga, 2011; but see also Barrett and Satpute, 2013). Bertolero et al. produced extensive evidence for a close correspondence between function and location in the brain and cited genetic evidence consistent with this conclusion. Neural research has found that LTM is functionally fragmented, its visual aspects in visual areas, motor aspects in motor areas, etc. (e.g. Fuster, 2009; Jonides et al., 2008; Rissman and Wagner, 2012). If WM is activation of LTM representations, it should show a parallel fragmentation: it should be modular. And neural research suggests that it is. WM tasks involving a given function specifically activate areas associated with that function (e.g. D’Esposito and Postle, 2015; Druzgal and D’Esposito, 2001; Zaksas et al., 2001).

2 Modularity and language

The modularity of language is highly controversial (critics include, amongst others, Fedorenko et al., 2006; Goldberg, 2009). The case for modularity starts with the poverty of the stimulus: Given the extreme complexity of natural language, the limited and sometimes misleading input, and the generally limited learning ability of small children, the universal success of acquisition is mysterious – unless we hypothesize a substantial innate foundation, universal grammar (UG). Children’s success is more striking in that it occurs in widely differing circumstances, involving both cultural differences and various conditions that should be expected to get in the way (see Bishop and Mogford, 1988), and even in children whose general learning ability (intelligence) is extremely weak (e.g. Flavell et al., 1993; Smith and Tsimpli, 1995). Further evidence comes from deaf children’s acquisition of sign languages, which shows striking parallels with acquisition of spoken language (Goldin-Meadow, 2003; Lust, 2006). In extreme cases children receive input from an impoverished version of the language, due to limitations of the speakers or signers around them, and go beyond this input, developing a full grammatical system. The success enjoyed by nearly all children under nearly all conditions contrasts with the problems experienced by a small population of otherwise normal children, even under the most favorable conditions: problems that have a clear genetic basis (Bishop, 2006). All these considerations point to an innate, domain-specific basis for language, UG.

This leaves the question of whether UG is involved in SLA. The standard argument for non-involvement is that learners typically have limited success. But this observation can be explained in any number of ways, including such factors as motivation, context, and cognitive maturity, and interactions among them. Perhaps more importantly, the existence of an established language inevitably influences processing and therefore learning of the L2. The high level of success achieved by many learners argues against any fundamental change from first to second language acquisition. Continuity is also the more parsimonious hypothesis.

1 Studying modularity

Studying modularity means first identifying the modules that make up the mind. Little disagreement exists regarding the modular character of the various sensory modalities. The modularity of spatial functions also faces little controversy, at least in the context of working memory. Other functions, including language, are more controversial. For understanding modularity specifically in relation to language, Jackendoff (1997, 2002) is probably the best starting point.

One general criterion for judging modularity is evolutionary justification: A proposed module should serve a function that is significant in natural selection and could plausibly have evolved. The extensive literature in evolutionary psychology (e.g. Cosmides and Tooby, 2013; Pinker, 1997) has much to say on these matters. Behavioral evidence plays a key role as well, particularly dual-task research looking at degree of interference between tasks carried out by two putative modules. No less important is neural work, identifying brain locations that are especially active in particular types of tasks.

2 Studying modular WMs

Within a modular conception of the mind, the function of making information available to cognitive processes implies that a distinct WM exists for each module, and this must be taken into consideration in empirical research and its interpretation. For the case of language, this includes modules that are specifically linguistic and others that process language-related information. An important aspect of a research program is thus study of the individual WMs for each of these modules, along with the interactions among them.

Probably the most important tool for this job is dual-task research, more or less as typically carried out but with a broader scope, recognizing all the modules that are involved. The published literature offers some highly sophisticated methodology (e.g. Cowan et al., 2014), which is fully applicable to modular SLA research. The main adjustment is that task design and selection should be based on hypothesized modules, with interpretation of results following. A ‘pure’ task relying on just one WM is unlikely, so comparison between results of related tasks is essential. The load on each putative WM can also be varied and resulting effects on performance observed, again as in familiar WM research. A challenge for task selection is that a single task can often be carried out in more than one way, as described above.

Neural research is also essential, looking for activation patterns occurring during WM tasks in regions that have been identified with particular modular functions. A crucial cautionary note for neural research and its interpretation is that there can be no simple equation of cognitive activation (WM) with neural activation. The former is simply the availability of a representation to its processor. Neurally, this availability can be in the form of sustained activation but can also be a transient change in relevant synapses, establishing a state of increased readiness for activation: a ‘quiescent working memory’ (Fusi, 2008; Mongillo et al., 2008). Thus, the absence of sustained (neural) activation does not imply that a representation is not cognitively active or is not ‘in working memory’. Here again task selection is crucial, and potentially challenging.

