2.1 Robot interaction

This first version of interaction emanates from monumental efforts in artificial intelligence to develop robots, or chatbots, which emulate human interaction. The Holy Grail in these efforts is a chatbot that passes the Turing Test – that is, which leads human beings to believe they're communicating with another human. Without further comment, then, let me show you a short video of two chatbots conversing with each other – not exactly what they're designed for, as just noted, but still a fairly good indicator of what kind of interactors they are. As you watch the video, I'd ask you simply to note what, in your judgment, is particularly humanlike and particularly nonhumanlike in this interaction. The video can be found at: www.youtube.com/watch?v=WnzlbyTZsQY- please watch it before continuing to read. (For further information about this interaction, see www.npr.org/2011/09/01/140124824/robot-to-robot-chat-yields-curious-conversation and http://creativemachines.cornell.edu/AI-vs-AI.)

Now that you've watched the video, here's my own comment: One thing that strikes me about this interaction is its largely non-cooperative character. The two participants – or should we call them contestants? – seem to be playing the linguistic version of competitive (versus cooperative) ping pong: They hit the ball at each other, but not in ways designed to make it easy to return. As cooperative conversationalists, then, these chatbots don't seem wildly successful.

2.2 Interaction in mainstream SLA

The second version of interaction I'd like to briefly describe is the one taken by the interactionist – or input and interaction – school of SLA, regarded by many as the single most influential SLA approach currently on offer (Block Reference Block2003). This approach is highly cognitivist in that it treats the brain as an information-processing device, or cognitive computer. Here, interaction is viewed as a form of input pre-processing – through input ‘modification’ or ‘negotiation’ via interaction, input becomes comprehensible and therefore available for further processing by the cognitive computer. This view also underlies the current focus on error feedback/correction in instructed SLA.

Gass's (Reference Gass1997: 3) influential model of the SLA process clearly indicates the role of interaction in this approach. Arranged vertically, the model has ‘input’ at its top and ‘output’ at its bottom, separated by dotted lines from the model's eight components arrayed sequentially in between. Of these, ‘Negotiation –> NS [i.e. Native Speaker] Modification’ is the only component that is not mind-internal, with the final component, ‘Personality, Mode (Oral/Written), Situation’, being a mixed case. Despite the centrality of the idea of interaction in this approach to SLA, then, it actually occupies a very small part of the overall model, the vast majority being mind-internal.

2.3 Interaction in a sociocognitive approach to SLA

It's not my intention to argue that the two versions of interaction just reviewed are wrong in any categorical sense; certainly, help from all quarters is needed in understanding something as complex as interaction or SLA. I would suggest, however, that the robot version seems to underestimate the depth and complexity of human interaction – two human beings designed by evolution (see below) to interact/talk with each other is not the same as two machines designed by humans to talk at each other – while the mainstream SLA version may undersell interaction in L2 development by giving it too limited a role.

The third version of interaction I want to discuss, then, is one that treats it more complexly, and one that can give it a more robust role in SLA. Luckily, human interaction – including but not limited to linguistic interaction – is a topic of serious interest among academics these days, especially those investigating the roots of human sociality (e.g. Goodwin Reference Goodwin2000, Reference Goodwin, Coupland and Gwin2003; Enfield & Levinson Reference Enfield and Levinson2006; Levinson Reference Levinson2006; Schegloff Reference Schegloff2006; Tomasello Reference Tomasello2008, Reference Kinsbourne and Jordan2009). I depend on their work in what follows.

There's little doubt that human interaction has been the ‘primordial site of [human] sociality’ (Schegloff Reference Schegloff2006: 70) over our species history, which began, by recent estimates, between two and eight million years ago. Schegloff notes that even as countless social structures (e.g. tribes, political and economic systems, and whole complex societies) have arisen and disappeared across human history, interaction as a social structure has continued largely unaffected. And until the advent of the telephone, at least, face-to-face interaction must have been the main means by which new tribes, political systems, and complex societies were established. I'd like to make three points here: First, if we've been engaging in close, usually face-to-face interaction for from two to eight million years, then we've probably become quite good at it – particularly if such interaction comprises a highly evolved, largely innate set of skills. In comparison, most of humanity has been interacting by telephone for 50 years at most, and by computer for 10 years at most (although much of humanity is probably still not online) – thinking in this way, we get a sense of just how highly developed face-to-face interaction may be. Of course, this is an unfair comparison since telephone and computer communication rely largely on the same interactive abilities developed through face-to-face interaction over the ages, so they're not separate things.

