Right from the start, the BAF was intended to be used as an evaluation resource in the development of general-purpose alignment methods. For this reason, we were interested in ``plain-text'' documents, i.e. text files that were not tied to a specific word-processing program, and that contained no formatting or structural mark-up. In many cases, the documents were explicitly converted from proprietary formats (WordPerfect, FrameMaker, etc.) to ``plain-text'' format. In other cases, mark-up (SGML and others) was eliminated.
In its current version, the BAF corpus takes the form of a collection of computer files. Because the corpus is available in three different formats, several files correspond to each pair of documents.
One important difference between the COAL and CES formats is that the CES assumes a complete segmentation of the texts into sentences, which is made explicit by the mark-up. The COAL format does not make that segmentation explicit, and nothing guarantees that the segmentation that is implicit in the alignment is complete. For example, if one sentence in the English text is translated as two sentences in the French text, the boundary between the first and second French sentences will not appear in the COAL alignment.
The CES version of the BAF was produced automatically from the COAL version. While performing the conversion, we ``completed'' the implicit segmentation, by using a number of sentence-boundary detection heuristics. A random sampling of the resulting sentences reveals that the segmentation is about 97.5% correct: less than 2.5% of the sentence boundaries in the corpus are incorrect, and less than 2.5% of the real boundaries are missing.