Vector-space models for PPDB paraphrase ranking in context

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2028–2034, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Vector-space models for PPDB paraphrase ranking in context Marianna Apidianaki LIMSI, CNRS, Universit´e Paris-Saclay 91403 Orsay, France marianna.apidianaki@limsi.fr Abstract The PPDB is an automatically built database which contains millions of paraphrases in dif- ferent languages. Paraphrases in this resource are associated with features that serve to their ranking and reflect paraphrase quality. This context-unaware ranking captures the seman- tic similarity of paraphrases but cannot serve to estimate their adequacy in specific con- texts. We propose to use vector-space se- mantic models for selecting PPDB paraphrases that preserve the meaning of specific text frag- ments. This is the first work that addresses the substitutability of PPDB paraphrases in con- text. We show that vector-space models of meaning can be successfully applied to this task and increase the benefit brought by the use of the PPDB resource in applications. 1 Introduction Paraphrases are alternative ways to convey the same information and can improve natural language pro- cessing by making systems more robust to lan- guage variability and unseen words. The paraphrase database (PPDB) (Ganitkevitch et al., 2013) contains millions of automatically acquired paraphrases in 21 languages associated with features that serve to their ranking. In PPDB’s most recent release (2.0), such features include natural logic entailment rela- tions, distributional and word embedding similari- ties, formality and complexity scores, and scores as- signed by a supervised ranking model (Pavlick et al., 2015b). These features serve to identify good qual- ity paraphrases but do not say much about their sub- stitutability in context. To judge the adequacy of paraphrases for specific instances of words or phrases, the surrounding con- text needs to be considered. This can be done using vector-space models of semantics which calculate the meaning of word occurrences in context based on distributional representations (Mitchell and La- pata, 2008; Erk and Pad´o, 2008; Dinu and Lapata, 2010; Thater et al., 2011). These models capture the influence of the context on the meaning of a target word through vector composition. More precisely, they represent the contextualised meaning of a target word w in context c by a vector obtained by com- bining the vectors of w and c using some operation such as component-wise multiplication or addition (Thater et al., 2011). We use this kind of represen- tations to rank the PPDB paraphrases in context and retain the ones that preserve the semantics of spe- cific text fragments. We evaluate the vector-based ranking models on data hand-annotated with lexical variants and compare the obtained ranking to confi- dence estimates available in the PPDB, highlighting the importance of context filtering for paraphrase se- lection. 2 Context-based paraphrase ranking 2.1 Paraphrase substitutability The PPDB1 provides millions of lexical, phrasal and syntactic paraphrases in 21 languages – acquired by applying bi- and multi-lingual pivoting on par- allel corpora (Bannard and Callison-Burch, 2005) – and is largely exploited in applications (Denkowski and Lavie, …

Content truncated to 500 words