Second, our high-level interactive abilities are probably the ultimate skill set separating us from other animals. Based on extensive research comparing apes with human infants, for instance, Michael Tomasello (e.g. Reference Tomasello2008) has shown that even at three months infants display interactive abilities far outdistancing those of their primate cousins – abilities that can only be innate given how early they appear.

Third, language in this view is the crowning achievement of human interaction rather than its underlying basis – our sophisticated interactive abilities are the foundation on which language is built rather than vice-versa. We simply couldn't have developed symbolic communication without robust, preexisting interactive abilities, and language in use still relies on those abilities at every turn (Levinson Reference Levinson2006; Tomasello Reference Tomasello2008).

The picture of interaction I want to adopt here, then, presents it as the fundamental basis for human sociality, and that includes language learning. In the words of the pioneer sociologist Max Weber, ‘A social relationship may be said to exist when several people reciprocally adjust their behavior to each other with respect to the meaning. . .they give. . .it, and when this reciprocal adjustment determines the form. . .it takes’ (Reference Weber and Runciman1922/1978: 30). Compared to the chatbot conversation shown earlier, interaction in this sense is like a cooperative (versus competitive) ping pong game, in which two or more partners coordinate or align their activities sensitively and ongoingly for their mutual benefit. And they do so in at least three ways.

First, the participants align cognitively, as interactants become intercognizers. That is, they use their profound, evolutionarily well-developed skills of shared intentionality (Tomasello et al. Reference Tomasello, Carpenter, Call, Behne and Moll2005), theory of mind – i.e. moment-by-moment insight into what the other is thinking (Premack & Woodruff Reference Premack and Woodruff1978), common knowledge – knowing what the other is thinking about what you're thinking (Lewis Reference Lewis1969), and cross-cognitive anticipation (Kinsbourne & Jordan Reference Kinsbourne and Jordan2009) whereby fluent social activity is enabled to ‘go on’ without perceptible breaks. The combined result of such abilities has various names: shared cognition and intersubjectivity are two of them.

Second, partners in interaction align affectively, minimally in the sense that the activity itself is presumably mutually beneficial, but typically in much deeper ways. Thus, research across fields reveals the many public signs of affective and affiliative alignment in cooperative human interaction, from enhanced eye contact (e.g. Kendon Reference Kendon1990), to rhythmic entrainment of body movement (e.g. Collins Reference Collins2004), to mutual smiling (e.g. Chartrand & Bargh Reference Chartrand and Bargh1999), to linguistic accommodation (e.g. Ylanne-McEwen & Coupland Reference Ylanne-McEwen, Coupland and Spencer-Oatley2000). Once again, we don't just affectively attune ourselves to the other in an ongoing, highly sensitive way, but additionally engage in ‘attunement to the attunement of the other’ (Rommetveit Reference Rommetveit and Braten1998: 360). The discovery of the human mirror neuron system, in which we neurally enact animate actions we are merely witness to (Rizzolatti & Sinigaglia Reference Rizzolatti and Sinigaglia2008), may provide one powerful key to understanding how human beings align naturally in the affective domain. That is, I may literally ‘feel your pain’.

Third, and finally, interactants align with each other physically in moment-to-moment, highly sensitive ways, with variations in alignment signifying varying roles in and degrees of commitment to the interaction. The natural history of an interaction can even be traced in terms of physical movements, the original coming together, for example, being marked by mutual presentation of a maximum of physical symptoms through a literally face-to-face engagement, often including touch. From these beginnings, an interaction might adopt an ‘L-structure’ (Kendon Reference Kendon1990) wherein interactants stand at right angles to each other, thereby allowing themselves physical and perceptual access to both their interlocutor and what (and who) else might be happening (or present) in the environment. Alternatively, the full face-to-face orientation might be maintained. Finally, as interactants prepare to withdraw, they make subtle adjustments of proximity and orientation to signal the approach of an ending, although such endings are typically negotiated and contingent on the other's acquiescence, usually at a subconscious level. But the main point I want to prepare the ground for here is simply this: Interaction underlies and supports human social behavior in profound and complex ways, and if language learning is a form of social (or, preferably, sociocognitive) behavior, then interaction supports language learning as well.

