YNU-HPCC at Semeval-2018 Task 11: Using an Attention-based CNN-LSTM for Machine Comprehension using Commonsense Knowledge

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 1058–1062 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics YNU-HPCC at SemEval-2018 Task 11: Using an Attention-based CNN-LSTM for Machine Comprehension Using Commonsense Knowledge Hang Yuan, Jin Wang and Xuejie Zhang School of Information Science and Engineering Yunnan University Kunming, P.R. China Contact:xjzhang@ynu.edu.cn Abstract This shared task is a typical question answer- ing (QA) task. Specially, this task must give the answer to the question based on the text provided. The essence of the problem is ac- tually reading comprehension. For each ques- tion, there are two candidate answers, and on- ly one of them is correct. Existing method for this task is to use convolutional neural network (CNN) and recurrent neural network (RNN) or their improved models, such as long short-term memory (LSTM). In this paper, an attention-based CNN-LSTM model is pro- posed for this task. By adding an attention mechanism and combining the two models, the experimental results have been significant- ly improved. The accuracy of our final sub- mission is 0.7143. 1 Introduction Question answering has long been an important re- search topic in the field of natural language pro- cessing. Prior to this, there have been many sim- ilar tasks, and many scholars have made very sig- nificant contributions to the research in this field. Such as the Allen AI Science Challenge on the K- aggle (Schoenick et al., 2016) and the IJCNLP- 2017 shared task 5: Multi-choice Question An- swering in Exams (Yuan et al., 2017). Machine comprehension using commonsense knowledge is required to answer multiple-choice questions based on narrative texts about daily ac- tivities of human beings. The answer to many questions does not appear directly in the text, but requires simple reasoning to achieve. In terms of the nature of the problem, this task can be consid- ered as a binary classification. That is, for each question, the candidate answers are divided into two categories: the correct answers and the wrong answers. In recent years, many achievements have been made in machine comprehension-based question answering. Among the existing methods, the main differences are in the data processing and the ap- plication of the model. A dataset for multi-choice question answering was released by Richardson et al. (2013). Clark (2015) described how to obtain more information from the background knowledge base by introducing the use of background knowl- edge to build the best scene. A large cloze-style dataset using CNN and Daily Mail news articles was created by Hermann et al. (2015). Unlike previous datasets, Rajpurkar et al. (2016) released a machine comprehension-based dataset (SQuAD dataset). It contains over 1M text-question-answer triples crawled from 536 Wikipedia articles, and the questions and answers are structured primarily through crowdsourcing. It also requires people to submit up to five article-based questions and pro- vide the correct answer that has appeared in the original text. For the open-domain QA dataset, it is even more challenging to get answers because it requires simple word matching …

Content truncated to 500 words