Talisman: Un Système Multi-Agents Gouverné Par Des Lois Linguistiques Pour Le Traitement De La Langue Naturelle

Natural language processing raises the problem of ambiguities and multiple solutions which follow frnm them. The knowledge gained when using the morphosyutactic atmlyser CRISSTAL showed how necessary it was to overcome this issue. The architecture with sequential levels, in which each module corresponds to a linguistic level (pretreatments, morphology, syntax, semantics) has shown its limits. A sequential architecture does not allow a real exchange between different modules. This le~als to the non availability of the linguistic information for the reduction of ambiguities, at the moment they are needed. The necessity for cooperation between different modules has lead us to envisage a new architecture which stems from the techniques of distributed artificial intelligence. The paper presents this new architecture which based on distributed artificial intelligence techniques treats the inherent problems of natural language processing. One of the originalities resides in the distributed treaunent of sentence ,'malysis (as apposed to a classic sequential treatment) and in the introduction of linguistic laws which allow the management of the communication between agents, without appealing to a central control. The Talisman system is an environment which iutegrates linguistic tools where different ,agents c,'m bring into use diflerent methods such as symbolic and/or statistic ones. The Talisman system contributes to the following points in the linguistic domain : • the restriction of ,ambiguities by agent cooperation, • rendering structures less complex by using local grsmlnar rnle.s, • the trealment of unce~dn intbnnation. It can : • function with partial analyses at diflerent classical levels of mndysis, • change strategies accordiug to the applications or the corpus useA, • use linguistic laws which are eaaily modifiable. At the implementation level, the system brings openess to the modification of dictionaries, grammars mid strategies of analysis, and the necessary mechanisms for the integration of new modules. Talisman is a linguistic environment based on the most recent techniques used in software engineering environments. It provides mechanisms for data and control integration of linguistic tools. This paper is organized as follows. After a short overview in section 2 about the problems of sequential architectures, we establish the main objectives of our work in section 3. We present, in section 4, the contribution of multi-agents systems "governed by laws". In section 5, we define the structure of an agent and of its society. The implementation is presented in section 6.