Modelling type-denoting concepts and words in a simulation of vocabulary development Andrew Webb, Simon McCallum and Alistair Knott Department of Computer Science, University of Otago (awebb@cs.otago.ac.nz) Steels (2000) proposes a mechanism by which a community of agents is able to negotiate a common vocabulary for referring to objects in their environment, using two simple games. In a 'discrimination game', an agent equipped with a set of simple sensory channels attempts to distinguish a target from a set of context objects. If it cannot do so, it subdivides one of its channels to make discrimination more likely in future. In a 'language game', a speaker agent identifies an object in the world, and consults a lexicon of mappings from object concepts to words to generate a word for this object, which is then sent to a hearer agent. The hearer uses its own lexicon to attempt to identify the object in question. If successful, the word-concept mapping is reinforced for both speaker and hearer; if not, the speaker indicates the intended object explicitly. Using these games, a group of agents can successfully develop a shared vocabulary. However, it is possible that the success of Steels' system is an artefact of the highly artificial classification and word-learning mechanisms which its agents use. The discrimination games make no reference to current biological theories of perception and discrimination, and the language games make no reference to psychological theories of vocabulary acquisition. The present paper describes a Steels-like system in which agents have more psychologically realistic categorisation and word-learning methods. One problem with Steels' discrimination games is that they are designed to identify objects in the environment as tokens, rather than types. Biological categorisation systems, on the other hand, break the world up into types of object for which the same set of behaviours is appropriate (see e.g. Rosch et al., 1976). To address this issue, an alternative to discrimination games was implemented, in which agents classify objects as types rather than tokens, using a self-organising map (Kohonen, 1982). We defined a set of objects which varied along a set of independent dimensions, and which could be made to cluster in different ways. An agent using a self-organising map is able to learn types which correspond to these clusters, and to recognise token objects as belonging to these types. We also defined a type-based analogue of Steels' language games, and showed that the new games allow a shared vocabulary to emerge. (Interestingly, in a parallel with Quine's (1977) observations about word learning and natural kinds, a shared vocabulary only emerges if there are genuine clusters of objects in the world.) A second, separate problem with Steels' system is with the language games themselves. In Steels' language games, the speaker only points to the target object if the hearer has not identified it correctly; this behaviour is not attested in studies on infant vocabulary acquisition in humans. The two dominant models of infant vocabulary acquisition are the joint-attention model of Baldwin (Baldwin, 1995) and the statistical learning model (see e.g. Saffran et al., 1996). We developed a formal model of each of these approaches, and implemented each model in our agent-based simulation. Again, both models allow the development of a shared vocabulary among agents. The main conclusion of this work is that Steels' system can be successfully reimplemented using more psychologically realistic object classification and word-learning methods. In addition, by focussing on type-denoting terms, the new implementation provides a sounder basis for studying the evolutionary emergence of real natural language parts of speech such as 'noun' and 'adjective'. Baldwin, D. A. (1995). Understanding the link between joint attention and language. In: Moore, C. & Dunham, P.J. (eds), Joint attention: its origins and role in development. Hillsdale, NJ: Lawrence Erlbaum. Kohonen, T. (1982). Self-organised formation of topologically correct feature maps. Biological Cybernetics, 43: 59-69. Quine, W. V. (1977). Natural Kinds. In Schwarz P. (ed) Naming, Necessity, and Natural Kinds. Cornell University Press. Rosch, E., Mervis, C. B., Gray, W., Johnson, D., & Boyes-Braem, P. (1976). Basic objects in natural categories. Cognitive Psychology, 8, 382--439. Saffran, J. R., Aslin, R. N., and Newport, E. L. (1996). Statistical learning by 8-month-old infants. Science, 274:1926 Steels, L. (2000) The puzzle of language evolution. In: Weber, G. (ed) Kognitionswissenschaft, vol. 8, nr. 4, pp. 143-150, Berlin: Springer-Verslag.