Context-Dependent Lexical Paraphrasing

Lexical paraphrasing is the task of extracting word-level paraphrases. Lexical paraphrases should be context dependent since a word may have different paraphrases in distinct contexts. This paper defines two issues, including dynamic lexical paraphrase extraction and paraphrase dictionary construction. With regard to the former, we investigates a novel framework for acquiring context-dependent lexical paraphrases, in which a web mining method is employed for extracting candidate paraphrases and a binary classification method is used in paraphrase validation. After analyzing the problems of the former, we produce the later issue, which includes syntax-based candidate paraphrases extracting and a classification method used in paraphrase validation. The two issues research from two aspects of lexical paraphrasing. The common between the two issues is that they both acquire context-dependent lexical paraphrases, and the difference includes: (1) the former is dynamic and the later is static; (2) the content used in the former is uncertain and the later not; (3) the target words of the former are arbitrary, and the ones of the later are verb.For the approaches of the two issues, the results show that: (1) for the dynamic lexical paraphrase extraction, the classifier for paraphrase validation is effective, which achieves an f-measure of 0.6023; (2) web mining method performs well in candidate paraphrase extraction, which extracts 2.3 correct paraphrases on average for each test word in each given context sentence; (3) for paraphrase dictionary construction, the classifier for paraphrase validation is effective, which achieves an precision of 0.7035; (4) the paraphrase dictionary we constructed is effective, which achieves a coverage of 0.3657; (5) most paraphrases extracted by our context-dependent method are context-dependent, cannot be recognized by the context-independent methods.