To solve the weakness of Chinese synonym dictionary Tongyici-Cilin’s，which can’t be used as a context-dependent paraphrase corpus, a word-level paraphrase method was presented to improved the Chinese paraphrase extraction accuracy. Based on its contextual sentence, the target word’s paraphrase candidates were identified and extracted from large-size corpuses. The target word was then paired up with each candidate, and a five-feature probability model captured the information of the target word, the context sentence, and the paraphrase candidates were established. Values of those five features were inputted to train a binary classifier which subsequently filtered out the paraphrase candidates. The experiment proved that through data mining the method for retrieving candidate paraphrases from large-size corpuses had pragmatic value, and on average 3.1 correct paraphrases were obtained for a word. Binary classifier was efficient in filtering out the paraphrases, with an accuracy rate of 0.65. 32% of the retrieved paraphrases could not be found in the Expanded Chinese Synonym Dictionary.
