ECNUCS: Recognizing Cross-lingual Textual Entailment Using Multiple Text Similarity and Text Difference Measures

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 118–123, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics ECNUCS: Recognizing Cross-lingual Textual Entailment Using Multiple Text Similarity and Text Difference Measures Jiang ZHAO Department of Computer Science and Technology East China Normal University Shanghai, P.R.China 51121201042@ecnu.cn Man LAN∗ Department of Computer Science and Technology East China Normal University Shanghai, P.R.China mlan@cs.ecnu.edu.cn Zheng-Yu NIU Baidu Inc. Beijing, P.R.China niuzhengyu@baidu.com Abstract This paper presents our approach used for cross-lingual textual entailment task (task 8) organized within SemEval 2013. Cross- lingual textual entailment (CLTE) tries to de- tect the entailment relationship between two text fragments in different languages. We solved this problem in three steps. Firstly, we use a off-the-shelf machine translation (MT) tool to convert the two input texts into the same language. Then after performing a text preprocessing, we extract multiple feature types with respect to surface text and gram- mar. We also propose novel feature types regarding to sentence difference and seman- tic similarity based on our observations in the preliminary experiments. Finally, we adopt a multiclass SVM algorithm for classification. The results on the cross-lingual data collec- tions provided by SemEval 2013 show that (1) we can build portable and effective systems across languages using MT and multiple ef- fective features; (2) our systems achieve the best results among the participants on two test datasets, i.e., FRA-ENG and DEU-ENG. 1 Introduction The Cross-lingual Textual Entailment (CLTE) task in SemEval 2013 consists in detecting the entail- ment relationship between two topic-related text fragments (usually called T(ext) and H(ypothesis)) in different languages, which is a cross-lingual ex- tension of TE task in (Dagan and Glickman, 2004). We say T entails H if the meaning of H can be in- ferred from the meaning of T. Mehdad et al. (2010b) firstly proposed this problem within a new challeng- ing application scenario, i.e., content synchroniza- tion. In consideration of the directionality, the task needs to assign one of the following entailment judg- ments to a pair of sentences (1) forward: unidirec- tional entailment from T to H; (2) backward: unidi- rectional entailment from H to T; (3) bidirectional: the two fragments entail each other (i.e., semantic equivalence); (4) non-entailment: there is no entail- ment between T and H. During the last decades, many researchers and communities have paid a lot of attention to resolve the TE detection (e.g., seven times of the Rec- ognizing Textual Entailment Challenge, i.e., from RTE1 to RET7, have been held) since identifying the relationship between two sentences is at the core of many NLP applications, such as text summa- rization (Lloret et al., 2008) or question answer- ing (Harabagiu and Hickl, 2006). For example, in text summarization, a redundant sentence should be omitted from the summary if this sentence can be entailed from other expressions in the summary. CLTE extends those tasks with lingual dimension- ality, where more than one language is involved. Although it …

Content truncated to 500 words