HDU: Cross-lingual Textual Entailment with SMT Features

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 467–471, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics HDU: Cross-lingual Textual Entailment with SMT Features Katharina W¨aschle and Sascha Fendrich Department of Computational Linguistics Heidelberg University Heidelberg, Germany {waeschle, fendrich}@cl.uni-heidelberg.de Abstract We describe the Heidelberg University system for the Cross-lingual Textual Entailment task at SemEval-2012. The system relies on fea- tures extracted with statistical machine trans- lation methods and tools, combining mono- lingual and cross-lingual word alignments as well as standard textual entailment dis- tance and bag-of-words features in a statisti- cal learning framework. We learn separate bi- nary classifiers for each entailment direction and combine them to obtain four entailment relations. Our system yielded the best overall score for three out of four language pairs. 1 Introduction Cross-lingual textual entailment (CLTE) (Mehdad et al., 2010) is an extension of textual entailment (TE) (Dagan and Glickman, 2004). The task of recog- nizing entailment is to determine whether a hypoth- esis H can be semantically inferred from a text T. The CLTE task adds a cross-lingual dimension to the problem by considering sentence pairs, where T and H are in different languages. The SemEval-2012 CLTE task (Negri et al., 2012) asks participants to judge entailment pairs in four language combina- tions1, defining four target entailment relations, for- ward, backward, bidirectional and no entailment. We investigate this problem in a statistical learn- ing framework, which allows us to combine cross- lingual word alignment features as well as common 1Spanish-English (es-en), Italian-English (it-en), French- English (fr-en) and German-English (de-en). monolingual entailment metrics, such as bag-of- words overlap, edit distance and monolingual align- ments on translations of T and H, using standard statistical machine translation (SMT) tools and re- sources. Our goal is to address this task without deep processing components to make it easily portable across languages. We argue that the cross-lingual entailment task can benefit from direct alignments between T and H, since a large amount of bilin- gual parallel data is available, which naturally mod- els synonymy and paraphrasing across languages. 2 Related Work With the yearly Recognizing Textual Entailment (RTE) challenge (Dagan et al., 2006), there has been a lot of work on monolingual TE. We therefore in- clude established monolingual features in our ap- proach, such as alignment scores (MacCartney et al., 2008), edit distance and bag-of-words lexical overlap measures (Kouylekov and Negri, 2010). So far, the only work on CLTE that we are aware of is Mehdad et al. (2010), where the problem is reduced to monolingual entailment using machine transla- tion, and Mehdad et al. (2011), which exploits par- allel corpora for generating features based on phrase alignments as input to an SVM. Our approach com- bines ideas from both, mostly resembling Mehdad et al. (2011). There are, however, several differ- ences; we use word translation probabilities instead of phrase tables and model monolingual and cross- lingual alignment separately. We also include addi- tional similarity measures derived from the MT eval- uation metric Meteor, …

Content truncated to 500 words