A Word-Class Approach to Labeling PSCFG Rules for Machine Translation

Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics, pages 1–11, Portland, Oregon, June 19-24, 2011. c⃝2011 Association for Computational Linguistics A Word-Class Approach to Labeling PSCFG Rules for Machine Translation Andreas Zollmann and Stephan Vogel Language Technologies Institute School of Computer Science Carnegie Mellon University Pittsburgh, PA 15213, USA {zollmann,vogel+}@cs.cmu.edu Abstract In this work we propose methods to label probabilistic synchronous context-free gram- mar (PSCFG) rules using only word tags, generated by either part-of-speech analysis or unsupervised word class induction. The proposals range from simple tag-combination schemes to a phrase clustering model that can incorporate an arbitrary number of features. Our models improve translation quality over the single generic label approach of Chiang (2005) and perform on par with the syntacti- cally motivated approach from Zollmann and Venugopal (2006) on the NIST large Chinese- to-English translation task. These results per- sist when using automatically learned word tags, suggesting broad applicability of our technique across diverse language pairs for which syntactic resources are not available. 1 Introduction The Probabilistic Synchronous Context Free Gram- mar (PSCFG) formalism suggests an intuitive ap- proach to model the long-distance and lexically sen- sitive reordering phenomena that often occur across language pairs considered for statistical machine translation. As in monolingual parsing, nonterminal symbols in translation rules are used to generalize beyond purely lexical operations. Labels on these nonterminal symbols are often used to enforce syn- tactic constraints in the generation of bilingual sen- tences and imply conditional independence assump- tions in the translation model. Several techniques have been recently proposed to automatically iden- tify and estimate parameters for PSCFGs (or related synchronous grammars) from parallel corpora (Gal- ley et al., 2004; Chiang, 2005; Zollmann and Venu- gopal, 2006; Liu et al., 2006; Marcu et al., 2006). While all of these techniques rely on word- alignments to suggest lexical relationships, they dif- fer in the way in which they assign labels to non- terminal symbols of PSCFG rules. Chiang (2005) describes a procedure to extract PSCFG rules from word-aligned (Brown et al., 1993) corpora, where all nonterminals share the same generic label X. In Galley et al. (2004) and Marcu et al. (2006), tar- get language parse trees are used to identify rules and label their nonterminal symbols, while Liu et al. (2006) use source language parse trees instead. Zoll- mann and Venugopal (2006) directly extend the rule extraction procedure from Chiang (2005) to heuristi- cally label any phrase pair based on target language parse trees. Label-based approaches have resulted in improvements in translation quality over the sin- gle X label approach (Zollmann et al., 2008; Mi and Huang, 2008); however, all the works cited here rely on stochastic parsers that have been trained on man- ually created syntactic treebanks. These treebanks are difficult and expensive to produce and exist for a limited set of languages only. In this work, we propose a labeling approach that is based merely on part-of-speech analysis of the source or target language (or even both). To- wards …

Content truncated to 500 words