Decompose, Fuse and Generate: A Formation-Informed Method for Chinese Definition Generation

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5524–5531 June 6–11, 2021. ©2021 Association for Computational Linguistics 5524 Decompose, Fuse and Generate: A Formation-Informed Method for Chinese Definition Generation Hua Zheng1,2∗, Damai Dai1,2∗, Lei Li1,2, Tianyu Liu1,2, Zhifang Sui1,2,3, Baobao Chang1,2,3, Yang Liu1,2† 1Department of Computer Science and Technology, Peking University 2Key Lab of Computational Linguistics (MOE), Peking University 3Peng Cheng Laboratory, China {zhenghua,daidamai,tianyu0421,szf,chbb,liuyang}@pku.edu.cn lilei@stu.pku.edu.cn Abstract In this paper, we tackle the task of Definition Generation (DG) in Chinese, which aims at au- tomatically generating a definition for a word. Most existing methods take the source word as an indecomposable semantic unit. However, in parataxis languages like Chinese, word mean- ings can be composed using the word forma- tion process, where a word ("桃花", peach- blossom) is formed by formation components ("桃", peach; "花", flower) using a formation rule (Modifier-Head). Inspired by this process, we propose to enhance DG with word forma- tion features. We build a formation-informed dataset and propose a model DeFT, which Decomposes words into formation features, dynamically Fuses different features through a gating mechanism, and generaTes word def- initions. Experimental results show that our method is both effective and robust.1 1 Introduction Definition Generation (DG) aims at automatically generating an explanatory text for a word. This task is of practical importance to assist dictionary construction, especially in highly productive lan- guages like Chinese (Yang et al., 2020). Most ex- isting methods take the source word as an indecom- posable lexico-semantic unit, using features like word embedding (Noraset et al., 2017) and con- text (Gadetsky et al., 2018; Ishiwatari et al., 2019). Recently, Yang et al. (2020) and Li et al. (2020) achieve improvement by decomposing the word meaning into different semantic components. In decomposing the word meaning, the word formation process is an intuitive and informative way that has not been explored in DG by far. For parataxis languages like Chinese, a word is formed by formation components, i.e., morphemes, and ∗Equal contribution. †Corresponding author. 1The code is available at https://github.com/ Hunter-DDM/DeFT-naacl2021. �2: �� (spend) �2: ��� (vainly) �1: �� (flower) �1: ��� (white) Modifier-Head Adverb-Verb ���������� (White flower.) ���� ������ (Vainly spend.) Formation Rule Word Definition �� Morphemes: Definitions Word Figure 1: Word formation process for the polysemous "白花". With morphemes and a formation rule spec- ified, the process can construct and distinguish each meaning. Definitions are simplified for ease of reading. a formation rule. As shown in Figure 1, the poly- semous word "白花" holds two meanings "白花1" and "白花2", which can be distinguished by dif- ferent morphemes ("白1;花1" vs. "白2;花2") and different rules (Modifier-Head vs. Adverb-Verb). Such intuitive formation process can clearly and unambiguously construct the word meaning. Inspired by the word formation process in Chi- nese, we propose to enhance DG with formation features. First, we build a formation-informed dataset under expert annotations. Next, we design a DG model DeFT, which Decomposes words into formation features, Fuses different features through …

Content truncated to 500 words