Better Alignments = Better Translations?

Proceedings of ACL-08: HLT, pages 986–993, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Better Alignments = Better Translations? Kuzman Ganchev Computer & Information Science University of Pennsylvania kuzman@cis.upenn.edu Jo˜ao V. Grac¸a L2F INESC-ID Lisboa, Portugal javg@l2f.inesc-id.pt Ben Taskar Computer & Information Science University of Pennsylvania taskar@cis.upenn.edu Abstract Automatic word alignment is a key step in training statistical machine translation sys- tems. Despite much recent work on word alignment methods, alignment accuracy in- creases often produce little or no improve- ments in machine translation quality. In this work we analyze a recently proposed agreement-constrained EM algorithm for un- supervised alignment models. We attempt to tease apart the effects that this simple but ef- fective modification has on alignment preci- sion and recall trade-offs, and how rare and common words are affected across several lan- guage pairs. We propose and extensively eval- uate a simple method for using alignment models to produce alignments better-suited for phrase-based MT systems, and show sig- nificant gains (as measured by BLEU score) in end-to-end translation systems for six lan- guages pairs used in recent MT competitions. 1 Introduction The typical pipeline for a machine translation (MT) system starts with a parallel sentence-aligned cor- pus and proceeds to align the words in every sen- tence pair. The word alignment problem has re- ceived much recent attention, but improvements in standard measures of word alignment performance often do not result in better translations. Fraser and Marcu (2007) note that none of the tens of papers published over the last five years has shown that significant decreases in alignment error rate (AER) result in significant increases in translation perfor- mance. In this work, we show that by changing the way the word alignment models are trained and used, we can get not only improvements in align- ment performance, but also in the performance of the MT system that uses those alignments. We present extensive experimental results evalu- ating a new training scheme for unsupervised word alignment models: an extension of the Expecta- tion Maximization algorithm that allows effective injection of additional information about the desired alignments into the unsupervised training process. Examples of such information include “one word should not translate to many words” or that direc- tional translation models should agree. The gen- eral framework for the extended EM algorithm with posterior constraints of this type was proposed by (Grac¸a et al., 2008). Our contribution is a large scale evaluation of this methodology for word alignments, an investigation of how the produced alignments dif- fer and how they can be used to consistently improve machine translation performance (as measured by BLEU score) across many languages on training cor- pora with up to hundred thousand sentences. In 10 out of 12 cases we improve BLEU score by at least 1 4 point and by more than 1 point in 4 out of 12 cases. After presenting the models and the algorithm in Sections 2 and 3, in Section 4 we examine how the new alignments differ …

Content truncated to 500 words