Language and Robots: A Distributed View
Language and Robots: A Distributed View
Language and Robots: A Distributed View
As noted in the closing session, the symposium united a multidisciplinary array of thinkers and doers. Generally, however, there was a surprising degree of agreement. First, robots can already learn how to map forms onto meanings. Second, new benchmarks are needed to inspire roboticists to do more than seek to achieve reference. Third, this raises problems which, it seems, compare with putting a man on the moon. Should we, following Deb Roy, work at the level of behaviour or, as Tony Belpaeme suggests, use the power of multimodal expression in seeking to design language aware robots. Do we work stepwise or, rather, aim to match humanlevel performance? As Rob Clowes pointed out, we can ask what models achieve. For example, modularity seems to be a necessary engineering principle. However, if used to do theoretical work, the same principle may undermine attempts to understand the emergence of cognitive powers. This is because agents gain from integrating the use of expressive resources. The issue is central to Mike Andersons hypothesis that brains function by redeploying resources. It is, moreover, echoed in Stephen Cowleys work on how human agents converge on co-action routines. As agents, we integrate artefacts and dynamic processes by finding ways of co-ordinating both language and other activity. To approach what Gerhard Segarer calls use cases, robots need new kinds of middleware. Indeed, as Roger Moore emphasises, it may be mistaken to think that speech recognition plugs in to a language system. The Call presented language as a dynamic cognitive process that uses artefacts in controlling embodied action. Contributors interpret this in many ways. Thus, while all emphasise the importance of physics, some rely on linguistic models of form and meaning. By contrast, others emphasise dynamics or how these patterns are integrated with speech, gesture, gaze and, above all, with the timing of expression. Embodied views such as those presented by Lakoff and Johnson (1999) or Tomasello (2003) are thus in competition with those based on the work of Hutchins (1995) and Clark (1997; 2006). Rather than posit that language re-represents individual experience, artefacts (including words) may become physical features that shape what we think, feel and do. As we participate in co-action, linguistic resources may prompt social agents to use each other in extending their cognitive powers. Appeal to narrow embodiment emphasises how philosophical views have shaped our sense of reference. In Towards a mechanistic model of referential semantics, Deb Roy refuted the widely held view that it is enough for reference to exploit correspondence relations based in causally mapping object categories to the world. Today robots use sensorimotor schema that encode different actions. In affordancebased semantics, robots slow dynamics to produce stable representations: these reifications index (inner) beliefs about a world. While illuminating, the work simply assumes that linguistic forms exist. By contrast, Luc Steels asks how forms can emerge by coming to set parameters. In Fluid Language Games and the Emergence of Grammar, he shows the sense in which he has solved the symbol grounding problem. Like Roys, his agents develop schema that encode symbols but, in these models, they do so socially. As a result they build autonomous systems that establish their own symbols. They rely not on sensorimotor dynamics alone but social, naming games. They usenot physical symbols but consistent patterns of relations. Implicitly, language may be grounded not in physical invariants but how agents learn from relating to each other in the world. Several other contributors model how
artificial agents can establish reference. Given their range of interests, they also seek to find ways of moving beyond the establishment of reference. Wouter van den Broek focuses on how (inner) meaning might emerge. Bringing learning into account, he showed how referentially grounded symbols can set off abductive relations. By contrast Henrik Zender presented the Saarbrucken model of reference in showing how the resulting categories can invoke not just perceived objects but categories in possible worlds. While having in mind the distributed/extended nature of language awareness, Chauhan focuses on the inner learning a memory processes that allow to establish referencial relations. In a practically oriented paper, he presented the state of the art in developing robots that use teachers in learning to index up to 68 referential relations. Finally, emphasising how language is distributed between people, Tony Belpaeme combines an interest in how language shapes conceptualizing with an emphasis on how we access each others experience. Provocatively, he concludes that the standard focus on inner embodiment (and meaning) may impede our work on language modelling. Formal patterns act as measures and controls that allow an agent to use sensorimotor and social factors in digitizing grounded categories. But how do the patterns emerge? While most assume the existence of linguistic symbols (Roy, van den Broek, Zender and Chauhan), their emergence can also be traced to naming games (Steels). This, however, leaves naming unexplained. It echoes the linguistic tradition which assumes that we know how to identify forms (and meanings). Language, it is posited, plugs into speech recognition. However, the papers in this field fuel scepticism about inputoutput models. Jonas Hornstein describes work with a synthesiser that uses mirror neuron theory in developing a motor