3.1 What is cognition for?

In a sociocognitive approach to SLA, human cognition is reconceptualized as adaptive intelligence (A. Clark Reference Clark, Nuñez and Freeman1999; Atkinson Reference Atkinson2012) – as having the major evolved purpose of promoting adaptation to our ecosocial environment. The argument goes as follows: Organisms which don't adapt to their environments typically die, and rather quickly; for this reason, most motile zoological organisms have at least proto-nervous systems. Single-cell organisms like the amoeba, for example, while lacking articulated nervous systems, have sensory organs that allow them to detect, on the one hand, danger (e.g. for amoebae, noxious substances in their aqueous environment) and, on the other, affordances (e.g. food sources). More advanced, multicellular organisms have correspondingly more advanced environmental sensing/nervous systems, with the ‘top of the line’ in the animal kingdom being the human nervous system, which is centrally based on our highly developed brain. Brains are where our millions of neural pathways converge, leading, at least theoretically, to the combination of every bit of sensory information with every other bit. The overall result is fine-grained, intimate, ongoing attunement to the ecosocial environment, at least where needed for survival.

While few if any cognitivists would deny that the brain has evolved for adaptive purposes, the cognitivist program has pushed biological adaptivity to the margins in an effort to develop the engineering-centered view that cognition operates like a digital computer (e.g. Boden Reference Boden2006; Noë Reference Noë2009). Cognitivists, in other words, treat cognition as the logic-based processing of abstract symbols and building of abstract representations, from which environmental actions and solutions (if any – see Chomsky's (Reference Chomsky1965) influential restriction of linguistic theory to the realm of competence) are then computed. Luckily, however, the ground has been shifting in the cognitive sciences: Instead of disembodied computation, biological/adaptive approaches to cognition emphasize a different skill set – one focusing on environmental pattern finding based on the organism's deep and interactive embedding in its environment. A. Clark (Reference Clark1997) states the argument in humorously memorable form: Humans are ‘good at frisbee, bad at logic’ (p. 60), where ‘logic’ denotes the computational representation and solution of logic-like problems in a cognitivist model of cognition, while ‘frisbee’ denotes complex pattern-finding abilities originally associated with evolutionarily crucial survival skills like hunting. Significantly, computational modeling of such skills has proven difficult, and a number of alternative cognitive theories, all emphasizing the ways in which minds orient bodies to adaptively align with the world, have arisen. Examples include ‘motor resonance’ enabled by the human mirror neuron system (Rizzolatti & Sinigaglia Reference Rizzolatti and Sinigaglia2008); adaptive, intuitive ‘rules of thumb’ (Gigerenzer Reference Gigerenzer2007); low-dimensional environmental sensing which ‘use[s] the world as its own best model’ (Brooks Reference Brooks1991: 139); and direct perception of environmental affordances (Gibson Reference Gibson1979). Such approaches vary in the degree to which they depend on internal representations, but Shanon's (in Glenberg Reference Glenberg1997: 3) statement provides a general guide:

3.2 What is learning for?

Cognitivist approaches to learning seem to see it as the accumulation and structuring of knowledge in isolated cognitive space – a so-called competence view of learning (e.g. Long Reference Long1997). This view also seems to underlie a substantial part of our educational practice – knowledge is to be squirreled away in the mind and then tested. It further seems to account for what is perhaps the reigning methodology in SLA research: An experimental group is exposed to an artificially high dose of some linguistic feature or other, and then given a discrete-point test on that feature to see what the subjects retain. How this accords with a view of acquisition as a largely implicit, gradual, non-linear developmental process is usually left unclear (Doughty Reference Doughty, Doughty and Long2003).

