Umelb: Cross-lingual Textual Entailment with Word Alignment and String Similarity Features

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 133–137, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics Umelb: Cross-lingual Textual Entailment with Word Alignment and String Similarity Features Yvette Graham Bahar Salehi Timothy Baldwin Department of Computing and Information Systems The University of Melbourne {ygraham,bsalehi,tbaldwin}@unimelb.edu.au Abstract This paper describes The University of Mel- bourne NLP group submission to the Cross- lingual Textual Entailment shared task, our first tentative attempt at the task. The ap- proach involves using parallel corpora and au- tomatic word alignment to align text fragment pairs, and statistics based on unaligned words as features to classify items as forward and backward before a compositional combination into the final four classes, as well as exper- iments with additional string similarity fea- tures. 1 Introduction Cross-lingual Textual Entailment (CLTE) (Negri et al., 2012) proposes the task of automatically iden- tifying the kind of relation that exists between pairs of semantically-related text fragments written in two distinct languages, a variant of the traditional Rec- ognizing Textual Entailment (RTE) task (Bentivogli et al., 2009; Bentivogli et al., 2010). The task tar- gets the cross-lingual content synchronization sce- nario proposed in Mehdad et al. (2010, 2011). Com- positional classification can be used by training two distinct binary classifiers for forward and backward entailment classification, before combining labels into the four final entailment categories that now in- clude bidirectional and no entailment labels. The most similar previous work to this work is the cross- lingual approach of the FBK system (Mehdad et al., 2012) from Semeval 2012 (Negri et al., 2012), in which the entailment classification is obtained without translating T1 into T2 for the Spanish– English language pair. We apply the cross-lingual approach to German–English and instead of cross- lingual matching features, we use Giza++ (Och et al., 1999) and Moses (Koehn et al., 2007) to auto- matically word align text fragment pairs to compute statistics of unaligned words. In addition, we in- clude some additional experiments using string sim- ilarity features. 2 Compositional Classification Given a pair of topically related fragments, T1 (Ger- man) and T2 (English), we automatically annotate it with one of the following entailment labels: bidi- rectional, forward, backward, no entailment. We take the compositional approach and separately train a forward, as well as a backward binary classifier. Each classifier is run separately on the set of text fragment pairs to produce two binary labels for for- ward and backward entailment. The two sets of la- bels are logically combined to produce a final clas- sification for each test pair of forward, backward, bidirectional or no entailment. 3 Word Alignment Features The test set of topically-related text fragments, T1 (German) and T2 (English) were added to Europarl German–English parallel text (Koehn, 2005) and Giza++ was used for automatic word alignment in both language directions. Moses (Koehn et al., 2007) was then used for symmetrization with the grow diag final and algorithm. This produces a …

Content truncated to 500 words