Sursin abstract

We aim to build a system for processing lexical ambiguities that are

generated by complex forms of polysemy. Our main objective is

robustness: this system must be able to interpret ambiguous words in

open context, even when contextual information is degenerated.

Starting from questions like „How to linguistically describe lexical

ambiguities?‰ and „Which computational treatments of ambiguities are

available?‰, we reach a double choice: we try to model the processing of

complex forms of polysemy (the least studied in computer science,

renowned as non-calculable) inside the frame of differential semantics.

We focus especially on usage polysemy, that has never been fully

studied, neither linguistically nor computationally, and we propose a

systematic characterisation for it. As a conclusion, these ambiguities‚

vagueness is necessary to the semantic cohesion of the statements in

which they appear: they must not be resolved.

Undertaking some existing computational systems, we propose a model of

dynamic lexicon inspired from the EDGAR model. Our model, PELEAS,

integrates ambiguity in its structures. For a given ambiguous

occurrence, it computes an analysis of the contributions of an attested

usage database for the occurrence‚ meaning in context. This lexicon is a

hybrid between a symbolic system (lexical structures) and a

connectionist network (algorithm).

It is implemented by a software pack (a set of ActiveX controls). Their

development implies techniques of formal specification, advanced object

oriented design and distributed programming.

PELEAS has been validated through a test phase on this software pack.

The model proved to be perfectly robust while keeping a reasonable level

of pertinence and efficiency. Results show that it is most efficient on

joker words, plays on words and double meanings.

Keywords : natural language ˆ written text understanding and