An incremental construction method of a large-scale thesaurus using co-occurrence information

摘要

A thesaurus is one of important knowledge in natural language processing and is manually made in general. However, as growth of the scale, frequent update is difficult because it takes huge time by hand. This paper aims to construct a hierarchical large-scale thesaurus by a clustering scheme based on co-occurrence information among words. In the proposed clustering algorithm, the Kullback-Leibler divergence is introduced as a similarity measurement in order to judge superordinate and subordinate relations. Besides, the thesaurus tree can be incrementally updated in each node for a minute change such as the addition of unknown words. In order to evaluate the presented method, a thesaurus consisting of about 60,000 words is made by using about 16 million co-occurrence relationships extracted from the Google N-gram. From random data in the thesaurus, it turns out that the proposed method for a large-scale thesaurus achieves high precision of 0.826.