Multilingual WSD-like Constraints for Paraphrase Extraction

Proceedings of the Seventeenth Conference on Computational Natural Language Learning, pages 202–211, Sofia, Bulgaria, August 8-9 2013. c⃝2013 Association for Computational Linguistics Multilingual WSD-like Constraints for Paraphrase Extraction Wilker Aziz Research Group in Computational Linguistics University of Wolverhampton, UK W.Aziz@wlv.ac.uk Lucia Specia Department of Computer Science University of Sheffield, UK L.Specia@sheffield.ac.uk Abstract The use of pivot languages and word- alignment techniques over bilingual cor- pora has proved an effective approach for extracting paraphrases of words and short phrases. However, inherent ambiguities in the pivot language(s) can lead to inade- quate paraphrases. We propose a novel ap- proach that is able to extract paraphrases by pivoting through multiple languages while discriminating word senses in the in- put language, i.e., the language to be para- phrased. Text in the input language is an- notated with “senses” in the form of for- eign phrases obtained from bilingual par- allel data and automatic word-alignment. This approach shows 62% relative im- provement over previous work in generat- ing paraphrases that are judged both more accurate and more fluent. 1 Introduction Paraphrases are alternative ways of expressing a given meaning. Generating paraphrases that go beyond morphological variants of the original text is a challenging problem and has been shown to be useful in many natural language applications. These include i) expanding the set of reference translations for Machine Translation (MT) eval- uation (Denkowski and Lavie, 2010; Liu et al., 2010) and parameter optimisation (Madnani et al., 2007), where multiple reference translations are important to accommodate for valid variations of system translations; ii) addressing the problem of out-of-vocabulary words or phrases in MT, either by replacing these by paraphrases that are known to the MT system (Mirkin et al., 2009) or by ex- panding the phrase table with new translation al- ternatives (Callison-Burch et al., 2006); and iii) expanding queries for improved coverage in ques- tion answering (Riezler et al., 2007). Bannard and Callison-Burch (2005) introduced an approach to paraphrasing which has shown par- ticularly promising results by pivoting through dif- ferent languages for which bilingual parallel data is available. The approach consists in aligning phrases in the bilingual parallel corpus to find pairs of phrases (e1, e2) in the input language, i.e., the language to be paraphrased, which typically align to the same foreign phrases F = {f : e1 → f → e2}. This intermediate language is called pivot language and the phrases f ∈ F that support the equivalence (e1, e2) are called pivot phrases. If there exists a non-empty set of pivots connect- ing e1 to e2, e2 is said to be a paraphrase of e1. The paraphrase is scored in terms of the conditional probabilities observed in the parallel corpus1 by marginalising out the pivot phrases that support the alignment (e1, e2) as shown in Equation 1. p(e2|e1) = � f∈F p(f|e1)p(e2|f) (1) Equation 1 allows paraphrases to be extracted by using multiple pivot languages such that these languages help discard inadequate paraphrases re- sulting from ambiguous pivot phrases. However in this …

Content truncated to 500 words