Type-Supervised Domain Adaptation for Joint Segmentation and POS-Tagging

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 588–597, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Type-Supervised Domain Adaptation for Joint Segmentation and POS-Tagging Meishan Zhang†, Yue Zhang‡ , Wanxiang Che†, Ting Liu†∗ †Research Center for Social Computing and Information Retrieval Harbin Institute of Technology, China {mszhang, car, tliu}@ir.hit.edu.cn ‡Singapore University of Technology and Design yue zhang@sutd.edu.sg Abstract We report an empirical investigation on type-supervised domain adaptation for joint Chinese word segmentation and POS-tagging, making use of domain- specific tag dictionaries and only un- labeled target domain data to improve target-domain accuracies, given a set of annotated source domain sentences. Pre- vious work on POS-tagging of other lan- guages showed that type-supervision can be a competitive alternative to token- supervision, while semi-supervised tech- niques such as label propagation are important to the effectiveness of type- supervision. We report similar findings using a novel approach for joint Chinese segmentation and POS-tagging, under a cross-domain setting. With the help of un- labeled sentences and a lexicon of 3,000 words, we obtain 33% error reduction in target-domain tagging. In addition, com- bined type- and token-supervision can lead to improved cost-effectiveness. 1 Introduction With accuracies of over 97%, POS-tagging of WSJ can be treated as a solved problem (Man- ning, 2011). However, performance is still well below satisfactory for many other languages and domains (Petrov et al., 2012; Christodoulopoulos et al., 2010). There has been a line of research on using a tag-dictionary for POS-tagging (Merialdo, 1994; Toutanova and Johnson, 2007; Ravi and Knight, 2009; Garrette and Baldridge, 2012). The idea is compelling: on the one hand, a list of lex- icons is often available for special domains, such as bio-informatics; on the other hand, compiling a ∗Corresponding author. lexicon of word-tag pairs appears to be less time- consuming than annotating full sentences. However, success in type-supervised POS- tagging turns out to depend on several subtle fac- tors. For example, recent research has found that the quality of the tag-dictionary is crucial to the success of such methods (Banko and Moore, 2004; Goldberg et al., 2008; Garrette and Baldridge, 2012). Banko and Moore (2004) found that the accuracies can drop from 96% to 77% when a hand-crafted tag dictionary is replaced with a raw tag dictionary gleaned from data, without any hu- man intervention. These facts indicate that careful considerations need to be given for effective type- supervision. In addition, significant manual work might be required to ensure the quality of lexicons. To compare type- and token-supervised tagging, Garrette and Baldridge (2013) performed a set of experiments by conducting each type of annota- tion for two hours. They showed that for low- resource languages, a tag-dictionary can be rea- sonably effective if label propagation (Talukdar and Crammer, 2009) and model minimizations (Ravi and Knight, 2009) are applied to expand and filter the lexicons. Similar findings were reported in Garrette et al. (2013). Do the above findings carry over to the Chi- nese language? …

Content truncated to 500 words