Line ’em up: advances in alignment technology and their impact on translation support tools

LINE 'EM UP: ADVANCES IN ALIGNMENT TECHNOLOGY AND THEIR IMPACT ON TRANSLATION SUPPORT TOOLS1 Elliott Macklovitch Marie-Louise Hannan macklovi@citi.doc.ca hannan@citi.doc.ca Centre for Information Technology Innovation (CITI) 1575 Chomedey Blvd., Laval (Quebec) CANADA H7V 2X2 Abstract We present a quantitative evaluation of one well-known word alignment algorithm, as well as an analysis of frequent errors in terms of this model's underlying assumptions. Despite error rates that range from 22% to 32%, we argue that this technology can be put to good use in certain automated aids for human translators. We support our contention by pointing to certain successful applications and outline ways in which text alignments below the sentence level would allow us to improve the performance of other translation support tools. 1. Introduction Considerable effort has been invested in the past few years in developing what has been called a "new generation" of translation support tools based on parallel text alignment (Brown et al. [2]; Gale and Church [8]; Isabelle et al. [10]). One of the questions we can ask at this stage, then, is: How well do these translation support tools (TST's) work? More specifically, can recent progress in word-level alignment make a significant contribution to their performance? In an attempt to address these questions, we designed a small-scale experiment to evaluate the performance of one well-known word alignment algorithm. The results of this evaluation, carried out on a bilingual corpus of French and English texts, point to certain weaknesses associated with the model, and our analysis suggests some explanations for these weaknesses. The investigation itself was motivated by a need for finer-grained text alignments than those at the sentence level, which would permit us to push the performance of existing translation support tools even further, and allow us to imagine novel ones. In the following sections, we will introduce and clarify some of the basic notions related to text alignment and translation analysis in order to situate our study within its larger theoretical context; then discuss progress in text alignment from the sentence to the word level; and, finally, present our experimental design and an analysis of results obtained. 1. We would like to thank George Foster, to whom we are indebted for his patient explanations of the statistical subtleties of alignment technology. Thanks also go to Pierre Plamondon, for providing us with output from TM- Align and input on its operation. Of course, neither is responsible for any of our errors of interpretation. 145 2. Translation Analysis 2.1 Definition of Terms and Concepts The term "translation analysis", introduced by Isabelle et al. [10], can be defined as a set of techniques for automatically reconstructing the correspondences between segments of a source text and the segments of its translation. Translational equivalence must be seen as compositional in some sense, with the global correspondences between source and translation being analyzable into sets of finer correspondences. The challenge for translation analysis is to discover the links between increasingly finer segments, motivated by the fact that existing translations contain a wealth …

Content truncated to 500 words