Lexical paraphrasing based on automatically constructed corpora

This paper presents a new method for lexical paraphrasing. The method first constructs a paraphrase corpus by automatically translating a bilingual parallel corpus into a monolingual parallel corpus, from which candidate paraphrases for words are extracted. After that, a new statistical model is proposed for lexical paraphrasing, which selects the best paraphrase for a word in a given context sentence. Experimental results show that the automatically constructed paraphrase corpus is effective for lexical paraphrasing. In addition, the presented paraphrasing model significantly outperforms two conventional models, enhancing precision and recall by about 10%, respectively.