3 Studying global WM

Global WM is, again, essentially what is normally treated as WM. Distinguishing it from local WMs is essentially the same issue as distinguishing central and peripheral capacities, a prominent theme in state-based WM research (e.g. Cowan et al., 2014). The sophisticated methodology developed there is thus available for study of global working memory (GWM). But it should be applied with awareness of the distinct roles played by the various modules, including perceptual, phonological, syntactic, and conceptual. One important caution comes from the observation, based on Carroll (1999), that the input to learning (the content of GWM) is not a sentence or an utterance but rather a perceptual representation.

Neurally, GWM is realized as broad patterns of synchronized activity over wide areas of the brain. Study of these patterns is again a common part of current WM research and so the methodology and interpretation tools are readily available. In the modular framework these broad patterns should include local patterns representing activity of the individual modules. This then is another way to approach the issues of local and global WM.

Study of individual differences in working memory capacity, as they are commonly studied, is primarily about global WM, as measurement of WMC involves deliberate, explicit tasks. But the role of individual modules and their WMs should be considered. Whether individual differences exist in linguistic WMs is unclear and constitutes another worthwhile research topic.

In state-based theories, attention is usually taken as the source of the domain-general activation that constitutes WM. So the vast literature on the topic is useful here. The concept of attention has its shortcomings, though (see Sharwood Smith and Truscott, 2014), so it may be better to focus directly on elements that determine what is attended: value, goals, context (as internally represented), and sensory input (i.e. active representations in the perceptual modules).

4 Studying the two types of linguistic knowledge and processing

The proposed WM research program connects to one of the most important and difficult topics in SLA, and beyond: the long-recognized distinction between two types of linguistic knowledge and use (e.g. R Ellis, 1988; Krashen, 1981; Lightbown, 1985; Long, 1977; Schwartz, 1986), its most popular current form being the implicit–explicit distinction (e.g. N Ellis, 2005). In each case the primary type is unconscious and automatic, the other largely conscious and playing a supporting role (compare the specialist–nonspecialist contrast).

The modular framework offers a useful way to understand and study this crucial distinction. The linguistic modules constitute the primary seat of (unconscious) linguistic knowledge, while the more conscious variety is conceptual, i.e. found in the conceptual module (see Truscott, 2015a, 2015b). We should expect the work of specialist modules performing their own function to be automatic; if knowledge outside these modules is to influence language-related activity, this use is likely to require GWM involvement. The role of local and global here can be studied, again, with familiar WM methods, notably dual-task studies and neural research. The main issue is task design.

Addressing this issue means recognizing the wealth of relevant findings in the SLA literature. The most important but perhaps least recognized finding of Norris and Ortega’s (2000) seminal meta-analysis on the effects of formal instruction was that very different results come from measurements of declarative (conceptual) linguistic knowledge and measurements that discourage the use of such knowledge, measuring instead the unconscious workings of the linguistic modules. This point can also be seen in meta-analytic reviews on the effects of error correction (e.g. Kang and Han, 2015; Li, 2010; Truscott, 2007; for discussion, see Truscott, 2016). These observations provide useful guidelines for task design. Their limits must be recognized, though. Many types of metalinguistic tasks can be carried out by intuitions, possibly based on knowledge in the linguistic modules. Conceptual knowledge can be used to some extent in spontaneous production, in the form of either deliberate focus on explicit knowledge or use of automatized conceptual knowledge.

This last point raises another significant research topic within the modular approach. While the distinction between modular and conceptual linguistic knowledge is closely associated with the standard implicit–explicit distinction, the two cannot be equated. Truscott (2015a, 2015b) offered an account of the distinction between implicit and explicit (and subliminal) in a modular perspective, providing a basis for study of this issue. The hypothesis is that knowledge/processing in the specialized linguistic modules is entirely implicit while its conceptual counterpart is typically explicit, consistent with Williams’ (2012) observation that SLA research has typically found explicit but not implicit knowledge and learning to be associated with WMC (which, again, is largely about GWM). This hypothesis also fits well with findings that the strongest relation between WMC and success occurs in the early stages of learning, especially in formal classroom learning (Gathercole, 2006; Hummel, 2009; Masoura and Gathercole, 2005; Winke, 2005). This association is readily explained by the use of conceptual knowledge, and therefore GWM, as a substitute for poorly developed modular knowledge in the early stages.

Probably the most important implication of this discussion is that we have to recognize, again, that two distinct kinds of processes are at work in language processing. In practical terms, this means that research requires a more refined and more theoretically based notion of task type than is often assumed and, no less important, greater concern with the implications that are to be drawn from experimental findings.