theory of production. Stressing caregiver imitations it was suggested that, like a synthesiser, children learn to produce speech sounds that fit phonological descriptions. In spite of the models simplicity, results are limited. A more radical challenge is posed by Simon Worgan who discards the view that brains represent symbols. Instead of positing (inner) phonology, he appeals to the physics of communication. If this is the basis for speech, the consequences are far reaching. Above all, one has to overthrow the assumption that we rely on an inner language system and, by extension, the view that reference grounds language. Roger Moore also challenges the orthodox. Rather than continue to base speech recognition on statistics, he regards speech as core cognitive behaviour. Our capacities depend not on invariances but speech based co-active interactive behaviour. Rather than appeal to (inner) phonology, he brings the users needs to the fore. Systems can be designed to produce not linguistic strings but continuous coupled continuous behaviours. In a PRESENCE model, words and naming can only result from speech. Ultimately, they depend on how this co-evolved with human agents in the time-scales of evolution, cultural history and development. Taking a larger view, Ee Sian Neo reported on work with a humanoid robot. Since moving agents such as this require constant updating of the world, they could gain much from better speech recognition systems. As noted above, progress in that field is urgently required. This is echoed in more applied fields. Thus Finale Doshi asks how models of dialogue management can be used in wheelchairs that rely on machine learning. In this, she stresses the importance of giving the user control of systems interpretations. In more theoretically oriented work, most of the remaining papers challenge the verbal focus of orthodox linguistics. In Henrik Janssons presentation, he complains that reference based models focus on language and vision. To get
beyond literal meaning, the Saarbrucken team have developed a system whose core functions use crossmodal concept binding. In elaborating this work, Jan Geert Krujff stressed that machines which exploit situated dialogue need an incremental model that constantly updates information. This, he suggested, requires a semeiotic view that can only be modelled on a dynamic integrated network. The proposed system generates complex relational structures that, to function efficiently, must draw on situational factors to simplify form-based processing. In Danijel Skoajs extension of this view, the amodal binder is shown to permit a two-way relationship with learning. Finally, in Gaze in Situated Dialogue, the group suggests that gaze is central to regulating the dynamics of dialogue. While in need of further research, gaze not only shows understanding but, in so doing, shapes the other agents speech. While not pursued, the thought implies that co-action integrating signals across modalities is crucial to language spread. It exemplifies co-action where one persons behaviour acts as a contextual control for the doings of the other. In examining the coexpressiveness of speech and gesture Katherina Rohlfing gives language spread more emphasis. She shows, for example, that interpretation of gesture is irreducible to kinetic information and, conversely, that the same movements are interpreted differently in robotic and human agents. On this view, robots can only be used in dialogical settings once they are able to engage with concurrent patterning of human speech and various forms of expressive dynamics. Finally, in Gerhard Segherers Towards learning by interacting , it was stressed that neither communication nor teaching could be construed around how humans use formal patterns. To learn from human agents, we need to explore how we teach and in so doing much can be learned from the methods of pattern recognition research. In the longer run, the approach may throw new light on the functions of human expression. This, he believes, has much to offer both the design of machines that learn from interacting and also in understanding of how perception and action (including speech) contribute to learning in different stages and settings. It is striking that, positing use cases, Segherers approach presupposes extended models that de-emphasise forms and encapsulated modules. In striking convergence, this fits Moores speech models and also Andersons massive redeployment hypothesis. Carlos Herrara described a model of how words might become cognitive. Stressing that we use cognitive resources from beyond the body, he explores the power of a pattern completing system that uses resources in the world beyond the agent. In this model a robots learning was enhanced by Hebbian learning that allowed it to talk to itself. Far from simply learning a persons way of solving a task, the robot was said to be reassembling action and perception in ways that led to shared understanding. In contextualizing this work Rob Clowes addressed how we might think about distributed language. He stressed that it was not only spread among persons as highlighted in naming and guessing games but also between people and artefacts. It is so radically distributed that it can be used to design talking video cameras and thus exemplifies how language can be embedded in both culture and artefacts. In the robotic model, agents use the world not only in talking to an incipient self but also in learning from a modeller. To pursue this, he focuses on what models achieve. Broadly, this fits Stephen Cowleys view that robots can be used, among other things, as a test-bed for exploring how external resources drive systems to self-organize their agency. In arguing that MacDormans (2007) person problem be made central to robotics, he suggests that systems be designed to self-construct as norm-following agents. As such, they can exploit our expectations by learning to orient to linguistic,
