Acquisition sur corpus d'informations lexicales fondées sur la sémantique différentielle
Mathias Rossignol · HAL (Le Centre pour la Communication Scientifique Directe) · 2005
Semantic lexicons are an essential resource to let many natural language processing applications (automatic summarization, information retrieval, automatic translation, etc.) penetrate the meaning of a text. The relevance of the information gathered by those lexicons raises a problematic question: the meaning of a word like soap, for example, varies considerably whether it is considered in a sanitary or televisual context. A linguistically motivated and cost-effective way of building semantic lexicons precisely adapted to a certain domain of expression consists in “learning” word meanings from their actual usage as observed in a representative collection of texts, or corpus. To answer this challenge, we propose in this document a three-stage methodology for the automatic acquisition of lexical semantic information from texts, based on the linguistic principles of F. Rastier's Interpretative semantics. Thanks to a statistical analysis of word uses, employing both classical and novel methods, we first manage to bring together words belonging to a same domain (for example data, transfer, network for IT), then to build classes of words having a similar meaning (data and information). We finally propose a first method to put to light fine-grained meaning distinctions between close words (data is more “concrete” than information), thus reaching a level of meaning refinement never before attained, to our knowledge, by automatic means.