Chinese Couplet Generation with Syntactic Information

Proceedings of the 29th International Conference on Computational Linguistics, pages 6436–6446 October 12–17, 2022. 6436 Chinese Couplet Generation with Syntactic Information Yan Song University of Science and Technology of China clksong@gmail.com Abstract Chinese couplet generation aims to generate a pair of clauses (usually generating a subse- quent clause given an antecedent one) with cer- tain rules (e.g., morphological and syntactical symmetry) adhered and has long been a chal- lenging task with cultural background. To gen- erate high-quality couplet (antecedent) clauses, it normally requires a model to learn the cor- respondences between antecedent and subse- quent clauses under aforementioned rules and constraint of few characters with their concise usage. To tackle this task, previous studies nor- mally directly adopt deep neural networks with- out explicitly taking into account fine-grained analysis of the clauses, in this paper, we pro- pose to enhance Chinese couplet generation by leveraging syntactic information, i.e., part-of- speech (POS) tags and word dependencies. In doing so, we identify word boundaries in the antecedent clause and then use a special atten- tion module to encode the syntactic information over the words for better generating the subse- quent clause. Experimental results on a dataset for Chinese couplet generation illustrate the va- lidity and effectiveness of our approach, which outperforms strong baselines with respect to automatic and manual evaluation metrics.1 1 Introduction A Chinese antithetical couplet is composed of a pair of two sentences (i.e., an antecedent and a subsequent clause) with particular meaning that are usually applied to festivals or certain circumstances, which can be seen as a special type of poetry with a history in China of more than one thousand years.2 Couplets are normally concise yet have profound 1Related code and resources are available at https:// github.com/synlp/ChiCoupletGen. 2People started to write paired sentences in Pre-Qin period over two thousand years ago, such as those ones widely used in 诗经 (Classic of Poetry) around 500 B.C. The first “special- ized” couplet emerged as recorded in 蜀祷杌 (Shu Tao Wu), which was composed in the Song Dynasty around 964 A.D. Figure 1: An example of a Chinese couplet pair marked with word dependencies and English translations. and expressive ability to convey a variety of mean- ings, thoughts and emotions with limited number of characters. Among all interesting characteristics, the most unique one of couplets is their dueling pattern that the antecedent and subsequent clauses should have an one-to-one correspondence adhere to strict rules covering multiple aspects including tone, length, word usage and even syntax. For automatic Chinese couplet generation, recent studies follow the conventional encoder-decoder paradigm (Sutskever et al., 2014), with an gen- eral attention mechanism to capture the corre- spondence between the antecedent and subsequent clauses (Zhang et al., 2018; Fan et al., 2019; Gao et al., 2021). Although satisfactory results are ob- tained via their approaches, reasonable granularity matching and explicit structural correspondence are still worthy of improvement with special treat- ment. One important reason is that conventional approaches apply character-level encoding and de- coding, larger granular text …

Content truncated to 500 words