Domain Adaptation for CRF-based Chinese Word Segmentation using Free Annotations

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 864–874, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Domain Adaptation for CRF-based Chinese Word Segmentation using Free Annotations Yijia Liu †‡, Yue Zhang †, Wanxiang Che ‡, Ting Liu ‡, Fan Wu † †Singapore University of Technology and Design ‡Research Center for Social Computing and Information Retrieval Harbin Institute of Technology, China {yjliu,car,tliu}@ir.hit.edu.cn {yue zhang,fan wu}@sutd.edu.sg Abstract Supervised methods have been the domi- nant approach for Chinese word segmen- tation. The performance can drop signif- icantly when the test domain is different from the training domain. In this paper, we study the problem of obtaining par- tial annotation from freely available data to help Chinese word segmentation on dif- ferent domains. Different sources of free annotations are transformed into a unified form of partial annotation and a variant CRF model is used to leverage both fully and partially annotated data consistently. Experimental results show that the Chi- nese word segmentation model benefits from free partially annotated data. On the SIGHAN Bakeoff 2010 data, we achieve results that are competitive to the best re- ported in the literature. 1 Introduction Statistical Chinese word segmentation gains high accuracies on newswire (Xue and Shen, 2003; Zhang and Clark, 2007; Jiang et al., 2009; Zhao et al., 2010; Sun and Xu, 2011). However, man- ually annotated training data mostly come from the news domain, and the performance can drop severely when the test data shift from newswire to blogs, computer forums and Internet literature (Liu and Zhang, 2012). Several methods have been proposed for solv- ing the domain adaptation problem for segmenta- tion, which include the traditional token- and type- supervised methods (Song et al., 2012; Zhang et al., 2014). While token-supervised methods rely on manually annotated target-domain sentences, type-supervised methods leverage manually as- sembled domain-specific lexicons to improve target-domain segmentation accuracies. Both . 浦 . 东 . 开 . 发 . 与 . 法 . 制 . 建 . 设 . b . b . b . b . b . b . b . b . b . m . m . m . m . m . m . m . m . m . e . e . e . e . e . e . e . e . e . s . s . s . s . s . s . s . s . s Figure 1: The segmentation problem, illustrated using the sentence “浦东 (Pudong) 开发 (devel- opment) 与 (and) 法制 (legal) 建设 (construc- tion)”. Possible segmentation labels are drawn un- der each character, where b, m, e, s stand for the beginning, middle, end of a multi-character word, and a single character word, respectively. The path shows the correct segmentation by choosing one label for each character. methods are competitive given the same amount of annotation effects (Garrette and Baldridge, 2012; Zhang et al., 2014). However, obtaining manually annotated data can be expensive. On the other hand, …

Content truncated to 500 words