CSReader at SemEval-2018 Task 11: Multiple Choice Question Answering as Textual Entailment

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 1053–1057 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics CSReader at SemEval-2018 Task 11: Multiple Choice Question Answering as Textual Entailment Zhengping Jiang Laboratory of Machine Perception Peking University tony.jiang.zhengping@gmail.com Qi Sun Laboratory of Machine Perception Peking University 1500012917@pku.edu.cn Abstract In this document we present an end-to-end machine reading comprehension system that solves multiple choice questions with a tex- tual entailment perspective. Since some of the knowledge required is not explicitly men- tioned in the text, we try to exploit common sense knowledge by using pretrained word em- beddings during contextual embeddings and by dynamically generating a weighted repre- sentation of related script knowledge. In the model two kinds of prediction structure are en- sembled, and the final accuracy of our system is 10 percent higher than the naiive baseline. 1 Introduction K. M. Arivuchelvan et al.(2017) stated that ma- chine reading comprehension can be defined as a task that deals with the automatic understand- ing of texts. In their paper, it was also men- tioned that machine comprehension can be eval- uated by two methods, namely (1) translating the text into formal language representations and eval- uating it using structured queries. (2) evaluat- ing it through natural language questions. Re- cently a lot of datasets are available for eval- uating machine reading comprehension systems, for example, there are SQuAD(Rajpurkar et al., 2016) and the MCTest(Richardson et al., 2013). On many of these datasets human-like perfor- mance has been achieved. However, one of the biggest challenges in machine comprehension is how to provide common sense knowledge regard- ing daily events to machines(Mostafazadeh et al., 2016). The SemEval2018-Task11(Ostermann et al., 2018) provides a dataset containing ques- tions that can only be answered with the help of common sense knowledge.To address this prob- lem, we first propose a model to solve normal reading comprehension problems and then try to modify the model to embody common sense knowledge. In section two, we give a brief in- troduction to ideas and models that might be use- ful to a comprehensive understanding of our work. Section three carefully describes our model imple- mentation, and why we chose this kind of model structure. And section four briefly examines the datasets used. Section five provides a simple eval- uation of our result. Finally, our conclusion can be found in section six. 2 Background Knowledge In this section, we present some basic knowledge required for a comprehensive understanding of our model. We first give a basic introduction to RNN models and the implementation of GRU Cell, then cast a little glance upon the textual entailment problems. 2.1 Recurent Neural Network Lipton et al.(2015) stated that Recurrent neural networks (RNNs) are “connectionist models with the ability to selectively pass information across sequence steps, while processing sequential data one element at a time.” An RNN model captures features of a sequence by updating a hidden set of variables at every input element, as illustrated in figure …

Content truncated to 500 words