Maximum Entropy Based Lexical Reordering Model for Hierarchical Phrase-based Machine Translation

Maximum Entropy Based Lexical Reordering Model for Hierarchical Phrase-based Machine Translation Zhongguang Zheng, Yao Meng, Hao Yu Fujitsu R&D Center CO., LTD. 15/F, Tower A, Ocean International Center, No.56 Dongsihuan Zhong Rd. Chaoyang District, Beijing, 100025, China {zhengzhg, mengyao, yu}@cn.fujitsu.com Abstract. The hierarchical phrase-based (HPB) model on the basis of a synchronous context-free grammar (SCFG) is prominent in solving global reorderings. However, the HPB model is inadequate to supervise the reordering process so that sometimes positions of dif- ferent lexicons are switched due to the incorrect SCFG rules. In this paper, we consider the order of two lexicons as a classification problem and propose a novel lexical reorder- ing model based on a maximum entropy classifier. Our model employs the word alignment and translation during the decoding process. Experimental results on the Chinese-to-English task showed that our method outperformed the baseline system in BLEU score significantly. Moreover, the translation results further proved the effectiveness of our approach. Keywords: hierarchical phrase-based model, reordering, maximum entropy model 1 Introduction Reordering is a big challenge for statistical machine translation (SMT). The hierarchical phrase- based (HPB) translation model (Chiang, 2005), which adopts a synchronous context-free grammar (SCFG), is considered to be prominent in capturing global reorderings. However, the HPB model is weak in controlling the reordering process. Arbitrary reorderings frequently come up during the decoding phase, worsening the translation quality, such as the example shown in Figure 1(a). The non-terminal “X2” is reordered with “X1” and “anshui rongye”, but the punctuation “�” indicates that the phrase should be translated monotonously. This kind of reordering error frequently occurs when target phrases contain similar lexicons, which include both terminals and non-terminals, with different permutations, e.g., the rules in Figure 1(b). We believe that there are mainly two reasons that the HPB model can not distinguish such ambiguous rules properly. • The extraction of SCFG rules is merely based on the word alignment information. Thus any form of rules could be obtained. As shown in Figure 1, the same Chinese part “X1 anshui rongye X2” corresponds to different target rules “X2 X1 ammonia solution” and “X1 am- monia solution X2”, which contain the same English words but differ in word order. The order of non-terminals should depend on their concrete translations and contexts. However, the decoder does not take into account those factors when reorderings happen. • There are not enough features to evaluate the correctness of word order for SCFG rules. Conventionally, the HPB model has 8 features (Chiang, 2005), including language model, constituent feature, word penalty, phrase penalty, bi-direction translation weights P(γ|α) and P(α|γ), bi-direction lexical weights Pw(γ|α) and Pw(α|γ), but none of them is responsible for the rationality of word order. Although language model evaluates the fluency of target Copyright 2011 by Zhongguang Zheng, Yao Meng, Hao Yu 25th Pacific Asia Conference on Language, Information and Computation, pages 216–225 216 jianjinshu , ru qingyanghuajia shuirongye 、 anshui rongye 、 tansuanqingna shuirongye Alkali metal, such as aqueous potassium hydroxide , aqueous sodium bicarbonate , ammonia …

Content truncated to 500 words