Enhancing Chinese Word Segmentation Using Unlabeled Data

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 970–979, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Enhancing Chinese Word Segmentation Using Unlabeled Data Weiwei Sun†‡ and Jia Xu‡ †Department of Computational Linguistics, Saarland University ‡German Research Center for Artificial Intelligence (DFKI) D-66123, Saarbr¨ucken, Germany wsun@coli.uni-saarland.de, Jia.Xu@dfki.de Abstract This paper investigates improving supervised word segmentation accuracy with unlabeled data. Both large-scale in-domain data and small-scale document text are considered. We present a unified solution to include features derived from unlabeled data to a discrimina- tive learning model. For the large-scale data, we derive string statistics from Gigaword to assist a character-based segmenter. In addi- tion, we introduce the idea about transductive, document-level segmentation, which is de- signed to improve the system recall for out-of- vocabulary (OOV) words which appear more than once inside a document. Novel features1 result in relative error reductions of 13.8% and 15.4% in terms of F-score and the recall of OOV words respectively. 1 Introduction Chinese sentences are written in continuous se- quence of characters without explicit delimiters such as space characters. To find the basic language units, i.e. words, segmentation is a necessary initial step for Chinese language processing. Previous research shows that word segmentation models trained on la- beled data are reasonably accurate. In this paper, we investigate improving supervised word segmen- tation with unlabeled data. We distinguish three types of unlabeled data, namely large-scale in-domain data, out-of-domain data and small-scale document text. Both large-scale 1You can download our derived features at http://www.coli.uni-saarland.de/˜wsun/ semi-cws-feats-emnlp11.tgz. in-domain and out-of-domain data is popular for en- hancing NLP tasks. Learning from these two types of unlabeled data normally involves semi-supervised learning. The difference between them is that out- of-domain data is usually used for domain adapta- tion. For a number of NLP tasks, there are relatively large amounts of labeled training data. In this sit- uation, supervised learning can provide competitive results, and it is difficult to improve them any further by using extra unlabeled data. Chinese word seg- mentation is one of this kind of tasks, since several large-scale manually annotated corpora are publicly available. In this work, we first exploit unlabeled in- domain data to improve strong supervised models. We leave domain adaptation for our future work. We introduce the third type of unlabeled data with a transductive learning, document-level view. Many applications of word segmentation involve process- ing a whole document, such as information retrieval. In this situation, the text of the current document can provide additional useful information to seg- ment a sentence. Take the word “氨纶丝/elastane” for example2. As a translated terminology word, it lacks compositionality. Moreover, this word appears rarely in general texts. As a result, if it does not ap- pear in the training data, it is very hard for statis- tical models to recognize this word. Nevertheless, when we deal with an article discussing an elastane company, this word may appear more than once in this article, and the document information can …

Content truncated to 500 words