One-shot Learning for Question-Answering in Gaokao History Challenge

Proceedings of the 27th International Conference on Computational Linguistics, pages 449–461 Santa Fe, New Mexico, USA, August 20-26, 2018. 449 One-shot Learning for Question-Answering in Gaokao History Challenge Zhuosheng Zhang1,2,, Hai Zhao1,2,∗ 1Department of Computer Science and Engineering, Shanghai Jiao Tong University 2Key Laboratory of Shanghai Education Commission for Intelligent Interaction and Cognitive Engineering, Shanghai Jiao Tong University, Shanghai, 200240, China zhangzs@sjtu.edu.cn, zhaohai@cs.sjtu.edu.cn Abstract Answering questions from university admission exams (Gaokao in Chinese) is a challenging AI task since it requires effective representation to capture complicated semantic relations be- tween questions and answers. In this work, we propose a hybrid neural model for deep question- answering task from history examinations. Our model employs a cooperative gated neural net- work to retrieve answers with the assistance of extra labels given by a neural turing machine labeler. Empirical study shows that the labeler works well with only a small training dataset and the gated mechanism is good at fetching the semantic representation of lengthy answers. Exper- iments on question answering demonstrate the proposed model obtains substantial performance gains over various neural model baselines in terms of multiple evaluation metrics. 1 Introduction Teaching a machine to pass admission exams is a hot AI challenge which has aroused a growing num- ber of research (Guo et al., 2017; Cheng et al., 2016; Clark, 2015; Fujita et al., 2014; Henaff et al., 2016). Among these studies, deep question-answering (QA) task (Ferrucci et al., 2010; Wang et al., 2014) is especially difficult, aiming to answer complex questions via deep feature learning from multi- source datasets. Recently, a highly challenging deep QA task is from the Chinese University Admission Examination (Gaokao in Chinese), which is a national-wide standard examination for all senior middle school students in China and has been known as its large scale and strictness. This work focuses on comprehensive question-answering in Gaokao history exams as shown in Table 11, which accounts for the major proportion in total scoring and is extremely difficult in the exams. (Cheng et al., 2016) made a preliminary attempt to take up the Gaokao challenge, trying to solve multiple-choice questions via re- trieving and ranking evidence from Wikipedia articles to determine the right choices. Differently, this task is to solve comprehensive questions and has to be based on knowledge representation and semantic computation rather than word form matching in the previous work. Although deep learning methods shine at various natural language processing tasks (Wang et al., 2015; Zhang et al., 2016b; Qin et al., 2017; Cai et al., 2018; Zhang et al., 2018b; Huang et al., 2018; Zhang et al., 2018a; He et al., 2018; Zhang et al., 2018c; Li et al., 2018), they usually rely on a large scale of dataset for effective learning. The concerned task, unfortunately, cannot receive sufficient training data under ordinary circumstances. Different from previous typical QA tasks such as community QA (Zhang et al., 2016a) which can enjoy the advantage of holding a very large known QA pair set, the concerned task is equal …

Content truncated to 500 words