Advancing Multi-Criteria Chinese Word Segmentation Through Criterion Classification and Denoising

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 6460–6476 July 9-14, 2023 ©2023 Association for Computational Linguistics Advancing Multi-Criteria Chinese Word Segmentation Through Criterion Classification and Denoising Tzu-Hsuan Chou∗and Chun-Yi Lin∗and Hung-Yu Kao Intelligent Knowledge Management Lab Institute of Computer Science and Information Engineering National Cheng Kung University Tainan, Taiwan ProFatXuanAll@gmail.com, NE6101050@gs.ncku.edu.tw, hykao@mail.ncku.edu.tw Abstract Recent research on multi-criteria Chinese word segmentation (MCCWS) mainly focuses on building complex private structures, adding more handcrafted features, or introducing com- plex optimization processes. In this work, we show that through a simple yet elegant input- hint-based MCCWS model, we can achieve state-of-the-art (SoTA) performances on sev- eral datasets simultaneously. We further pro- pose a novel criterion-denoising objective that hurts slightly on F1 score but achieves SoTA recall on out-of-vocabulary words. Our result establishes a simple yet strong baseline for fu- ture MCCWS research. Source code is avail- able at https://github.com/IKMLab/ MCCWS. 1 Introduction Chinese word segmentation (CWS) is a prelimi- nary step for performing Chinese NLP tasks. Re- searchers have proposed many CWS datasets to enhance word segmentation performance in dif- ferent text domains. However, due to the diver- gence in linguistic perspectives, the same text pas- sage can be segmented in entirely different ways across datasets. For example, in their written forms, Chinese human names have no spaces in between. Some datasets segment human names into last and first names, while others leave human names as a whole (see Table 1). The simplest way to ad- dress such an issue is through single-criterion CWS (SCCWS) model, i.e., to train different models for different datasets. But the cost of maintaining mul- tiple versions of the same model becomes cumber- some as recent deep learning models get deeper and larger. Thus, recent CWS works started to shift their focuses to multi-criterion Chinese word seg- mentation (MCCWS), which aims to fit one model for all CWS datasets (Chen et al., 2017; He et al., ∗Equally contributed. Dataset Samples Labels PKU 江-泽民 S-BE MSRA 江泽民 BME AS 何-樂-而-不-為 S-S-S-S-S CITYU 何樂而不為 BMMME Table 1: Actual samples from SIGHAN bakeoff 2005 datasets (Emerson, 2005) demonstrating labeling incon- sistency. The hyphen “-” denotes segmentation. Labels are defined in Section 3.1. In the first two rows, the human name 江泽民 (Jiang Zemin) in PKU dataset is segmented into the last name 江 (Jiang) and the first name 泽民 (Zemin), but not in MSRA dataset. In the last two rows, the idiom 何樂而不為 (Why not do some- thing?) is segmented in AS dataset but not in CITYU dataset. More examples can be found in these datasets. 2019; Gong et al., 2019; Huang et al., 2020b,a; Ke et al., 2020; Qiu et al., 2020; Ke et al., 2021). MCCWS can be seen as a multi-task learning problem (Chen et al., 2017) that benefits from leveraging large amounts of heterogeneous data, meanwhile dealing with subtle linguistic diver- gence. Prior works are mainly divided into private- structure-based and input-hint-based models. In a typical SCCWS workflow, …

Content truncated to 500 words