next up previous
Next: References Up: The BAF: A Corpus Previous: Programs

Conclusion and Future Work

We have described the BAF, a corpus of English and French translations, hand-aligned at the sentence level. The corpus, which totals approximately $800\ 000$ words, is primarily intended as an evaluation tool in the development of automatic bilingual text alignment method. It is currently available from the RALI's Web site, at URL http://www-rali.iro.umontreal.ca. A number of computer tools for manipulating bitext alignments were also developed in the course of the project, the most important of which is the Manual program, which allows one to visualize and edit an alignment within the Emacs editor. All of these programs are also available on the RALI's Web site.

The BAF corpus has already been put to use: in the context of the AUPELF-UREF's ARC-A2, a friendly competition between alignment programs was organized. The BAF was used, along with a similar corpus, as a common testbed: the corpus files were submitted to each of the alignment programs, and the output of each was compared to the hand-made alignments. Performances were then measured, using various evaluation metrics of the kind proposed in [Isabelle and Simard, 1996].

Yet, while the corpus is already usable, there is still much room left for improvement. For instance, we did not initially believe it necessary to first produce a complete segmentation of the texts, and then describe correspondences with regard to this segmentation: in practice, both segmentation and alignment were performed in parallel. Although this way of proceeding was probably more economical, it had one major drawback: when two or more sentences translated to a single one, the internal segmentation was not recorded. As a result, the segmentation of the corpus in the BAF was partial. In version 1.1 of the corpus, the segmentation was completed by means of automatic segmentation methods. This automatic segmentation contains errors, and should be manually verified, something that we plan to do in the short term.

Also, the BAF does not record crossing alignments. While the vast majority of alignments are non-crossing, it is currently impossible to use the corpus to evaluate alignment methods that do crossing alignments. Correcting this situation would probably not be too costly, because once the segmentation of the text is verified, crossing alignments can be easily detected by examining exclusively alignment patterns that deviate from the usual ``1 to 1''.

Finally, we have recently begun experimenting with word-level alignments. At this level, the questions of ``what to align?'' and ``how to align them?'' become much more complex. In spite of these difficulties, we hope to have at least a small part of the BAF aligned at this level before the end of the ARC-A2 collaboration.


next up previous
Next: References Up: The BAF: A Corpus Previous: Programs
Michel Simard
4/24/1998