A Semi-Supervised Approach to the Construction of Semantic Lexicons

Abstract

A growing number of applications require dictionaries of words belonging to semantic classes present in specialized domains. Manually constructed knowledge bases often do not provide sufficient coverage of specialized vocabulary and require substantial effort to build and keep up-to-date. In this thesis, we propose a semi-supervised approach to the construction of domain-specific semantic lexicons based on the distributional similarity hypothesis. Our method starts with a small set of seed words representing the target class and an unannotated text corpus. It locates instances of seed words in the text and generates lexical patterns from their contexts; these patterns in turn extract more words/phrases that belong to the semantic category in an iterative manner. This bootstrapping process can be continued until the output lexicon reaches the desired size. We explore employing techniques such as learning lexicons for multiple semantic classes at the same time and using feedback from competing lexicons to increase the learning precision. Evaluated for extraction of dish names and subjective adjectives from a corpus of restaurant reviews, our approach demonstrates great flexibility in learning various word classes, and also performance improvements over state of the art bootstrapping and distributional similarity techniques for the extraction of semantically similar words. Its shallow lexical patterns also prove to perform superior to syntactic patterns in capturing the semantic class of words.