Multi-grained Chinese Word Segmentation with Weakly Labeled Data

Proceedings of the 28th International Conference on Computational Linguistics, pages 2026–2036 Barcelona, Spain (Online), December 8-13, 2020 2026 Multi-grained Chinese Word Segmentation with Weakly Labeled Data Chen Gong, Zhenghua Li∗, Bowei Zou, Min Zhang Institute of Artificial Intelligence, School of Computer Science and Technology, Soochow University, China cgong@stu.suda.edu.cn, {zhli13, minzhang}@suda.edu.cn, zou bowei@i2r.a-star.edu.sg Abstract In contrast with the traditional single-grained word segmentation (SWS), where a sentence cor- responds to a single word sequence, multi-grained Chinese word segmentation (MWS) aims to segment a sentence into multiple word sequences to preserve all words of different granularities. Due to the lack of manually annotated MWS data, previous work train and tune MWS models only on automatically generated pseudo MWS data. In this work, we further take advantage of the rich word boundary information in existing SWS data and naturally annotated data from dictio- nary example (DictEx) sentences, to advance the state-of-the-art MWS model based on the idea of weak supervision. Particularly, we propose to accommodate two types of weakly labeled data for MWS, i.e., SWS data and DictEx data by employing a simple yet competitive graph-based parser with local loss. Besides, we manually annotate a high-quality MWS dataset according to our newly compiled annotation guideline, consisting of over 9,000 sentences from two types of texts, i.e., canonical newswire (NEWS) and non-canonical web (BAIKE) data for better evalua- tion. Detailed evaluation shows that our proposed model with weakly labeled data significantly outperforms the state-of-the-art MWS model by 1.12 and 5.97 on NEWS and BAIKE data in F1. 1 Introduction As a preliminary but critical processing step for Chinese language processing, word segmentation (WS) has been extensively studied for decades and made great progress (Zheng et al., 2013; Pei et al., 2014; Zhang et al., 2016; Yang et al., 2019; He et al., 2020). However, most of previous works adopt the single-grained word segmentation (SWS) formulation, where a sentence corresponds to a single word sequence according to some pre-defined annotation guidelines. As shown in Figure 1 (left), the SWS annotations of the sentence are different according to the guidelines of Penn Chinese Treebank (CTB) (Xue et al., 2005), the People Daily Corpus of the Peking University (PPD) (Yu et al., 2003), and the Microsoft Research WS Corpus (MSR) (Huang et al., 2006). This is largely due to the fact that the boundary between words is usually subtle and vague (Jernudd and Shapiro, 1989) and there are various underlying linguistic theories. Sproat et al. (1987) show that the consensus ratio over word boundaries is only 76% even among Chinese native speakers without training on any guideline. In order to guarantee the annotation consistency, people turn to detailed annotation guidelines ac- cording to specific tasks or applications. For example, CTB usually prefers fine-grained words over coarse-grained words to facilitate further annotation of syntax and semantics, while PPD accommodates more coarse-grained words with information extraction and retrieval tasks in mind. This poses a strong challenge in Chinese WS since words of different granularities are necessary for a variety of …

Content truncated to 500 words