Maximum Entropy based Rule Selection Model for Syntax-based Statistical Machine Translation

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 89–97, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Maximum Entropy based Rule Selection Model for Syntax-based Statistical Machine Translation Qun Liu1 and Zhongjun He1,2 and Yang Liu1 and Shouxun Lin1 1Key Laboratory of Intelligent Information Processing Institute of Computing Technology, Chinese Academy of Sciences Beijing, 100190, China 2Graduate University of Chinese Academy of Sciences Beijing, 100049, China {liuqun,zjhe,yliu,sxlin}@ict.ac.cn Abstract This paper proposes a novel maximum en- tropy based rule selection (MERS) model for syntax-based statistical machine transla- tion (SMT). The MERS model combines lo- cal contextual information around rules and information of sub-trees covered by variables in rules. Therefore, our model allows the de- coder to perform context-dependent rule se- lection during decoding. We incorporate the MERS model into a state-of-the-art linguis- tically syntax-based SMT model, the tree- to-string alignment template model. Experi- ments show that our approach achieves signif- icant improvements over the baseline system. 1 Introduction Syntax-based statistical machine translation (SMT) models (Liu et al., 2006; Galley et al., 2006; Huang et al., 2006) capture long distance reorderings by us- ing rules with structural and linguistical information as translation knowledge. Typically, a translation rule consists of a source-side and a target-side. How- ever, the source-side of a rule usually corresponds to multiple target-sides in multiple rules. Therefore, during decoding, the decoder should select a correct target-side for a source-side. We call this rule selec- tion. Rule selection is of great importance to syntax- based SMT systems. Comparing with word selec- tion in word-based SMT and phrase selection in phrase-based SMT, rule selection is more generic and important. This is because that a rule not only contains terminals (words or phrases), but also con- NP DNP NP X 1 DEG NPB NN X 2 NN NP DNP NP X 1 DEG NPB NN X 2 NN X 1 X 2 levels X 2 standard of X 1 Figure 1: Example of translation rules tains nonterminals and structural information. Ter- minals indicate lexical translations, while nontermi- nals and structural information can capture short or long distance reorderings. See rules in Figure 1 for illustration. These two rules share the same syntactic tree on the source side. However, on the target side, either the translations for terminals or the phrase re- orderings for nonterminals are quite different. Dur- ing decoding, when a rule is selected and applied to a source text, both lexical translations (for terminals) and reorderings (for nonterminals) are determined. Therefore, rule selection affects both lexical transla- tion and phrase reordering. However, most of the current syntax-based sys- tems ignore contextual information when they se- lecting rules during decoding, especially the infor- mation of sub-trees covered by nonterminals. For example, the information of X 1 and X 2 is not recorded when the rules in Figure 1 extracted from the training examples in Figure 2. This makes the decoder hardly distinguish the two rules. Intuitively, information of sub-trees covered by nonterminals as well as contextual …

Content truncated to 500 words