expressive and cultural conventions. Turning from how language emerges, greater attention will thus be given to the time scales of development, conscious experience and biomechanics. In Segherers terms what matters is how multimodal activity enables nonverbal communication to be integrated with speech in ways that, among other things, produce patterns that permit verbal analysis. Finally, Mike Anderson also warns us against appeal to narrow embodiment. In Circuit sharing for action grounded meaning he too emphasises that gestures and other kinds of expression are inseparable from speech. Physical activity is part of meaning: language is not the use of forms- but intrinsic to social (and, at times, solo) action. Anderson offers a hypothesis that can be squared with this view. Instead of positing that the brain represents quasi-linguistic forms (a mental lexicon), he suggests that we depend on massive redeployment of neural resources. The brain is a multi-use system that can be defined by the functions that develop in its bodys world. Further, new functions (e.g. those associated with language) are thus very scattered. In using robots to explore language and in developing linguistic theory relevant to roboticists we can ask what language does for motor control. Instead of assuming that language is fundamentally communicative or representational, its main function may be that of enabling us to set (and reset) parameters used to control how we move and, thus, assess the doings of others. The symposium asked what robotics shows the language sciences. First, unless robots can use the dynamics of human expression, the achievement of reference remains rather toothless. Until they use the cognitive dynamics of language (Cowley, 2007), robots are caught in an artificial world. We need to follow Clowes with: What have current approaches achieved? The results are impressive. Strikingly, they show that language does much more than bi-directionally map forms onto (referential) meanings. But current roboticist approaches are also focused on very small parts of the full language picture. That is the result of practical factors (technology, engineering team organization) as well as of not having a widely accepted theory on how the full language picture can be implemented in robots. It is thus necessary to ask how we wish to apply language. In spite of linguistic tradition, it matters that human expression is multi-modal, irreducible to sensory input, unites learning and interaction and, among other things, can inhibit thinking. Linguistic expression functions in adapting to each other, to how we perceive tasks and to how each manifestly construes events. Verbal patterns (and reference) are, it seems, only part of the mix. For this reason, we must model the simultaneous deployment of neural and dialogical resources. This is non-trivial. To move beyond the focus on reference, roboticists face stark choices. To ask how language connects agents whether by learning, cognition or communication they need to separate design requirements from hypotheses about internal models. Equally, design reasons may demand that emphasis be given to variability and timing of speech and other expression. Do roboticists then need an extended embodied view where verbal patterns are integral for multimodal dynamics? This is an issue for the future. There are plenty of others. For example, it is striking that, in spite of the Call for papers, little was said about 1st person phenomenology, how joint experience is embodied, or silent thought. Are these issues beyond robotics? Is this a field where we know what tasks are simple? Or can we ask how interaction shapes learning as agents redeploy sensorimotor skills to cognitive ends? Are speech recognition systems that approximate human performance beyond us? If so, why? Finally, to address such
issues, it seems likely that more attention may need to be given to how biological agents solve such problems. Why do living systems apparently not use referring signals? How does human signalling compare with that of other social species? To pursue such matters, we face challenges that can fuel debates and, we hope, lead to new kinds of modelling. While the theoretical and engineering fronts do not meet, robotics and the language sciences will certainly continue to influence each other, inspiring new developments. We aim to pursue these and related issues at a follow up symposium by the Distributed Language Group in 2009.
References Clark, A., (1997). Being There: Putting Brain, Body, and World Together Again . Cambridge MA: MIT Press. Clark, A. (2006). Language, embodiment and the cognitive niche. Trends in Cognitive Sciences, 10/8: 370-374. Cowley, S.J., 2007. The cognitive dynamics of distributed language. Language Sciences, 16/1: 575-583. Hutchins, E., 1995. Cognition in the Wild. MIT Press, Cambridge MA. Lakoff, G., Johnson. M. (1999). Philosophy In The Flesh: The Embodied Mind and Its Challenge to Western Thought. New York: Basic Books. MacDorman, K. F. (2007). Life after the symbol system metaphor. Interation Studies, 8(1), 143-158. Seabra Lopes, L., Connell, J.H. (2001) Semisentient Robots: Routes to Integrated Intelligence, IEEE Intelligent Systems, vol. 16, n. 5, Computer Society, p. 10-14. Tomasello, M. (2003). Constructing a Language: A Usage-based Theory of Language Acquisition. Cambridge MA: Harvard University Press.