Syntactic Constraints on Paraphrases Extracted from Parallel Corpora

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 196–205, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Syntactic Constraints on Paraphrases Extracted from Parallel Corpora Chris Callison-Burch Center for Language and Speech Processing Johns Hopkins University Baltimore, Maryland ccb cs jhu edu Abstract We improve the quality of paraphrases ex- tracted from parallel corpora by requiring that phrases and their paraphrases be the same syn- tactic type. This is achieved by parsing the En- glish side of a parallel corpus and altering the phrase extraction algorithm to extract phrase labels alongside bilingual phrase pairs. In or- der to retain broad coverage of non-constituent phrases, complex syntactic labels are intro- duced. A manual evaluation indicates a 19% absolute improvement in paraphrase quality over the baseline method. 1 Introduction Paraphrases are alternative ways of expressing the same information. Being able to identify or gen- erate paraphrases automatically is useful in a wide range of natural language applications. Recent work has shown how paraphrases can improve question answering through query expansion (Riezler et al., 2007), automatic evaluation of translation and sum- marization by modeling alternative lexicalization (Kauchak and Barzilay, 2006; Zhou et al., 2006; Owczarzak et al., 2006), and machine translation both by dealing with out of vocabulary words and phrases (Callison-Burch et al., 2006) and by expand- ing the set of reference translations for minimum er- ror rate training (Madnani et al., 2007). While all ap- plications require the preservation of meaning when a phrase is replaced by its paraphrase, some addi- tionally require the resulting sentence to be gram- matical. In this paper we examine the effectiveness of placing syntactic constraints on a commonly used paraphrasing technique that extracts paraphrases from parallel corpora (Bannard and Callison-Burch, 2005). The paraphrasing technique employs various aspects of phrase-based statistical machine transla- tion including phrase extraction heuristics to obtain bilingual phrase pairs from word alignments. En- glish phrases are considered to be potential para- phrases of each other if they share a common for- eign language phrase among their translations. Mul- tiple paraphrases are frequently extracted for each phrase and can be ranked using a paraphrase proba- bility based on phrase translation probabilities. We find that the quality of the paraphrases that are generated in this fashion improves significantly when they are required to be the same syntactic type as the phrase that they are paraphrasing. This con- straint: • Eliminates a trivial but pervasive error that arises from the interaction of unaligned words with phrase extraction heuristics. • Refines the results for phrases that can take on different syntactic labels. • Applies both to phrases which are linguistically coherent and to arbitrary sequences of words. • Results in much more grammatical output when phrases are replaced with their para- phrases. A thorough manual evaluation of the refined para- phrasing technique finds a 19% absolute improve- 196 ment in the number of paraphrases that are judged to be correct. This paper is structured as follows: Section 2 describes related work …

Content truncated to 500 words