next up previous
Next: Guidelines and Definitions Up: The BAF: A Corpus Previous: Introduction

Background

For some years now, the international scientific community has shown interest in automated techniques that reproduce a multilingual speaker's ability to align a text with its translation, i.e. to identify the correspondences that exist between the segments of the two texts. Members of our laboratory, the RALI, have been actively involved in this area since 1991. Our interest in this question stems from the conviction that accurate alignment methods are the required basis for a whole set of computer tools for human translators [Isabelle et al., 1993]. The simplest example of such a tool is probably the TransSearch bilingual concordancing system [Simard et al., 1993], which allows a user to query a large archive of existing translations, in order to find ready-made solutions to specific translation problems. Such a tool has proved extremely useful not only for translators, but also for bilingual lexicographers [Langlois, 1996] and terminologists [Dagan and Church, 1994]. More sophisticated applications based on alignment technology have also been the object of recent work, such as the automatic building of bilingual lexical resources [Melamed, 1996,Klavans and Tzoukermann, 1995], the automatic verification of translations [Macklovitch, 1996,Macklovitch, 1995], the automatic dictation of translations [Brousseau et al., 1995] and even interactive machine translation [Foster et al., 1997].

Enthusiasm for this relatively new field of work was sparked early on by the apparent demonstration that very simple techniques could yield almost perfect results. For instance, to produce sentence alignments, Brown et al. [Brown et al., 1991] and Gale and Church [Gale and Church, 1991] both proposed methods that completely ignored the lexical content of the texts, and relied almost entirely on the intuition that short sentences tend to translate into short sentences, while longer sentences tend to translate into longer ones. With simple programs in which this observation was encoded into a statistical model, both teams were able to achieve accuracy levels exceeding 98%.

However, it quickly became apparent that this kind of performance could not be obtained with just any type of text, and that in general, the level of success that can be expected from automatic text alignment programs is highly dependent on the specific pair of texts under consideration. The truth is that, while text alignment is mostly an easy problem, especially when considered at the sentence level, there are situations where even humans have a hard time making the right decisions. In fact, the argument could be made that, ultimately, text alignment is no easier than the more general problem of natural language understanding.

Substantial work remains to be done, therefore, if the alignment technology is to achieve its full potential. Given the number of groups working on this problem, there is a pressing need for tools and resources that make it possible to evaluate and compare the performance of the various methods proposed. One of the things that is required is a common testbed, in the form of reference alignment corpora. This is precisely what the BAF is meant to be.


next up previous
Next: Guidelines and Definitions Up: The BAF: A Corpus Previous: Introduction
Michel Simard
4/24/1998