Paraphrase Identification by Text Canonicalization

Proceedings of the Australasian Language Technology Workshop 2005, pages 160–166, Sydney, Australia, December 2005. Paraphrase Identification by Text Canonicalization Yitao Zhang and Jon Patrick School of Information Technology University of Sydney Sydney, Australia, 2006 {yitao, jonpat}@it.usyd.edu.au Abstract This paper proposes an approach to sentence- level paraphrase identification by text canon- icalization. The source sentence pairs are first converted into surface text that approxi- mates canonical forms. A decision tree learn- ing module which employs simple lexical match- ing features then takes the output canonical- ized texts as its input for a supervised learn- ing process. Experiments on the Microsoft Re- search (MSR) Paraphrase Corpus give compa- rable performance to other systems that are equipped with more sophisticated lexical se- mantic and syntactic matching components, with a Confidence-weighted Score of 0.791. An ancillary experiment using the occurrence of nominalizations suggests that the MSR Para- phrase Corpus might not be a rich source for learning paraphrasing patterns. 1 Introduction Paraphrase identification is the task of recog- nizing text fragments with approximately the same meaning within a specific context. has been recently proposed as an application-It independent framework for measuring semantic equivalence in text, which is critical to many natural language systems like Question Answer- ing, Information Extraction, Information Re- trieval, Document Summarization, and Machine Translation. This paper proposes an approach to identi- fying sentence-level paraphrase pairs by trans- forming source sentences into more canonical- ized text forms. By “canonical form”, we mean a transformed text which is more generic and simpler in someway than the original text, fol- lowing the idea of restricted languages. For ex- ample, the sentence Remaining shares will be held by QVC’s management. is transformed into a more canonicalized form by changing it from the passive to active voice producing QVC ’s management will hold Re- maining shares. which is more common in Subject-Verb-Object (SVO) languages like English, while the Passive Voice in the source sentence usually occurs in scientific and business text where a more formal writing style is used. This approach is consistent with Chomsky’s Transformational Grammar, in which syntacti- cally different, but semantically equivalent sen- tences can be related by their identical deep se- mantic structures (Chomsky, 1957). However, it is generally difficult to efficiently analyze any corpus by using the Transformational Grammar due to its high complexity and computational overhead (Hausser, 2001). In our approach, we only attempt to transform parts of the surface structure into a more generic text representa- tion within the context of the paraphrase identi- fication problem. The underlying hypothesis of this approach is that if two sentences are para- phrases of each other, they have a higher chance of being transformed into similar surface texts than a pair of non-paraphrase sentences. In this paper, only a set of limited canoni- calization rules have been applied as a prelim- inary attempt to evaluate the effectiveness of the methodology. The objective is not to cre- ate grammatically correct text sequences from source sentences, but to enable the true para- phrases to …

Content truncated to 500 words