Lattice-based System Combination for Statistical Machine Translation

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1105–1113, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Lattice-based System Combination for Statistical Machine Translation Yang Feng, Yang Liu, Haitao Mi, Qun Liu, Yajuan L¨u Key Laboratory of Intelligent Information Processing Institute of Computing Technology Chinese Academy of Sciences P.O. Box 2704, Beijing 100190, China {fengyang, yliu, htmi, liuqun, lvyajuan}@ict.ac.cn Abstract Current system combination methods usu- ally use confusion networks to find consensus translations among different systems. Requir- ing one-to-one mappings between the words in candidate translations, confusion networks have difficulty in handling more general situa- tions in which several words are connected to another several words. Instead, we propose a lattice-based system combination model that allows for such phrase alignments and uses lattices to encode all candidate translations. Experiments show that our approach achieves significant improvements over the state-of- the-art baseline system on Chinese-to-English translation test sets. 1 Introduction System combination aims to find consensus transla- tions among different machine translation systems. It has been proven that such consensus translations are usually better than the output of individual sys- tems (Frederking and Nirenburg, 1994). In recent several years, the system combination methods based on confusion networks developed rapidly (Bangalore et al., 2001; Matusov et al., 2006; Sim et al., 2007; Rosti et al., 2007a; Rosti et al., 2007b; Rosti et al., 2008; He et al., 2008), which show state-of-the-art performance in benchmarks. A confusion network consists of a sequence of sets of candidate words. Each candidate word is associated with a score. The optimal consensus translation can be obtained by selecting one word from each set to maximizing the overall score. To construct a confusion network, one first need to choose one of the hypotheses (i.e., candidate translations) as the backbone (also called “skeleton” in the literature) and then decide the word align- ments of other hypotheses to the backbone. Hy- pothesis alignment plays a crucial role in confusion- network-based system combination because it has a direct effect on selecting consensus translations. However, a confusion network is restricted in such a way that only 1-to-1 mappings are allowed in hypothesis alignment. This is not the fact even for word alignments between the same languages. It is more common that several words are connected to another several words. For example, “be capa- ble of” and “be able to” have the same meaning. Although confusion-network-based approaches re- sort to inserting null words to alleviate this problem, they face the risk of producing degenerate transla- tions such as “be capable to” and “be able of”. In this paper, we propose a new system combina- tion method based on lattices. As a more general form of confusion network, a lattice is capable of describing arbitrary mappings in hypothesis align- ment. In a lattice, each edge is associated with a sequence of words rather than a single word. There- fore, we select phrases instead of words in each candidate set and minimize the chance to produce unexpected translations such as …

Content truncated to 500 words