Paraphrase Recognition via Dissimilarity Significance Classification

Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing (EMNLP 2006), pages 18–26, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Paraphrase Recognition via Dissimilarity Significance Classification Long Qiu, Min-Yen Kan and Tat-Seng Chua Department of Computer Science National University of Singapore Singapore, 117543 {qiul,kanmy,chuats}@comp.nus.edu.sg Abstract We propose a supervised, two-phase framework to address the problem of para- phrase recognition (PR). Unlike most PR systems that focus on sentence similarity, our framework detects dissimilarities be- tween sentences and makes its paraphrase judgment based on the significance of such dissimilarities. The ability to differenti- ate significant dissimilarities not only re- veals what makes two sentences a non- paraphrase, but also helps to recall addi- tional paraphrases that contain extra but insignificant information. Experimental results show that while being accurate at discerning non-paraphrasing dissimilar- ities, our implemented system is able to achieve higher paraphrase recall (93%), at an overall performance comparable to the alternatives. 1 Introduction The task of sentence-level paraphrase recognition (PR) is to identify whether a set of sentences (typ- ically, a pair) are semantically equivalent. In such a task, “equivalence” takes on a relaxed meaning, allowing sentence pairs with minor semantic dif- ferences to still be considered as paraphrases. PR can be thought of as synonym detection ex- tended for sentences, and it can play an equally important role in natural language applications. As with synonym detection, applications such as summarization can benefit from the recognition and canonicalization of concepts and actions that are shared across multiple documents. Automatic construction of large paraphrase corpora could mine alternative ways to express the same con- cept, aiding machine translation and natural lan- guage generation applications. In our work on sentence-level PR, we have iden- tified two main issues through observation of sam- ple sentences. The first is to identify all discrete in- formation nuggets, or individual semantic content units, shared by the sentences. For a pair of sen- tences to be deemed a paraphrase, they must share a substantial amount of these nuggets. A trivial case is when both sentences are identical, word for word. However, paraphrases often employ dif- ferent words or syntactic structures to express the same concept. Figure 1 shows two sentence pairs, in which the first pair is a paraphrase while the second is not. The paraphrasing pair (also denoted Paraphrase (+pp): Authorities said a young man injured Richard Miller. Richard Miller was hurt by a young man. Non-Paraphrase (-pp): The technology-laced Nasdaq Composite Index .IXIC added 1.92 points, or 0.12 percent, at 1,647.94. The technology-laced Nasdaq Composite Index .IXIC dipped 0.08 of a point to 1,646. Figure 1: Examples: Paraphrasing & Non- paraphrasing as the +pp class) use different words. Focusing just on the matrix verbs, we note differences be- tween “injured” and “hurt”. A paraphrase recogni- tion system should be able to detect such semantic similarities (despite the different syntactic struc- tures). Otherwise, the two sentences could look even less similar than two non-paraphrasing sen- tences, such as the two in the …

Content truncated to 500 words