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

The Alignment Protocol

The definition of alignment given above suggested a very straightforward way of producing alignments by hand: read both texts in parallel, and segment them as you go along, in such a way that:

1.
segment boundaries always coincide with sentence boundaries;
2.
the nth segment in one text and the nth segment in the other are translations of one another;
3.
segments are always as small as possible.

Of course, given the relative vagueness of the definitions of sentence and translation given above, it was clear that in many situations, arbitrary decisions would have to be made. Our human ``aligners''[*]were instructed to be as consistent as possible, and when in doubt, to try to do the most ``useful'' thing. But even then, because of the repetitive nature of the task, errors had to be expected.

For these reasons, it was decided that all the texts would be aligned twice, each time by a different aligner. The resulting alignments would then be compared, so as to detect any discrepancies between the two. The aligners were then asked to conciliate these differences together. Because all of the BAF corpus was aligned by the same two aligners, this way of proceeding not only minimized the number of errors, it also ensured that both aligners had the same understanding of the guidelines.

In this regard, it is interesting to note that the vast majority of disagreements between the two aligners revolved around questions of sentence segmentation rather than questions of translational equivalence. Considering that, as discussed earlier, the actual segmentation on which the alignment is based is not crucially important, this suggests that it would probably have been a good idea to first have the texts segmented by a single person, and then have the aligners produce the alignments based on that segmentation.


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