Synonym Knowledge Enhanced Reader for Chinese Idiom Reading Comprehension

Proceedings of the 28th International Conference on Computational Linguistics, pages 3684–3695 Barcelona, Spain (Online), December 8-13, 2020 3684 Synonym Knowledge Enhanced Reader for Chinese Idiom Reading Comprehension Siyu Long1∗, Ran Wang1∗, Kun Tao1, Jiali Zeng2, Xin-Yu Dai1 1National Key Laboratory for Novel Software Technology, Nanjing University 2Smart Platform Product Department of Tencent Inc., China {longsy, wangr, taok}@smail.nju.edu.cn lemonzeng@tencent.com, daixinyu@nju.edu.cn Abstract Machine reading comprehension (MRC) is the task that asks a machine to answer questions based on a given context. For Chinese MRC, due to the non-literal and non-compositional semantic characteristics, Chinese idioms pose unique challenges for machines to understand. Previous studies tend to treat idioms separately without fully exploiting the relationship among them. In this paper, we first define the concept of literal meaning coverage to measure the consistency between semantics and literal meanings for Chinese idioms. With the definition, we prove that the literal meanings of many idioms are far from their semantics, and we also verify that the syn- onymic relationship can mitigate this inconsistency, which would be beneficial for idiom com- prehension. Furthermore, to fully utilize the synonymic relationship, we propose the synonym knowledge enhanced reader. Specifically, for each idiom, we first construct a synonym graph according to the annotations from a high-quality synonym dictionary or the cosine similarity be- tween the pre-trained idiom embeddings and then incorporate the graph attention network and gate mechanism to encode the graph. Experimental results on ChID, a large-scale Chinese idiom reading comprehension dataset, show that our model achieves state-of-the-art performance.1 1 Introduction Machine reading comprehension is the task that asks a machine to answer questions based on a given context. Although various advanced neural models have been proposed in recent years (Hermann et al., 2015; Chen et al., 2016). Chinese reading comprehension is still a challenging task (He et al., 2018; Sun et al., 2020). One of the reasons is the existence of Chinese idioms, also known as “chengyu”, whose non-literal and non-compositional characteristics make them difficult for machines to understand (Shao et al., 2018; Zheng et al., 2019). Recently, some studies (Jiang et al., 2018; Zheng et al., 2019) claimed that good idiom representations are essential for idiom comprehension. Meanwhile, many methods have been proposed to obtain idiom representations, which can be divided into two categories: First, a straightforward way is to treat each Chinese idiom as a single token and obtain its representation by training on a large-scale corpus (Song et al., 2018; Zheng et al., 2019). However, compared to common words, Chinese idioms appear less frequently. These methods are likely not sufficient to learn good enough idiom representations (Bah- danau et al., 2017); Second, since most Chinese idioms consist of four Chinese characters, another direct approach is to treat each Chinese idiom as a regular multi-word expression. In this category, most studies combine the distributional word embeddings (Mikolov et al., 2013a; Mikolov et al., 2013b) to learn rep- resentations of multi-word expressions. Additionally, some methods try to enhance the representations by employing a more powerful compositionality function …

Content truncated to 500 words