Could we automatically reproduce semantic relations of an information retrieval thesaurus?

Abstract

A well constructed thesaurus is recognized as a valuable source of semantic information for various applications, especially for Information Retrieval. The main hindrances to using thesaurus-oriented approaches are the high complexity and cost of manual thesauri creation. This paper addresses the problem of automatic thesaurus construction, namely we study the quality of automatically extracted semantic relations as compared with the semantic relations of a manually crafted thesaurus. The vector-space model based on syntactic contexts was used to reproduce relations between the terms of a manually constructed thesaurus. We propose a simple algorithm for representing both single word and multiword terms in the distributional space of syntactic contexts. Furthermore, we propose a method for evaluation quality of the extracted relations. Our experiments show significant difference between the automatically and manually constructed relations: while many of the automatically generated relations are relevant, just a small part of them could be found in the original thesaurus.