next up previous
Next: The Alignment Protocol Up: The BAF: A Corpus Previous: Background

Subsections

Guidelines and Definitions

The first step in the process of building a corpus of hand-aligned bitext is to clarify what we understand by the term alignment . Essentially, this entails describing the objects that the alignment connects, and defining how the alignment connects them. Based on the answers to these questions, a set of guidelines can then be devised, which the human aligners will be instructed to follow when producing the alignments.

What is an Alignment?

A bitext alignment describes the relations that exists between a text and its translation. These relations can be viewed at various levels of granularity: between text divisions, paragraphs, sentences, propositions, words, even characters. While it would certainly have been interesting to produce finer-grain alignments, it was decided that the BAF would record correspondences at the level of sentences. This decision was based on a number of factors.

First, sentence-level alignments have so far proved very useful in a number of applications, which could be characterized as high recall, low precision applications, i.e. applications where it is more important to have all the answers to a specific question than to have only the ``good'' ones.

One example of such an application is bilingual lexicography. When a lexicographer is examining a bilingual concordance, with a view to mapping out the various meanings or contexts of use of a particular term or expression, he seeks exhaustivity. In other words, he is willing to tolerate a relatively high number of irrelevant or redundant examples (``noise''), in order to make sure that he doesn't overlook anything (``silence'').

Automatic or machine-assisted translation verification is another such application. A system that does translation verification will look for specific translation errors, such as omissions on the part of the translator, the use of faux-amis (false cognates), inconsistent use of terminology, etc. If translation verification is anything like spelling or grammar checking, we can expect users to be ready to tolerate a fair amount of noise, just to make sure they don't miss out on glaring errors.

A final example is the automatic acquisition of information about translation, as was proposed in [Brown et al., 1993] as part of a project to build a machine translation system entirely based on statistical knowledge. While such ambitious projects now seem to have been abandoned, the statistical models at the heart of these projects are still around, for example in less ambitious interactive MT projects [Foster et al., 1997] and text alignment systems [Simard and Plamondon, 1996]. Such statistical models need to be ``trained'' with large quantities of bitext. Intuitively, the ideal training material for this task would be bitext aligned at the level of words. Yet, because these models picture the translation process in an extremely simplified manner, reliable statistical estimates can nevertheless be obtained from much less precise data, such as pairs of sentences.

This explains why a lot of the research effort in this domain has so far focussed on sentence-level alignments. Of course, this is not to say that reference alignments at a finer level would not be a useful thing, in the contrary. Besides, a word-level alignment could be made to incorporate the sentence-level alignment as a by-product.

Unfortunately, producing such a thing as a word-level alignment turns out to be a much more difficult problem: while there is often a one-to-one correspondence between the sentences of a text and its translation, matters get a lot more complicated when we get down to the level of ``words''. The main reason is that, at this level, syntactic and stylistic constraints in the target language affect the content and structure of the translated text at least as much as does the source text. As a result, in order to accurately describe the complex relations that exist between the words of a text and its translation, we will likely need a fairly elaborate alignment scheme. Finally, it is clear that producing hand-made word-alignments for more than a few sentences is going to be a very costly proposition.

For all these reasons, we decided that it would be more appropriate initially to concentrate on sentence-level alignments. Furthermore, we decided to restrict ourselves to ``non-crossing'' alignments:

We refer to such alignments as ``non-crossing'' because of the impossibility to explicitly account for inversions, i.e. situations where the order of sentences is not the same in the two texts. This type of alignment nonetheless covers the vast majority of situations encountered in real-life texts. Furthermore, this is the type of output that is actually produced by most existing sentence alignment programs.

What is a Sentence?

If we are going to align sentences, then obviously we must clarify what we understand by sentence : while most people have strong intuitions about what is a sentence and what is not, there is no universal definition of that notion. Before we set out on devising one, however, it should be noted that because the BAF is primarily intended to be used as a testbed for alignment methods, neither the exact definition, nor the actual segmentation of the text that results are crucially important: if the evaluation process focuses on alignment, the tested methods should all work on the same prior segmentation of the text. It is unlikely that a particular segmentation will favor one alignment method over another.

Therefore, our main concern in this regard was to come up with some guidelines for segmentation that would be both practical for the aligners and useful for the end-users of the corpus. We started out with something relatively straightforward, which we then expanded as needed. Essentially, these were the guiding principles:

What is a translation?

Finally, we needed to provide the aligners with is some criteria for determining what constitutes a translation. In general, we found it satisfactory to say that segments of text A and B were translations of one another if they conveyed the same ``ideas'' or ``concepts'', at least to an acceptable point. The main practical problems we had to solve revolved around situations where the translation deviated from its usual ``linear'' progression.

First, there were the cases of omissions and insertions, i.e. situations where some segment in one text does not appear to have a corresponding counterpart in the other text. In these cases, we allowed for the existence of ``empty'' segments in the alignment. This way, a sentence that does not have an equivalent in the other text can be aligned with an empty segment.

There were some situations where we chose to ignore an omission (or insertion), for the benefit of recording a larger correspondence. This would happen, for example, if a single sentence A in one language was translated as two sentences A1' and A2', between which a third, untranslated sentence B' was interpolated. In this case, we would simply align A with the sequence A1'B'A2', regardless of the fact that B' has no equivalent in A.

Then, there was the case of inversions. This happens when the order of the sentences is not the same in the source and translated texts. As mentioned earlier, our definition of alignment makes it impossible to explicitly account for inversions. Two different strategies were adopted, depending on the nature of the inversion.

For simple inversions, we opted for a strategy of ``under-segmentation'': when a pair of contiguous sentences AB appeared as B'A' in the other text, we chose not to segment the texts after sentences A and B', but rather to keep A and B together within the same segment, and then do the same for B' and A'.

For more complex inversions, we usually chose to treat the inverted segments as omissions. For example, given some sequence of sentences A1 A2 A3 ... An translated as A2' A3' ... An' A1', we would consider A1 and A1' to be ``omitted'' segments (align them with empty segments), and then align A2 with A2', A3 with A3', etc. Although this was clearly not the correct way of aligning the texts, it was felt that in the end, such an alignment would be more ``useful''.


next up previous
Next: The Alignment Protocol Up: The BAF: A Corpus Previous: Background
Michel Simard
4/24/1998