The approach to learning adopted in sociocognitive SLA, on the other hand, is that, like cognition itself, learning is primarily adaptive – our formidable learning abilities have evolved to help us more sensitively and efficiently adapt to our environments. Learning can therefore be defined, at a first approximation, as the ability to adapt to our environment better (than formerly) through progressive interaction with/experience in and alignment with that environment. As a spectacularly mobile and wide-ranging species, humans need to adapt to a spectacular range of environments, as well as to retain traces of that adaptation in the mind-body-world ecology. In Wenger's (Reference Wenger1998) words, learning is therefore ‘as much a part of our human nature as eating or sleeping’ (p. 3).

3.3 What is language (and language learning) for?

Arguably the major means by which humans adapt to and exist within their ecosocial worlds is language. While research on language evolution is inevitably speculative, this much seems clear: For language to have evolved as both/either part of the human genome (Hauser, Chomsky & Fitch Reference Hauser, Chomsky and Fitch2002) and/or part of our human cultural heritage (Tomasello Reference Tomasello2003), it must first have provided us with a selective adaptive advantage. While some cognitivists see the resulting phenomenon as having primarily an internal function – Chomsky (in Andor Reference Andor2004), for example, believes that the primary function of language is to mediate cognition – the most obvious adaptive advantage would have been in socially mediating the relationship with our natural world. In other words, language has provided human beings with a selective advantage over our predators and competitors by allowing us to work together more effectively in groups. How else could such weak, ecologically unspecialized creatures have survived and prospered in competition with predators and others which were stronger, faster, or better at climbing trees?

To put it differently, language is for social action. If so, then language learning means learning how to operate adaptively in our social environments, which have the primary role of mediating our relationship with nature. From this viewpoint, learning a natal language is not first and foremost learning linguistic forms, although these are clearly important. Hopper (Reference Hopper and Tomasello1998) put it radically: ‘Children do not seem to learn sentences, but rather. . .to adapt their behavior to increasingly complex surroundings. It is an enterprise that does not stop at the age of 6, but continues throughout a lifetime’ (pp. 161−162). My own preferred version would be that children learn language, which certainly includes and is based on multiple formal systems, yet the purpose of such learning is to enable children not to become linguists, but to survive and prosper in the world. Even today, humans without language live exiguous existences, made possible in most cases only through massive social support. Until quite recently in our species history they must not have survived at all.

3.4 Why do we learn second/additional languages?

Second/additional languages have the same evolved function as first languages: They support adaptation, or social action-for-survival, but in L2 environments – settings in which the L2 is a primary medium of such action. From this viewpoint, human beings learn L2s naturally, at least to the degree needed to function adaptively in L2 worlds. This is why so much of the world's population is multilingual, at least to the level people need to be. The monolingual bias of SLA studies (e.g. Firth & Wagner Reference Firth and Wagner1997) has played a large role in supporting the apparent belief that L2 learning takes place primarily through formal education.

The paradigm example of an L2 environment is obviously a city or country where the L2 functions as the primary linguistic tool for adaptive action in everyday life. Just as obviously, however, there are many other possible L2 environments: Intense personal relationships, well-designed language classrooms, the internet, music, or an engrossing novel can all be rich and supportive environments for SLA. Interaction as I use the term should thus be extended beyond its face-to-face version to include a wide range of social activities of more mediated types. While face-to-face interaction may have special qualities that make it ideal for some kinds of learning, any social action or artifact that promotes interactive engagement with/in L2 environments potentially provides support for SLA.

3.5 How do we learn second/additional languages?

The correct answer to this question can only be ‘in many ways’. As mentioned at the beginning of my talk, no theory can claim to be anything like a total account of the complex and multifaceted phenomenon known as SLA, and most are very partial indeed. In this sense, Cook & Seidlhofer's (Reference Cook, Seidlhofer, Cook and Seidlhofer1995: 4) vision of the complex nature of language must hold equally for SLA:

Within this scheme, my main claim today has been that interactive alignment, the means by which and processes through which individuals build moment-to-moment social relations and cooperative social action, is a major facilitator of SLA. If the primary evolved purpose of language is to promote cooperative activity by which we adapt to and function in our worldly environments, then how we come to perform such action in L2 settings should be prominently included in SLA theory.

