Improving Dependency Parsing with Subtrees from Auto-Parsed Data

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 570–579, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Improving Dependency Parsing with Subtrees from Auto-Parsed Data Wenliang Chen, Jun’ichi Kazama, Kiyotaka Uchimoto, and Kentaro Torisawa Language Infrastructure Group, MASTAR Project National Institute of Information and Communications Technology 3-5 Hikari-dai, Seika-cho, Soraku-gun, Kyoto, Japan, 619-0289 {chenwl, kazama, uchimoto, torisawa}@nict.go.jp Abstract This paper presents a simple and effective approach to improve dependency parsing by using subtrees from auto-parsed data. First, we use a baseline parser to parse large-scale unannotated data. Then we ex- tract subtrees from dependency parse trees in the auto-parsed data. Finally, we con- struct new subtree-based features for pars- ing algorithms. To demonstrate the ef- fectiveness of our proposed approach, we present the experimental results on the En- glish Penn Treebank and the Chinese Penn Treebank. These results show that our ap- proach significantly outperforms baseline systems. And, it achieves the best accu- racy for the Chinese data and an accuracy which is competitive with the best known systems for the English data. 1 Introduction Dependency parsing, which attempts to build de- pendency links between words in a sentence, has experienced a surge of interest in recent times, owing to its usefulness in such applications as machine translation (Nakazawa et al., 2006) and question answering (Cui et al., 2005). To ob- tain dependency parsers with high accuracy, super- vised techniques require a large amount of hand- annotated data. While hand-annotated data are very expensive, large-scale unannotated data can be obtained easily. Therefore, the use of large- scale unannotated data in training is an attractive idea to improve dependency parsing performance. In this paper, we present an approach that ex- tracts subtrees from dependency trees in auto- parsed data to improve dependency parsing. The auto-parsed data are generated from large-scale unannotated data by using a baseline parser. Then, from dependency trees in the data, we extract dif- ferent types of subtrees. Finally, we represent subtree-based features on training data to train de- pendency parsers. The use of auto-parsed data is not new. How- ever, unlike most of the previous studies (Sagae and Tsujii, 2007; Steedman et al., 2003) that im- proved the performance by using entire trees from auto-parsed data, we exploit partial information (i.e., subtrees) in auto-parsed data. In their ap- proaches, they used entire auto-parsed trees as newly labeled data to train the parsing models, while we use subtree-based features and employ the original gold-standard data to train the mod- els. The use of subtrees instead of complete trees can be justified by the fact that the accuracy of par- tial dependencies is much higher than that of en- tire dependency trees. Previous studies (McDon- ald and Pereira, 2006; Yamada and Matsumoto, 2003; Zhang and Clark, 2008) show that the accu- racies of complete trees are about 40% for English and about 35% for Chinese, while the accuracies of relations between two words are much higher: about 90% for English and about 85% for …

Content truncated to 500 words