GIST at SemEval-2018 Task 12: A network transferring inference knowledge to Argument Reasoning Comprehension task

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 773–777 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics GIST at SemEval-2018 Task 12: A network transferring inference knowledge to Argument Reasoning Comprehension task HongSeok Choi, Hyunju Lee Department of Electrical Engineering and Computer Science, Gwangju Institute of Science and Technology, Gwangju, Republic of Korea Data Mining and Computational Biology Laboratory {hongking9,hyunjulee}@gist.ac.kr Abstract This paper describes our GIST team sys- tem that participated in SemEval-2018 Argu- ment Reasoning Comprehension task (Task 12). Here, we address two challenging fac- tors: unstated common senses and two lexi- cally close warrants that lead to contradicting claims. A key idea for our system is full use of transfer learning from the Natural Language Inference (NLI) task to this task. We used En- hanced Sequential Inference Model (ESIM) to learn the NLI dataset. We describe how to use ESIM for transfer learning to choose cor- rect warrant through a proposed system. We show comparable results through ablation ex- periments. Our system ranked 1st among 22 systems, outperforming all the systems more than 10%. 1 Introduction Argument Reasoning Comprehension is a task that choose correct warrant from two options given a claim and a reason. The Argument Reasoning Comprehension is a very important task because “argument comprehension requires not only lan- guage understanding and logic skills, but it also heavily depends on common sense”, as mentioned by Habernal et al. (2018). There are two challeng- ing factors. One is a certain part of an argument is left unstated (Habernal et al., 2018). Because of the unstated part, humans or machines need reasoning ability about that part. Human can re- construct the unstated part depending on common knowledge. However, it has still remained diffi- cult to machines. Another is that “both options are plausible and lexically very close while leading to contradicting claims”, as mentioned by Habernal et al. (2018). To address these factors, we have two assumptions. One is that similar and large datasets may help to address the unstated common sense by learning various cases. Another is that an in- ferrence model to distinguish semantic differences between two sentences may help to choose one of two lexically close warrants that lead to contra- dicting claims. There are two suitable datasets in the Natural Language Inference (NLI) task, Stan- ford NLI (SNLI) (Bowman et al., 2015) and Multi NLI (MNLI) (Williams et al., 2017) datasets. NLI is a task choosing one of relationships (Entail- ment, Contradiction, Neutral) between two sen- tences. Both SNLI and MNLI are very large cor- pus (each 0.5M sentence pairs). In addition, there is a good performance model for the task, En- hanced Sequential Inference Model (ESIM) (Chen et al., 2017). To make use of other datasets for our task, we use transfer learning. About trans- fer learning, Conneau et al. (2017) showed a good precedent, using SNLI dataset. By learning the NLI task, the model can obtain inference knowl- edge. Therefore, we propose a network transfer- ring …

Content truncated to 500 words