_______________________

Gérard Sabah

The Language and Cognition group (some fifteen senior researchers and about fifteen graduate students) is an interdisciplinary team that covers various aspects of natural language understanding systems: in addition to artificial intelligence specialists, it is composed of researchers from various other disciplines (semantics, linguistics, cognitive psychology and mathematics). It studies language as an object (linguistics) as well as the main tool of communication (pragmatics). Automatic processes are compared with human mechanisms of understanding and generation (psycholinguistics) and allows for the simulation of such behaviours by computers (artificial intelligence, with formal aspects of representation formalisms and computer architectures, which are implemented in various applications). Many people in the group also play an important role for the development of cognitive science.

The ongoing research of the group is distributed over three axes, each containing two topics: software architectures, knowledge representation and use, and discourse understanding and generation.

* architecture

reflective systems and distributed artificial intelligence (Gérard Sabah)

connectionism (Daniel Memmi)

* knowledge

formalisms (Gérard Ligozat)

automatic acquisition (Brigitte Grau)

* discourse

understanding and production  (Violaine Prince)

dialogue. (Anne Vilnat)

The first axis is concerned with new techniques in AI for knowledge representation and communication (blackboards, distributed artificial intelligence, parallel implementations, connectionism) and their relations with human behaviour (neural level, reflexivity and consciousness for the Caramel model). These theoretical models are implemented in various man-machine dialogue applications (STANDIA-DIABOLO, TEDDI-MIEL), and studies of semantic flexibility (EDGAR) and spatio-temporal aspects. Moreover, recent research along these axes account for non conscious processes, underlying natural understanding with non explicit control.

The second axis implies two distinct aspects: on one hand, the study of the formalisms best suited for representing the semantic and pragmatic structures built by the understanding and generative processes (our main options are: Sowa's conceptual graphs, generalised intervals and temporal logic). On the other hand, the automatic acquisition of knowledge is studied along two main points: a dynamic use of encyclopedic knowledge, and the implementation of an hybrid model combining connectionist and symbolic techniques, in order to acquire semantic and pragmatic knowledge from experience.

The third axisinvolves the analysis of the reasoning mechanisms used to implement the above studies. These mechanisms intervene at various stages: sentence analysis and generation, text understanding and production, and to dialogue management (with focus on planning processes and belief handling). The introduction of flexibility and robustness in all these levels should also be mentioned. This leads to several applications, that allow us to test the feasibility of these theoretical ideas and to validate them in real world applications.

Thus, the research of the group takes into account the main symbolic aspects relevant within "language engineering" (with a special focus on semantics and its relations to lexicon), and also tryiesto tackle the issue of grounding these symbols, as well as the relations between meaning and perception. These basic links between form and meaning are also complemented: communicative context, connotations and the argumentative effect of utterances should be taken into account, more or less independently of their truth value. Hence, meaning cannot be an a priori representation to be computed from the message alone. Instead, meaning modifies the cognitive context of the listener. In the same way, memory is not simply a place to store and retrieve information: the accessibility of knowledge varies dynamically. New theoretical ideas are necessary for these viewpoints to be considered within our man-machine communication models.