A Thematic Segmentation Procedure for Extracting Semantic Domains from Texts
Résumé
Thematic analysis is essential for a lot of Natural Language Processing (NLP) applications, such as text summarization or information extraction. It is a two-dimensional process which has both to identify the thematic segments of a text and to recognize the semantic domain concerned by each of them. This second task requires having a representation of these domains. Such representations are built in Information Retrieval or Text Categorization fields by grouping together the words of a set of texts which have been manually linked to the same domain. We claim that this kind of method can only be apply to characterize very general topics. We propose here a method for building the representation of narrower semantic domains without any manual intervention. First, we present a procedure for the thematic segmentation of texts which relies on lexical cohesion evaluated from a collocation network. This procedure allows us to have basic units that are more thematically coherent than a whole text. Then, we show how these units can be aggregated together, according to a similarity measure, to build the representation of semantic domains in an incremental and unsupervised way.
Domaines
Informatique et langage [cs.CL]
Origine : Fichiers produits par l'(les) auteur(s)
Loading...