Improving Cross-Domain Chinese Word Segmentation with Word Embeddings

Proceedings of NAACL-HLT 2019, pages 2726–2735 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 2726 Improving Cross-Domain Chinese Word Segmentation with Word Embeddings Yuxiao Ye1∗, Yue Zhang2∗†,Weikang Li3, Likun Qiu1†, Jian Sun1 1Alibaba Group, 2Peking University, 3Westlake University {yuxiao.yyx, likun.qlk, jian.sun}@alibaba-inc.com yue.zhang@wias.org.cn wavejkd@pku.edu.cn Abstract Cross-domain Chinese Word Segmentation (CWS) remains a challenge despite recent progress in neural-based CWS. The limited amount of annotated data in the target do- main has been the key obstacle to a satisfac- tory performance. In this paper, we propose a semi-supervised word-based approach to im- proving cross-domain CWS given a baseline segmenter. Particularly, our model only de- ploys word embeddings trained on raw text in the target domain, discarding complex hand- crafted features and domain-specific dictionar- ies. Innovative subsampling and negative sam- pling methods are proposed to derive word embeddings optimized for CWS. We conduct experiments on five datasets in special do- mains, covering domains in novels, medicine, and patent. Results show that our model can obviously improve cross-domain CWS, espe- cially in the segmentation of domain-specific noun entities. The word F-measure increases by over 3.0% on four datasets, outperform- ing state-of-the-art semi-supervised and unsu- pervised cross-domain CWS approaches with a large margin. We make our code and data available on Github. 1 Introduction Chinese Word Segmentation (CWS) is the first step for many Chinese Natural Language Process- ing (NLP) tasks (Cai and Zhao, 2016; Zhao et al., 2017). Approaches to CWS could be categorized into two categories: character-based and word- based. The former treats CWS as a sequence la- beling problem, labeling each character in a se- quence with B/I/E/S (Beginning, Internal, End, Single) labels (Tseng et al., 2005). Traditional character-based approaches often use Conditional work.∗ Yuxiao Ye and Yue Zhang contributed equally to this †Corresponding authors. Random Fields (CRF) models to label sequences, with complex hand-crafted discrete features (Peng et al., 2004; Tseng et al., 2005). Unlike char- acter based CWS, word-based CWS operates on a word-level, directly exploiting word-level fea- tures. Typical CRF models are replaced with semi-CRF models, in which labels are assigned to subsequences instead of characters (Sarawagi and Cohen, 2005; Liu et al., 2014). Transition- based approaches have also been used to exploit larger feature contexts (Zhang and Clark, 2007). More recent approaches exploit neural networks including Recurrent Neural Networks (RNN) to replace hand-crafted discrete features with real- valued features (Cai and Zhao, 2016; Chen et al., 2015, 2017). Existing studies have achieved satisfactory re- sults for in-domain CWS, with F-scores over 96.0% in the newspaper domain (Chen et al., 2017). Nevertheless, cross-domain CWS remains a big challenge (Liu et al., 2014; Liu and Zhang, 2012). The main reason is the lack of anno- tated data in the target domain, which makes su- pervised approaches less useful. To tackle this problem, some unsupervised and semi-supervised approaches have been proposed. One way is to exploit complex features including character types, lexical features and accessor varieties (Wu et al., 2014), which requires much efforts on fea- …

Content truncated to 500 words