Machine Translation System Combination with Flexible Word Ordering

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 56–60, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics Machine Translation System Combination with Flexible Word Ordering Kenneth Heafield, Greg Hanneman, Alon Lavie Language Technologies Institute, Carnegie Mellon University 5000 Forbes Avenue, Pittsburgh, PA 15213, USA {kheafiel,ghannema,alavie}@cs.cmu.edu Abstract We describe a synthetic method for com- bining machine translations produced by different systems given the same input. One-best outputs are explicitly aligned to remove duplicate words. Hypotheses follow system outputs in sentence order, switching between systems mid-sentence to produce a combined output. Experi- ments with the WMT 2009 tuning data showed improvement of 2 BLEU and 1 METEOR point over the best Hungarian- English system. Constrained to data pro- vided by the contest, our system was sub- mitted to the WMT 2009 shared system combination task. 1 Introduction Many systems for machine translation, with dif- ferent underlying approaches, are of competitive quality. Nonetheless these approaches and sys- tems have different strengths and weaknesses. By offsetting weaknesses with strengths of other sys- tems, combination can produce higher quality than does any component system. One approach to system combination uses con- fusion networks (Rosti et al., 2008; Karakos et al., 2008). In the most common form, a skele- ton sentence is chosen from among the one-best system outputs. This skeleton determines the or- dering of the final combined sentence. The re- maining outputs are aligned with the skeleton, pro- ducing a list of alternatives for each word in the skeleton, which comprises a confusion network. A decoder chooses from the original skeleton word and its alternatives to produce a final output sen- tence. While there are a number of variations on this theme, our approach differs fundamentally in that the effective skeleton changes on a per-phrase basis. Our system is an enhancement of our previous work (Jayaraman and Lavie, 2005). A hypothesis uses words from systems in order, switching be- tween systems at phrase boundaries. Alignments and a synchronization method merge meaning- equivalent output from different systems. Hy- potheses are scored based on system confidence, alignment support, and a language model. We contribute a few enhancements to this pro- cess. First, we introduce an alignment-sensitive method for synchronizing available hypothesis ex- tensions across systems. Second, we pack similar partial hypotheses, which allows greater diversity in our beam search while maintaining the accuracy of n-best output. Finally, we describe an improved model selection process that determined our sub- missions to the WMT 2009 shared system combi- nation task. The remainder of this paper is organized as fol- lows. Section 2 describes the system with empha- sis on our modifications. Tuning, our experimen- tal setup, and submitted systems are described in Section 3. Section 4 concludes. 2 System The system consists of alignment (Section 2.1) and phrase detection (Section 2.2) followed by de- coding. The decoder constructs hypothesis sen- tences one word at a time, starting from the left. A partially constructed hypothesis comprises: Word The most recently decoded …

Content truncated to 500 words