Indirect-HMM-based Hypothesis Alignment for Combining Outputs from Machine Translation Systems

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 98–107, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Indirect-HMM-based Hypothesis Alignment for Combining Outputs from Machine Translation Systems Xiaodong He†, Mei Yang‡ *, Jianfeng Gao†, Patrick Nguyen†, and Robert Moore† †Microsoft Research ‡Dept. of Electrical Engineering One Microsoft Way University of Washington Redmond, WA 98052 USA Seattle, WA 98195, USA {xiaohe,jfgao, panguyen, bobmoore}@microsoft.com yangmei@u.washington.edu Abstract This paper presents a new hypothesis alignment method for combining outputs of multiple machine translation (MT) systems. An indirect hidden Markov model (IHMM) is proposed to address the synonym matching and word ordering issues in hypothesis alignment. Unlike traditional HMMs whose parameters are trained via maximum likelihood estimation (MLE), the parameters of the IHMM are estimated indirectly from a variety of sources including word semantic similarity, word surface similarity, and a distance-based distortion penalty. The IHMM-based method significantly outperforms the state-of-the-art TER-based alignment model in our experiments on NIST benchmark datasets. Our combined SMT system using the proposed method achieved the best Chinese-to-English translation result in the constrained training track of the 2008 NIST Open MT Evaluation. 1 Introduction* System combination has been applied successfully to various machine translation tasks. Recently, confusion-network-based system combination algorithms have been developed to combine outputs of multiple machine translation (MT) systems to form a consensus output (Bangalore, et al. 2001, Matusov et al., 2006, Rosti et al., 2007, Sim et al., 2007). A confusion network comprises a sequence of sets of alternative words, possibly including null’s, with associated scores. The consensus output is then derived by selecting one word from each set of alternatives, to produce the sequence with the best overall score, which could be assigned in various ways such as by voting, by * Mei Yang performed this work when she was an intern with Microsoft Research. using posterior probability estimates, or by using a combination of these measures and other features. Constructing a confusion network requires choosing one of the hypotheses as the backbone (also called “skeleton” in the literature), and other hypotheses are aligned to it at the word level. High quality hypothesis alignment is crucial to the performance of the resulting system combination. However, there are two challenging issues that make MT hypothesis alignment difficult. First, different hypotheses may use different synonymous words to express the same meaning, and these synonyms need to be aligned to each other. Second, correct translations may have different word orderings in different hypotheses and these words need to be properly reordered in hypothesis alignment. In this paper, we propose an indirect hidden Markov model (IHMM) for MT hypothesis alignment. The HMM provides a way to model both synonym matching and word ordering. Unlike traditional HMMs whose parameters are trained via maximum likelihood estimation (MLE), the parameters of the IHMM are estimated indirectly from a variety of sources including word semantic similarity, word surface similarity, and a distance- based distortion penalty, without using large amount of training data. Our combined SMT system using the proposed …

Content truncated to 500 words