Extracting Part-Whole Relations from Online Encyclopedia

Abstract : Automatic discovery of part-whole relations is a fundamental problem in the area of information extraction. In this paper, we present an unsupervised approach to learning lexical patterns from online encyclopedia for extracting part-whole relations. The only input is a few part-whole instances. To tackle the term recognition problem, terms from the domain of the seeds are extracted taking use of the semantic information contained in the online encyclopedia. Instead of collecting sentences that contain relation instances from the seeds, we introduce a novel process to select sentences that may indicate part-whole relations. Patterns are produced from these sentences with terms replaced by Part and Whole tags. A similarity measurement based on a new edit distance is used and an algorithm is described to cluster similar patterns. We rank the pattern clusters according to their frequencies, and patterns from the top-k clusters are chosen to be applied to identify the new part-whole relations. Experimental results show that our method can extract abundant part-whole relations and achieve a preferable precision compared to the other state-of-the-art approaches.