Adopting the Word-Pair-Dependency-Triplets with Individual Comparison for Natural Language Inference

Proceedings of the 27th International Conference on Computational Linguistics, pages 414–425 Santa Fe, New Mexico, USA, August 20-26, 2018. 414 Adopting the Word-Pair-Dependency-Triplets with Individual Com- parison for Natural Language Inference Qianlong Du1,2, Chengqing Zong1,2,3, Keh-Yih Su4 1National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Science 2University of Chinese Academy of Science 3CAS Center for Excellence in Brain Science and Intelligence Technology 4Institute of Information Science, Academia Sinica {qianlong.du, cqzong}@nlpr.ia.ac.cn kysu@iis.sinica.edu.tw Abstract This paper proposes to perform natural language inference with Word-Pair-Dependency- Triplets. Most previous DNN-based approaches either ignore syntactic dependency among words, or directly use tree-LSTM to generate sentence representation with irrelevant infor- mation. To overcome the problems mentioned above, we adopt Word-Pair-Dependency- Triplets to improve alignment and inference judgment. To be specific, instead of comparing each triplet from one passage with the merged information of another passage, we first pro- pose to perform comparison directly between the triplets of the given passage-pair to make the judgment more interpretable. Experimental results show that the performance of our approach is better than most of the approaches that use tree structures, and is comparable to other state- of-the-art approaches. 1 Introduction Natural language inference (NLI) refers to the following task: given a text passage P (Premise) (which might have more than one sentence) and a text passage H (Hypothesis), whether we can infer H from P, i.e., identifying a specific relationship among entailment, neutral and contradiction. It has many applications such as question answering (Bhaskar et al., 2013; Harabagiu and Hickl, 2006), infor- mation extraction (Romano et al., 2006), machine translation (Pado et al., 2009), automatic text sum- marization (Harabagiu et al., 2007) and so on. Some evaluations about this task have been organized in the past decades, such as the PASCAL Recognizing Textual Entailment (RTE) Challenge (Dagan et al., 2005), SemEval-2014 (Marelli et al., 2014) and RITE (Shima et al., 2011). Many previous approaches adopt statistical frameworks (Heilman et al., 2010; Kouylekov and Magnini, 2005). However, neural network approaches have emerged after Stanford Natural Language Inference (SNLI) dataset (Bowman et al., 2015) was released. Most of them adopt an increasingly complicated network structure to represent text passages, and then predict the relationship between them (Bowman et al., 2016; Liu et al., 2016b). However, P might include extra words which are not directly related to H. Actually, only the words in P that are associated with the words in H should be paid attention to. Those relevant words should be emphasized more while the irrelevant words should be less weighted during decision making. Therefore, some approaches (Parikh et al., 2016; Chen et al., 2017a) adopt attention mechanism to implicitly align the words between two passages to yield a better performance. This idea is very similar to how human make the entailment judgment, and the result shows that it is very effective for performing natural language inference on SNLI corpus in which most words in H can find their corresponding ones in P.  This work is licensed under …

Content truncated to 500 words