Addressing Domain Adaptation for Chinese Word Segmentation with Global Recurrent Structure

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 184–193, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Addressing Domain Adaptation for Chinese Word Segmentation with Global Recurrent Structure Shen Huang and Xu Sun and Houfeng Wang∗ Key Laboratory of Computational Linguistics, Ministry of Education School of Electronics Engineering and Computer Science, Peking University Beijing, P.R.China, 100871 huangshenno1,xusun,wanghf@pku.edu.cn Abstract Boundary features are widely used in traditional Chinese Word Segmentation (CWS) methods as they can utilize un- labeled data to help improve the Out-of- Vocabulary (OOV) word recognition per- formance. Although various neural net- work methods for CWS have achieved per- formance competitive with state-of-the-art systems, these methods, constrained by the domain and size of the training corpus, do not work well in domain adaptation. In this paper, we propose a novel BLSTM- based neural network model which incor- porates a global recurrent structure de- signed for modeling boundary features dy- namically. Experiments show that the proposed structure can effectively boost the performance of Chinese Word Seg- mentation, especially OOV-Recall, which brings benefits to domain adaptation. We achieved state-of-the-art results on 6 do- mains of CNKI articles, and competitive results to the best reported on the 4 do- mains of SIGHAN Bakeoff 2010 data. 1 Introduction Since Chinese writing system does not have ex- plicit word delimiters, word segmentation be- comes an essential first step for further Chinese language processing. In recent years, Chinese Word Segmentation (CWS) has experienced great advancement. One mainstream method is to re- gard word segmentation task as a sequence label- ing problem (Xue, 2003; Peng et al., 2004) where each character is assigned a tag indicating its po- sition in the word. This method has been proved ∗Corresponding author effective as it turns word segmentation into a struc- tured discriminative learning task which can be handled by supervised learning algorithms such as Maximum Entropy (ME) (Berger et al., 1996) and Conditional Random Fields (CRF) (Lafferty et al., 2001). Furthermore, rich features can be incorpo- rated into these systems to improve their perfor- mances and most state-of-the-art systems are still based on feature-based models. Recently, neural network models are drawing increasing attention in Natural Language Process- ing (NLP) tasks. They significantly reduced fea- ture engineering effort and achieved competitive or state-of-the-art results in many NLP tasks. Col- lobert et al. (2011) developed a general neural net- work architecture for sequence labeling tasks. Fol- lowing this work, many neural network model- s (Zheng et al., 2013; Pei et al., 2014; Chen et al., 2015a,b) have been applied to CWS and some ap- proached state-of-the-art performance. However, these neural network models, as well as other supervised methods, do not work well in domain adaptation. In recent years, manual- ly annotated training corpus mostly come from the news domain. When it shifts to other do- mains such as literature or medicine, where there are many domain-related words that rarely ap- pear in other domains, Out-of-Vocabulary (OOV) word recognition becomes an important problem. Moreover, different …

Content truncated to 500 words