Cognitivist approaches to SLA see it as a largely invisible process – once the data enter the cognitive processor there is no direct way to study what is going on (Doughty Reference Doughty, Doughty and Long2003). A sociocognitive approach diverges from this view in positing that the SLA process is at least partly visible. That is, as individuals interact in the L2, and mutually adapt, adjust, and align their behavior in the course of interaction, signs or tools of their alignment become publicly available. If, for instance, while you and I are interacting in your second/additional language, I use a word you don't know as I point to or otherwise indicate an object in the environment, you have a pretty sound basis on which to associate the word with the object, and therefore possibly to learn that word or at least some incremental aspect of it (see Churchill Reference Churchill2008 for a detailed case study of such learning). Although a comprehensive list of sociocognitive tools of this sort does not yet exist, the following may indicate some of the possibilities:

1. 1. Gaze following: Human beings know intuitively that if they don't understand what someone is saying, and the latter is referring to the here-and-now environment, then a good way to figure it out is to follow the other's gaze. Infants certainly know this – they learn to ‘word the world’ in part by following their caregivers’ gaze while the latter talks about the environment (Baldwin et al. Reference Baldwin, Markman, Bill, Desjardins, Irwin and Tidball1996). Yu, Ballard & Aslin (Reference Yu, Ballard and Aslin2005) have experimentally demonstrated the facilitative effect of gaze-following on word learning and segmentation skills in SLA.

2. 2. Gesture: As mentioned above, pointing can be a powerful public index of what in the environment is being referred to linguistically. Yet there are many kinds of gestures providing many in-the-world affordances for SLA (e.g. Churchill et al. Reference Churchill, Nishino, Okada and Atkinson2010; Gullberg Reference Gullberg2013; Smotrova & Lantolf in press). Mori & Hayashi (Reference Mori and Hayashi2006) studied the use of ‘embodied completions’ – gestures completing a turn at talk, as when someone says ‘I want’ and then points to the desired object – in ways which are sensitively calibrated to the linguistic proficiency of L2 speakers. Although acquisition per se was not demonstrated in this study, it certainly captures the semiotic richness and availability of in-the-world gestural support for SLA.

3. 3. Facial expression: Among the richest sources of affective information are the faces of social others. We use them continuously to interpret meaning in interaction. Although no SLA research focusing specifically on facial expressions exists that I know of, it strikes me as a potentially significant area of study: Smiles often accompany linguistic expressions of affiliation, for instance, while frowns do the opposite. Even in the absence of visual data – for example during phone calls or when listening to the radio – ‘smile voices’ can be reliably identified and therefore used as indexes of meaning (Drahota, Costall & Reddy Reference Drahota, Costall and Reddy2008).

4. 4. Body positioning: In Goodwin's (Reference Goodwin, Coupland and Gwin2003) words, ‘The positioning, actions, and orientation of the body in the environment are crucial to how participants understand what is happening and build action together’ (p. 20). Although not yet widely studied, the role of bodily orientation in SLA is suggested by Seo & Koshik (Reference Seo and Koshik2010), who identified distinctive head actions used by ESL tutors to prompt repair of non-targetlike utterances by their tutees. These head actions appear to be natural, automatic, and largely unconscious, and to be substantially the same ones used to signal momentary lack of comprehension in native speaker interaction.

5. 5. Physical objects and artifacts: Aspects of the physical (but often humanly designed) environment provide rich semiotic support for language learning and use, frequently allowing the possibility of forms of interaction – and therefore learning opportunities – that could not otherwise occur. To give just one example from a very wide range of phenomena: Churchill et al. (Reference Churchill, Nishino, Okada and Atkinson2010; see Section 5 below for related work) showed how an English grammar review worksheet used by a Japanese junior high school student provided a suite of sociocognitive affordances for learning, thus acting as a ‘cognitive technology’ (A. Clark Reference Clark2001: 240):

