Gathering and Generating Paraphrases from Twitter with Application to Normalization

We present a new and unique paraphrase resource, which contains meaningpreserving transformations between informal user-generated text. Sentential paraphrases are extracted from a comparable corpus of temporally and topically related messages on Twitter which often express semantically identical information through distinct surface forms. We demonstrate the utility of this new resource on the task of paraphrasing and normalizing noisy text, showing improvement over several state-of-the-art paraphrase and normalization systems 1.