Improving Alignments for Better Confusion Networks for Combining Machine Translation Systems

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 33–40 Manchester, August 2008 Improving Alignments for Better Confusion Networks for Combining Machine Translation Systems Necip Fazil Ayan and Jing Zheng and Wen Wang SRI International Speech Technology and Research Laboratory (STAR) 333 Ravenswood Avenue Menlo Park, CA 94025 {nfa,zj,wwang}@speech.sri.com Abstract The state-of-the-art system combination method for machine translation (MT) is the word-based combination using confu- sion networks. One of the crucial steps in confusion network decoding is the align- ment of different hypotheses to each other when building a network. In this paper, we present new methods to improve alignment of hypotheses using word synonyms and a two-pass alignment strategy. We demon- strate that combination with the new align- ment technique yields up to 2.9 BLEU point improvement over the best input sys- tem and up to 1.3 BLEU point improve- ment over a state-of-the-art combination method on two different language pairs. 1 Introduction Combining outputs of multiple systems perform- ing the same task has been widely explored in various fields such as speech recognition, word sense disambiguation, and word alignments, and it had been shown that the combination approaches yielded significantly better outputs than the in- dividual systems. System combination has also been explored in the MT field, especially with the emergence of various structurally different MT systems. Various techniques include hypothesis selection from different systems using sentence- level scores, re-decoding source sentences using phrases that are used by individual systems (Rosti et al., 2007a; Huang and Papineni, 2007) and 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. word-based combination techniques using confu- sion networks (Matusov et al., 2006; Sim et al., 2007; Rosti et al., 2007b). Among these, confu- sion network decoding of the system outputs has been shown to be more effective than the others in terms of the overall translation quality. One of the crucial steps in confusion network decoding is the alignment of hypotheses to each other because the same meaning can be expressed with synonymous words and/or with a different word ordering in different hypotheses. Unfortu- nately, all the alignment algorithms used in confu- sion network decoding are insensitive to synonyms of words when aligning two hypotheses to each other. This paper extends the previous alignment approaches to handle word synonyms more effec- tively to improve alignment of different hypothe- ses. We also present a two-pass alignment strategy for a better alignment of hypotheses with similar words but with a different word ordering. We evaluate our system combination approach using variants of an in-house hierarchical MT sys- tem as input systems on two different language pairs: Arabic-English and Chinese-English. Even with very similar MT systems as inputs, we show that the improved alignments yield up to an abso- lute 2.9 BLEU point improvement over the best input system and up to an absolute 1.3 BLEU point improvement over the old alignments in a confusion-network-based combination. The rest of this paper is organized …

Content truncated to 500 words