Self-Training PCFG Grammars with Latent Annotations Across Languages

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 832–841, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Self-Training PCFG Grammars with Latent Annotations Across Languages Zhongqiang Huang1 1Laboratory for Computational Linguistics and Information Processing Institute for Advanced Computer Studies University of Maryland, College Park zqhuang@umiacs.umd.edu Mary Harper1,2 2Human Language Technology Center of Excellence Johns Hopkins University mharper@umiacs.umd.edu Abstract We investigate the effectiveness of self- training PCFG grammars with latent anno- tations (PCFG-LA) for parsing languages with different amounts of labeled training data. Compared to Charniak’s lexicalized parser, the PCFG-LA parser was more ef- fectively adapted to a language for which parsing has been less well developed (i.e., Chinese) and benefited more from self- training. We show for the first time that self-training is able to significantly im- prove the performance of the PCFG-LA parser, a single generative parser, on both small and large amounts of labeled train- ing data. Our approach achieves state- of-the-art parsing accuracies for a single parser on both English (91.5%) and Chi- nese (85.2%). 1 Introduction There is an extensive research literature on build- ing high quality parsers for English (Collins, 1999; Charniak, 2000; Charniak and Johnson, 2005; Petrov et al., 2006), however, models for parsing other languages are less well developed. Take Chi- nese for example; there have been several attempts to develop accurate parsers for Chinese (Bikel and Chiang, 2000; Levy and Manning, 2003; Petrov and Klein, 2007), but the state-of-the-art perfor- mance, around 83% F measure on Penn Chinese Treebank (achieved by the Berkeley parser (Petrov and Klein, 2007)) falls far short of performance on English (∼90-92%). As pointed out in (Levy and Manning, 2003), there are many linguistic dif- ferences between Chinese and English, as well as structural differences between their corresponding treebanks, and some of these make it a harder task to parse Chinese. Additionally, the fact that the available treebanked Chinese materials are more limited than for English also increases the chal- lenge of building high quality Chinese parsers. Many of these differences would also tend to apply to other less well investigated languages. In this paper, we focus on English and Chinese because the former is a language for which ex- tensive parsing research has been conducted while the latter is a language that has been less exten- sively studied. We adapt and improve the Berke- ley parser, which learns PCFG grammars with la- tent annotations, and show through comparative studies that this parser significantly outperforms Charniak’s parser, which was initially developed for English and subsequently ported to Chinese. We focus on answering two questions: how well does a parser perform across languages and how much does it benefit from self-training? The first question is of special interest when choosing a parser that is designed for one language and adapting it to another less studied language. We improve the PCFG-LA parser by adding a language-independent method for handling rare words and adapt it to another language, Chinese, by creating a method to better model …

Content truncated to 500 words