Paraphrase Identification by Text Canonicalization

This paper proposes an approach to sentencelevel paraphrase identification by text canonicalization. The source sentence pairs are first converted into surface text that approximates canonical forms. A decision tree learning module which employs simple lexical matching features then takes the output canonicalized texts as its input for a supervised learning process. Experiments on the Microsoft Research (MSR) Paraphrase Corpus give comparable performance to other systems that are equipped with more sophisticated lexical semantic and syntactic matching components, with a Confidence-weighted Score of 0.791. An ancillary experiment using the occurrence of nominalizations suggests that the MSR Paraphrase Corpus might not be a rich source for learning paraphrasing patterns.