ALTN: Word Alignment Features for Cross-lingual Textual Entailment

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 128–132, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics ALTN: Word Alignment Features for Cross-lingual Textual Entailment Marco Turchi and Matteo Negri Fondazione Bruno Kessler Trento, Italy {turchi,negri}@fbk.eu Abstract We present a supervised learning approach to cross-lingual textual entailment that explores statistical word alignment models to predict entailment relations between sentences writ- ten in different languages. Our approach is language independent, and was used to participate in the CLTE task (Task#8) or- ganized within Semeval 2013 (Negri et al., 2013). The four runs submitted, one for each language combination covered by the test data (i.e. Spanish/English, German/English, French/English and Italian/English), achieved encouraging results. In terms of accuracy, performance ranges from 38.8% (for Ger- man/English) to 43.2% (for Italian/English). On the Italian/English and Spanish/English test sets our systems ranked second among five participants, close to the top results (re- spectively 43.4% and 45.4%). 1 Introduction Cross-lingual textual entailment (CLTE) is an ex- tension of the Textual Entailment task (Dagan and Glickman, 2004) that consists in deciding, given two texts T and H written in different languages (respectively called text and hypothesis), if H can be inferred from T (Mehdad et al., 2010). In the case of SemEval 2013, the task is formulated as a multi-class classification problem in which there are four possible relations between T and H: for- ward (T → H), backward (T ← H), bidirectional (T ↔ H) and “no entailment”. Targeting the identification of semantic equiva- lence and information disparity between topically related sentences, CLTE recognition can be seen as a core task for a number of cross-lingual applications. Among others, multilingual content synchronization has been recently proposed as an ideal framework for the exploitation of CLTE components and the in- tegration of semantics and machine translation (MT) technology (Mehdad et al., 2011; Mehdad et al., 2012b; Bronner et al., 2012; Monz et al., 2011). In the last few years, several methods have been proposed for CLTE. These can be roughly divided in two main groups (Negri et al., 2012): i) those us- ing a pivoting strategy by translating H into the lan- guage of T and then using monolingual TE compo- nents1, and those directly using cross-lingual strate- gies. Among this second group, several sources of cross-lingual knowledge have been used, such as dictionaries (Kouylekov et al., 2012; Perini, 2012), phrase and paraphrase tables (Mehdad et al., 2012a), GIZA++ (Och and Ney, 2003) word alignment mod- els (W¨aschle and Fendrich, 2012), MT of sub- segments (Espl`a-Gomis et al., 2012), or semantic Wordnets (Castillo, 2011). In this work we propose a CLTE detection method based on a new set of features using word align- ment as a source of cross-lingual knowledge. This set, which is richer than the one by (W¨aschle and Fendrich, 2012), is aimed not only at grasping infor- mation about the proportion of aligned words, but also about the distribution of the …

Content truncated to 500 words