Paraphrase Identification as Probabilistic Quasi-Synchronous Recognition

Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 468–476, Suntec, Singapore, 2-7 August 2009. c⃝2009 ACL and AFNLP Paraphrase Identification as Probabilistic Quasi-Synchronous Recognition Dipanjan Das and Noah A. Smith Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213, USA {dipanjan,nasmith}@cs.cmu.edu Abstract We present a novel approach to decid- ing whether two sentences hold a para- phrase relationship. We employ a gen- erative model that generates a paraphrase of a given sentence, and we use proba- bilistic inference to reason about whether two sentences share the paraphrase rela- tionship. The model cleanly incorporates both syntax and lexical semantics using quasi-synchronous dependency grammars (Smith and Eisner, 2006). Furthermore, using a product of experts (Hinton, 2002), we combine the model with a comple- mentary logistic regression model based on state-of-the-art lexical overlap features. We evaluate our models on the task of distinguishing true paraphrase pairs from false ones on a standard corpus, giving competitive state-of-the-art performance. 1 Introduction The problem of modeling paraphrase relation- ships between natural language utterances (McK- eown, 1979) has recently attracted interest. For computational linguists, solving this problem may shed light on how best to model the semantics of sentences. For natural language engineers, the problem bears on information management sys- tems like abstractive summarizers that must mea- sure semantic overlap between sentences (Barzi- lay and Lee, 2003), question answering modules (Marsi and Krahmer, 2005) and machine transla- tion (Callison-Burch et al., 2006). The paraphrase identification problem asks whether two sentences have essentially the same meaning. Although paraphrase identification is defined in semantic terms, it is usually solved us- ing statistical classifiers based on shallow lexical, n-gram, and syntactic “overlap” features. Such overlap features give the best-published classifi- cation accuracy for the paraphrase identification task (Zhang and Patrick, 2005; Finch et al., 2005; Wan et al., 2006; Corley and Mihalcea, 2005, in- ter alia), but do not explicitly model correspon- dence structure (or “alignment”) between the parts of two sentences. In this paper, we adopt a model that posits correspondence between the words in the two sentences, defining it in loose syntactic terms: if two sentences are paraphrases, we expect their dependency trees to align closely, though some divergences are also expected, with some more likely than others. Following Smith and Eis- ner (2006), we adopt the view that the syntactic structure of sentences paraphrasing some sentence s should be “inspired” by the structure of s. Because dependency syntax is still only a crude approximation to semantic structure, we augment the model with a lexical semantics component, based on WordNet (Miller, 1995), that models how words are probabilistically altered in generating a paraphrase. This combination of loose syntax and lexical semantics is similar to the “Jeopardy” model of Wang et al. (2007). This syntactic framework represents a major de- parture from useful and popular surface similarity features, and the latter are difficult to incorporate into our probabilistic model. We use a product of experts (Hinton, 2002) to bring together …

Content truncated to 500 words