Until the recent past, there have been two extreme approaches toward extracting knowledge from text. On the one hand, artificial intelligence systems have long operated under the assumption that an a priori structuring of the text domain is necessary before treating the text. On the other, the information retrieval community has traditionally limited itself to simple document cooccurrence statistics as sole indicator of word semantics. The first approach is difficult to extend, while the second relies on context which is too coarse-grained. We present here SEXTANT, a complete system that uses finer-grained syntactic contexts to discover similarities between words. The system is based on the hypothesis that words that are used in a similar way throughout a corpus are indeed semantically similar. This system takes raw text in input, performs syntactic analysis to extract fine-grained contexts, compares contexts of words, and produces a list of similar words as output. Though no domain structuring is needed, the low-level semantic information that SEXTANT extracts can be thought of as an approximation to a domain-specific thesaurus. As one application, these similar words have been shown to improve classical information retrieval, through use in query expansion. This article will present a detailed description of the SEXTANT system.
Tijdschrift beschrijving