Improving Statistical Machine Translation using Lexicalized Rule Selection

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 321–328 Manchester, August 2008 Improving Statistical Machine Translation using Lexicalized Rule Selection Zhongjun He1,2 and Qun 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 {zjhe,liuqun,sxlin}@ict.ac.cn Abstract This paper proposes a novel lexicalized ap- proach for rule selection for syntax-based statistical machine translation (SMT). We build maximum entropy (MaxEnt) mod- els which combine rich context informa- tion for selecting translation rules dur- ing decoding. We successfully integrate the MaxEnt-based rule selection models into the state-of-the-art syntax-based SMT model. Experiments show that our lexical- ized approach for rule selection achieves statistically significant improvements over the state-of-the-art SMT system. 1 Introduction The syntax-based statistical machine translation (SMT) models (Chiang, 2005; Liu et al., 2006; Galley et al., 2006; Huang et al., 2006) use rules with hierarchical structures as translation knowl- edge, which can capture long-distance reorderings. Generally, a translation rule consists of a left-hand- side (LHS) 1and a right-hand-side (RHS). The LHS and RHS can be words, phrases, or even syn- tactic trees, depending on SMT models. Transla- tion rules can be learned automatically from par- allel corpus. Usually, an LHS may correspond to multiple RHS’s in multiple rules. Therefore, in sta- tistical machine translation, the rule selection task is to select the correct RHS for an LHS during de- coding. The conventional approach for rule selection is to use precomputed translation probabilities which c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. 1In this paper, we use LHS and source-side interchange- ably (so are RHS and target-side). are estimated from the training corpus, as well as a n-gram language model which is trained on the tar- get language. The limitation of this method is that it ignores context information (especially on the source-side) during decoding. Take the hierarchi- cal model (Chiang, 2005) as an example. Consider the following rules for Chinese-to-English transla- tion 2: (1) X → ⟨ � X 1 � X 2 , X 2 in X 1 ⟩ (2) X → ⟨ � X 1 � X 2 , at X 1 ’s X 2 ⟩ (3) X → ⟨ � X 1 � X 2 , with X 2 of X 1 ⟩ These rules have the same source-side, and all of them can pattern-match all the following source phrases: (a) � in [�� economic ��]1 field ’s� [��]2 cooperation [cooperation]2 in [the economic field]1 (b) � at [��]1 today ’s� [�� meeting �]2 on at [today]1 ’s [meeting]2 (c) � with [��]1 people ’s� [�� support �]2 under with [the support]2 of [the people]1 Given a source phrase, how does the decoder know which rule is suitable? In fact, rule (1) and rule (2) have different syntactic structures (the left two trees of Figure 1). Thus rule (1) can be used for translating noun …

Content truncated to 500 words