An Unsupervised Approach for Constructing Word Similarity Network

Abstract

To evaluate how much a pair of entities or documents are similar is a common problem for current applications. Most approaches for this problem are based on the co-occurrence. However, different terms or words may represent the same entity or similar semantic in the real world since a concept often has more than one way of expression. Existing works always focus on computing semantic relatedness of words. But relatedness cannot reflect the similarity most of the time, on the other hand, most of their corpus are from common data sources such as Wikipedia and are not useful for the specialized vocabulary. In this paper, we propose a novel unsupervised approach for evaluating the semantic similarity between words by mapping texts to vector space and computing prior information. In our approach, we construct a model that can identify the words representing the same entity in special context even though they don't belong to the same concept. At last, we construct a network of words in which paths between words can reflect the evolution process of concepts. Our experimental results show that that our approach gives an effective solution to discover the semantic relationship between words, especially for words in specialty domains.