Re-training Monolingual Parser Bilingually for Syntactic SMT

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 854–862, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Re-training Monolingual Parser Bilingually for Syntactic SMT †Shujie Liu*, ‡Chi-Ho Li, ‡Mu Li and ‡Ming Zhou †School of Computer Science and Technology Harbin Institute of Technology, Harbin, China shujieliu@mtlab.hit.edu.cn ‡Microsoft Research Asia, Beijing, China {chl, muli, mingzhou}@microsoft.com Abstract The training of most syntactic SMT approaches involves two essential components, word alignment and monolingual parser. In the current state of the art these two components are mutually independent, thus causing problems like lack of rule generalization, and violation of syntactic correspondence in translation rules. In this paper, we propose two ways of re-training monolingual parser with the target of maximizing the consistency between parse trees and alignment matrices. One is targeted self-training with a simple evaluation function; the other is based on training data selection from forced alignment of bilingual data. We also propose an auxiliary method for boosting alignment quality, by symmetrizing alignment matrices with respect to parse trees. The best combination of these novel methods achieves 3 Bleu point gain in an IWSLT task and more than 1 Bleu point gain in NIST tasks. 1 Introduction There are many varieties in syntactic statistical machine translation (SSMT). Apart from a few attempts to use synchronous parsing to produce the tree structure of both source language (SL) and target language (TL) simultaneously, most SSMT approaches make use of monolingual parser to produce the parse tree(s) of the SL and/or TL sentences, and then link up the information of the two languages through word alignment. In the current state of the art, word aligner and monolingual parser are trained and applied separately. On the one hand, an average word aligner does not consider the syntax information of both languages, and the output links may violate syntactic correspondence. That is, some SL words yielded by a SL parse tree node may not be traced to, via alignment links, some TL words with legitimate syntactic structure. On the other hand, parser design is a monolingual activity and its impact on MT is not well studied (Ambati, 2008). Many good translation rules may thus be filtered by a good monolingual parser. In this paper we will focus on the translation task from Chinese to English, and the string-to-tree SSMT model as elaborated in (Galley et al., 2006). There are two kinds of translation rules in this model, minimal rules, and composed rules, which are composition of minimal rules. The minimal rules are extracted from a special kind of nodes, known as frontier nodes, on TL parse tree. The concept of frontier node can be illustrated by Figure 1, which shows two partial bilingual sentences with the corresponding TL sub-trees and word alignment links. The TL words yielded by a TL parse node can be traced to the corresponding SL words through alignment links. In the diagram, each parse node is represented by a rectangle, …

Content truncated to 500 words