Chinese Character Expansion for Retrieving Japanese Paraphrases

This paper proposes two methods of query expansion for retrieving paraphrase candidates indexed by Kanzi (Chinese) characters. The idea is to calculate similarity between Kanzi characters based on an ordinary thesaurus defining relations between words. The local analysis method calculates similarity of Kanzi characters based on a semantic class to which words in a query belong, in contrast, the global analysis method calculates similarity based on whole semantic classes in the thesaurus. The methods were evaluated by using the EDR concept dictionary which defines about 410,000 concepts. In the experiments, both headwords and concept descriptions of dictionary entries were indexed by Kanzi characters, and concept descriptions were retrieved by giving a headword as a query. The experiments showed that Kanzi character expansion is significantly effective, and the global analysis method is better than the local analysis method in recall.