Construire un Espace Sémantique Pour Représenter la Polysémie D'un Marqueur Grammatical

The study presented here constitutes an important part of a general project of modeling polysemy, which involves three steps: linguistic theory, mathematical formalization, and computer implementation. In this approach, two continuous spaces are to be constructed to model the behaviour of a polysemous expression: a contextual space to represent the context in which the expression appears and a semantic space to represent the different values that can be taken by this expression. We focus here on the problems involved in the construction of the semantic space, illustrating our general method on the example of the French adverb encore. After a brief description of our framework, we present a general methodology to represent the different meanings of the polysemous expression as regions of a continuous space. Analysis of a real corpus of utterances containing encore revealed a small number of typical values. There appear to be both (1) a unique semantic kernel of the expression, which defines the basic abstract operation associated with the expression, and (2) different features which have to be instanciated in every occurrence of the expression in context, and which determine the different dimensions of the semantic space. Then we show how different types of interpretative figures of a polysemous expression can be represented in this space.