Learning and Evaluating Chinese Idiom Embeddings

Proceedings of Recent Advances in Natural Language Processing, pages 1387–1396 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_155 1387 Learning and Evaluating Chinese Idiom Embeddings Minghuan Tan and Jing Jiang School of Computing and Information Systems Singapore Management University mhtan.2017@phdcs.smu.edu.sg jingjiang@smu.edu.sg Abstract We study the task of learning and evaluating Chinese idiom embeddings. We first construct a new evaluation dataset that contains idiom synonyms and antonyms. Observing that ex- isting Chinese word embedding methods may not be suitable for learning idiom embeddings, we further present a BERT-based method that directly learns embedding vectors for individ- ual idioms. We empirically compare represen- tative existing methods and our method. We find that our method substantially outperforms existing methods on the evaluation dataset we have constructed. 1 Introduction Chengyu (成语, literally meaning “set phrases”) are a type of idiomatic expressions in Chinese that usually consist of four Chinese characters. They are mostly derived from ancient Chinese litera- ture and many of them are based on historical sto- ries. The semantic meanings of Chengyu are often non-compositional and sometimes metaphoric. For example, the Chengyu 瓜田李下 literally means “melon field, beneath the plums,” but its idiomatic meaning is to warn people to avoid situations where a person may be easily suspected of wrongdoing. Chengyu are commonly used in modern Chinese language, and using computational methods to un- derstand Chengyu plays an important role in Chi- nese language understanding. For example, a re- cent work studied how to improve essay writing with recommending Chinese idioms (Liu et al., 2019), and others studied how to improve read- ing comprehension by correcting usage of Chinese idioms (Wang et al., 2020) and differentiating syn- onyms of Chinese idioms (Long et al., 2020). In this paper, we refer to Chengyu as Chinese idioms, although there are also other types of idioms in Chinese. Recent years have witnessed the success of deep neural networks for many NLP tasks. A central idea behind deep neural networks for NLP is to use dense embedding vectors to represent language units including words, phrases and sentences, and such embeddings have been shown to be useful for many tasks such as sentiment analysis (Yu et al., 2017), question answering (Hao et al., 2017) and machine translation (Zhou et al., 2016). We there- fore believe that it is also desirable to derive embed- ding vectors for Chinese idioms that can accurately capture their semantic meanings. However, it is not clear whether existing methods for Chinese word embeddings are effective in deriving good Chinese idiom embeddings, and there are at least two reasons for this. First, existing Chinese word embedding eval- uation datasets do not have sufficient coverage of idioms. For example, in the commonly used WordSim-240 (Wang et al., 2011) and WordSim- 296 (Chen et al., 2015) datasets for Chinese word relatedness, no idiom is found. More recently, Huang et al. (2019) released a COS960 dataset with similarities of Multiword Expressions (MWEs). Al- though COS960 covers 150 Chinese idioms, this is still a relatively small number, and only 20 MWE pairs …

Content truncated to 500 words