Multilingual Dependency Parsing using Bayes Point Machines

Proceedings of the Human Language Technology Conference of the North American Chapter of the ACL, pages 160–167, New York, June 2006. c⃝2006 Association for Computational Linguistics Multilingual Dependency Parsing using Bayes Point Machines Simon Corston-Oliver Microsoft Research One Microsoft Way Redmond, WA 98052 simonco@microsoft.com Anthony Aue Microsoft Research One Microsoft Way Redmond, WA 98052 anthaue@microsoft.com Kevin Duh Dept. of Electrical Eng. Univ. of Washington Seattle, WA 98195 duh@ee.washington.edu Eric Ringger Computer Science Dept. Brigham Young Univ. Provo, UT 84602 ringger@cs.byu.edu Abstract We develop dependency parsers for Ara- bic, English, Chinese, and Czech using Bayes Point Machines, a training algo- rithm which is as easy to implement as the perceptron yet competitive with large margin methods. We achieve results com- parable to state-of-the-art in English and Czech, and report the first directed depen- dency parsing accuracies for Arabic and Chinese. Given the multilingual nature of our experiments, we discuss some issues regarding the comparison of dependency parsers for different languages. 1 Introduction Dependency parsing is an alternative to constituency analysis with a venerable tradition going back at least two millenia. The last century has seen at- tempts to formalize dependency parsing, particu- larly in the Prague School approach to linguistics (Tesni`ere, 1959; Melˇcuk, 1988). In a dependency analysis of syntax, words directly modify other words. Unlike constituency analysis, there are no intervening non-lexical nodes. We use the terms child and parent to denote the dependent term and the governing term respectively. Parsing has many potential applications, rang- ing from question answering and information re- trieval to grammar checking. Our intended ap- plication is machine translation in the Microsoft Research Treelet Translation System (Quirk et al., 2005; Menezes and Quirk, 2005). This system ex- pects an analysis of the source language in which words are related by directed, unlabeled dependen- cies. For the purposes of developing machine trans- lation for several language pairs, we are interested in dependency analyses for multiple languages. The contributions of this paper are two-fold: First, we present a training algorithm called Bayes Point Machines (Herbrich et al., 2001; Harrington et al., 2003), which is as easy to implement as the per- ceptron, yet competitive with large margin meth- ods. This algorithm has implications for anyone interested in implementing discriminative training methods for any application. Second, we develop parsers for English, Chinese, Czech, and Arabic and probe some linguistic questions regarding depen- dency analyses in different languages. To the best of our knowledge, the Arabic and Chinese results are the first reported results to date for directed depen- dencies. In the following, we first describe the data (Section 2) and the basic parser architecture (Section 3). Section 4 introduces the Bayes Point Machine while Section 5 describes the features for each lan- guage. We conclude with experimental results and discussions in Sections 6 and 7. 2 Data We utilize publicly available resources in Arabic, Chinese, Czech, and English for training our depen- dency parsers. For Czech we used the Prague Dependency Tree- bank version 1.0 (LDC2001T10). This is …

Content truncated to 500 words