Labeled Alignment for Recognizing Textual Entailment

International Joint Conference on Natural Language Processing, pages 605–613, Nagoya, Japan, 14-18 October 2013. Labeled Alignment for Recognizing Textual Entailment ∗ Xiao-Lin Wang Hai Zhao Bao-Liang Lu Center for Brain-Like Computing and Machine Intelligence MOE-Microsoft Key Laboratory for Intelligent Computing and Intelligent Systems Department of Computer Science and Engineering, Shanghai Jiao Tong University 800 Dongchuan Road, Shanghai 200240, China arthur.xl.wang@gmail.com {zhaohai,blu}@cs.sjtu.edu.cn Abstract Recognizing Textual Entailment (RTE) is to predict whether one text fragment can semantically infer another, which is re- quired across multiple applications of nat- ural language processing. The conven- tional alignment scheme, which is devel- oped for machine translation, only marks the paraphrases and hyponyms to justify the entailment pairs, while provides less support for the non-entailment ones. This paper proposes a novel alignment scheme, named labeled alignment, to address this problem, which introduces negative links to explicitly mark the contradictory ex- pressions to justify the non-entailment pairs. Thus the alignment-based RTE method employing the proposed scheme, compared with those employing the con- ventional one, can gain accuracy improve- ment through actively detecting the signals of non-entailment. The experimental re- sults on the data sets of two shared RTE tasks indicate the implemented system sig- nificantly outperforms both the baseline system and all the other submitted sys- tems. 1 Introduction Textual Entailment (TE) is a directional relation between two text fragments. One natural-language premise, noted as P, entails one natural-language hypothesis, noted as H, if typically a human read- ∗ B. L. Lu and X. L. Wang are supported by the National Natural Science Foundation of China (Grant No. 61272248), the National Basic Research Program of China (Grant No.2013CB329401), and the Science and Tech- nology Commission of Shanghai Municipality (Grant No. 13511500200). H. Zhao is supported by the National Nat- ural Science Foundation of China (Grant No. 60903119 and Grant No. 61170114). ing P would infer that H is most likely true (Da- gan et al., 2006). Recognizing Textual Entailment (RTE) is pro- posed as a generic task that captures the seman- tic inference need across a wide range of natural language processing applications. For example, a question answering system should identify the texts that entail a hypothesized answer, e.g., given the question “What does Peugeot manufacture?”, the text “Chr´etien visited Peugeot’s newly reno- vated car factory” entails the hypothesized answer form “Peugeot manufactures cars” (Dagan et al., 2006). Similarly, in Machine Translation (MT) evaluation, a correct translation should be seman- tically equivalent to the gold translation, that is, both translations should entail each other (Pad´o et al., 2009). RTE has attracted extensive attention ever since it was proposed. A wide range of methods have been proposed, and quite a few success- ful approaches treat RTE as an alignment prob- lem. Alignment is originally developed for MT to bridge two languages (Brown et al., 1993). Alignment is to establish links between the seman- tically equivalent atom expressions in two sen- tences. (Marsi and Krahmer, 2005) first advocates pipelined system architectures that contain dis- tinct alignment components. This latter …

Content truncated to 500 words