Linguistic interaction is studied here within a multi-agent framework that we call "simulated pragmatics". Adapting G. M. Edelman's neural darwinism to build an adaptive agent is a first step towards naturalistic bases of language for an agent. Our simulation of the emergence of language conventions in an agent population then demonstrates the advantage of using a lexicalized tree adjoining grammar.

These achievements do not investigate, however, the essential link between the communication code and the entities it might denote, mainly the problem of linguistic reference. After presenting the existent theories, models and systems, we identify eleven referential cases and their expression in French, from a functionalist point of view. We propose a model for the referring capacity of an agent, which links the referring cases to perceptual states. In order to study reference on real texts, a pragmatics-inspired model is also proposed, relying on the agents' internal representations. This model is adapted to build a reference solver; the related tools and the results are given further on.

These results are based on an original frame for natural language processing systems evaluation, together with coherence criteria for quality measures. Several measures, some existing and some original, are analyzed; we define "referring information" for a measure adapted to agent communication. These measures allow us to compare our solver with other approaches, to optimize its parameters, and to support some of the previous results on the expression of referential cases.

June 21, 2000