   This dense, multiliterate technology is likely to play several crucial roles in learning and teaching. Most generally, it reduces a complex and messy reality – the relatively less determinate uses of verb tenses in real-world (often spoken) communication – to a ‘simplified world’ – a determinate and highly learnable system of relations among verb forms, adverbials, and semantic functions, all presented in written form. In doing so, it helps make the target syntactic relations publicly visible and available. . . More specifically, the cognitive technology of the worksheet affords a systematic work routine: Segregation of like items into sections and graduated levels of difficulty within each section allow students to proceed through the items in a predictable order. Similarly, teachers and tutors can use the worksheet's elaborate [numerical, alphabetical, and spatial] indexing system. . .for pedagogical purposes – for example, to establish cross-cognitive attention and action, to select items that illustrate a grammar point particularly well, or to provide instructive contrasts. (p. 240)

6. 6. Theory of mind and common knowledge: These concepts – popular in psychology and cognitive science in recent years – have a range of meanings and applications. By theory of mind I mean what Tomasello (Reference Tomasello2008) calls ‘mindreading’, and what Levinson (Reference Levinson2006) describes as ‘intention attribution: given the observed behavior, the [interactor] must be able to infer likely goals that would have motivated that behavior’ (p. 48). In other words, based on our innate, highly refined social abilities, we are often able to predict what the other is thinking, at least in so far as how they might proceed in an interaction. Accompanying such skills is the ability to ascertain common ground, or common knowledge (Lewis Reference Lewis1969; H. Clark Reference Clark1996) – knowing what the other is thinking about what you may be thinking yourself. Levinson gives the example of two people getting separated in a department store, and then determining a common place to reunite based on such knowledge. How theory of mind and common knowledge might facilitate SLA is not yet known, but if minds are in fact designed this way it could certainly simplify SLA processes: If innate abilities to understand the other's likely actions in the course of interaction exist, then it might be substantially easier to determine how the language being used is related to those actions.

7. 7. Interaction skills: Schegloff (Reference Schegloff2006) describes the rich repertoire of apparently universal abilities humans employ to conduct interaction in orderly, comprehensible ways. Turn-taking, sequential organization, and repair, for example, provide a flexible range of tools for building interaction both linguistically and non-verbally – the two typically function together as part of a larger communicative whole – that at the same time provides a predictable character to social interaction. Without such ‘interaction-management’ techniques, both speaking and learning to speak would be substantially more difficult, if not impossible. Conversation analysts (e.g. Kasper & Wagner Reference Kasper and Wagner2011) have begun to show how these universal features of interaction support SLA.

8. 8. ‘Natural pedagogy’: It is widely agreed, at least in theory, that humans are natural learners. Recent research under the rubric of natural pedagogy (Csibra & Gergely Reference Csibra and Gergely2009), however, has established the possibility that we are also natural teachers. That is, we appear to have innate abilities to structure our social activities in ways that make them more transparent, parsable, and therefore learnable, and we appear to do so especially in the presence of ‘learners’ – typically younger and less experienced members of our social groups. The phenomenon of foreigner talk, which Zuengler (Reference Zuengler, Giles, Coupland and Coupland1991) interpreted as a form of sociolinguistic accommodation (and therefore alignment), is a notable case in point as regards SLA.

9. 9. Activity frameworks: Most cooperative human activities have conventional structures. While these are in part culturally shaped, it is likely that they also have universal attributes. Service encounters, for instance, seem to share characteristics the world over: There are two parties – seller and buyer; the buyer usually initiates the transaction; the seller responds, and the specifics of the transaction are then transacted; the seller and buyer then complete the transaction as the goods are delivered and the compensation paid. Obviously, this is not a perfect template, yet the number of options in service encounters is probably quite limited. The behavioral predictability and shared goal orientation that such activity frameworks provide may constitute a rich base of environmental support for SLA.

10. 10. Language: In the foregoing list, I have purposely downplayed the role of ‘the tool of tools’ (Wells Reference Wells1999: 9) for sociocognitive alignment and interaction – language itself. This is because language and interaction are too often elided, rather than viewing more basic and long-lived forms of interaction as fundamentally underlying language use. Yet as Krashen (e.g. Reference Krashen1985) pointed out some years ago, one of the best ways to make hitherto unknown linguistic forms known/comprehensible is to embed them in already comprehensible language. While I have alluded to examples of this phenomenon in my talk today, I will briefly cite two explicitly here: Codeswitching and/or L1 use can often provide rich support at the right moment for SLA.