Abstract
Paraphrases are alternative ways to express the same meaning. Automatically generating paraphrases can be applied in many of National Language Processing tasks. We propose a method for generating paraphrases which preserve the meaning and the syntax of a given phrase. In our approach, the paraphrasing problem is transformed into a graph representing direct and indirect paraphrase relations. The method involves incorporating various linguistically motivated features to reflect the similarities of paraphrase candidates, and using Weighted PageRank Algorithm to evaluate the relevance of paraphrase candidates. Evaluation on a set of phrases commonly used in research articles shows that our method significantly outperforms the state-of-the-art methods under both semantic and syntactic considerations.