Natural Language Inference using Neural Network and Tableau Method

Natural Language Inference Using Neural Network and Tableau Method Ayahito Saji1, Daiki Takao1, Yoshihide Kato2, Shigeki Matsubara1,2 1Graduate School of Informatics, Nagoya University 2Information and Communications, Nagoya University Furo-cho, Chikusa-ku, Nagoya, 464-8601, Japan saji.ayahito.y7@s.mail.nagoya-u.ac.jp Abstract Natural language inference (NLI) is the task of identifying the inferential relation between a text pair. Although neural-based approaches have recently achieved high performance on the NLI task, they are unable to explain their reasoning processes. Symbolic ap- proaches, on the other hand, have the ad- vantage that the reasoning process is under- standable to humans. This paper proposes a method for integrating a neural NLI model and tableau proof system, with the latter ex- plaining the reasoning processes. The stan- dard tableau method consists of decomposing logical formulas by applying inferential rules, and checking whether or not there exists a val- uation that satisfies the given constraints. Un- like the standard tableau method, our method uses dependency structures as its ingredients rather than logical formulas, and it employs a neural NLI model for the latter process. To an- alyze the behavior of a neural NLI model, we conducted an experiment on the neural NLI model alone and proposed method using the SNLI corpus. 1 Introduction Natural language inference (NLI) is the task of iden- tifying the inferential relation between a text pair: premise and hypothesis. If a hypothesis can be in- ferred from a premise using logical and common sense knowledge, it is judged as entailment; if they are incompatible, it is judged as contradiction; and if neither of them is the case, it is judged as neu- tral. It is expected to be used in a wide range of fields such as question answering, information re- trieval, and text summarization. In recent years, neural-based approaches have achieved high performance on the NLI task. For ex- ample, Chen et al. (2017) proposed a model based on word embedding and bidirectional LSTM (Hochre- iter and Schmidhuber, 1997). Dispite its simplic- ity, the model achieves a high accuracy in experi- ments with the SNLI corpus (Bowman et al., 2015). On the other hand, neural-based approaches have the limitation that the models cannot explain the reason- ing processes that lead to judgment. The model is a black box, and it is difficult to analyze what kind of inference was performed. Furthermore, Gururangan et al. (2018) demonstrated that NLI datasets such as the SNLI corpus and MultiNLI corpus (Williams et al., 2018) have a hidden bias in that inferential re- lations can be determined only from a hypothesis, and they highlighted the risk that neural models are simply identifying inferential relations based on the biases. On the other hand, symbolic approaches to the NLI task have been proposed. These approaches have the advantage that the reasoning process lead- ing to the result is understandable to humans, un- like the neural model approach. In addition, sym- bolic manipulations in these approaches are gen- erally founded on formal logic or linguistic analy- ses, which allow us to understand the reasoning …

Content truncated to 500 words