Abstract
A paraphrase is to express the same semantic content using different words. The use of paraphrases has been widely discussed in both the literature of teaching English writing and Natural Language Processing (NLP). In this paper, we introduce a new method for extracting synonyms and paraphrases for a given word or phrase based on Web-scale n-grams. In our approach, we use surface patterns to extract trigram over the Web, and filter out noises with rank ratio, overlap coefficient with Pointwise mutual information (PMI). Furthermore, we derive phrasal paraphrases from refined synonyms. In our experiments, we applied system to find phrase-level paraphrases, and trained a classifier for about 200 phrases. The experimental results show that the method has the potential to generate good paraphrases of a given phrase.