next up previous contents index
Next: Intelligibility Up: Evaluating MT Systems Previous: Some Central Issues

Evaluation of Engine Performance


Substantial long-term experience with particular MT systems in particular circumstances shows that productivity improvements and cost-savings actually achieved can be very variable. Not all companies can apply MT as successfully as the following:



Different organizations experience different results with MT. The above examples indicate that the kind of input text is one of the important factors for getting good results. A sound system evaluation is therefore one which is executed within the company itself. An MT vendor might provide you with translated material which shows what their system can do. There is, however, no guarantee that the system will do the same in a different company setting, with different texts. Only a company specific evaluation will provide the client with the feedback she ultimately wants. Information provided by the MT vendor can be useful though, e.g. if system specifications indicate what sort of text type  it can or cannot handle or what sort of language constructions are problematic for their system.

In evaluating MT systems one should also take into account the fact that system performance will normally improve considerably during the first few months after its installation, as the system is tuned to the source materials, as discussed in Chapter gif. It follows that performance on an initial trial with a sample of the sort of material to be translated can only be broadly indicative of the translation quality  that might ultimately be achieved after several months or years of work.

Something similar holds for those stages of the translation process which involve the translator, like dictionary updating and post-editing  of the output. Times needed for these tasks will reduce as translators gain experience.

So how do we evaluate a system? Early evaluation studies were mainly concerned with the quality  of MT. Of course, assessing translation quality is not just a problem for MT: it is a practical problem that human translators face, and one which translation theorists have puzzled over. For human translators, the problem is that there are typically many possible translations, some of them faithful to the original in some respects (e.g. literal meaning), while others try to preserve other properties (e.g. style , or emotional impact).gif  

In MT, the traditional transformer  architecture introduces additional difficulties, since its output sentences often display structures and grammar that are unknown to the target language. It is the translator's task to find out what the correct equivalent is for the input sentence and its ill-formed translation. And, in turn, the evaluator's task is to find out how difficult the translator's task is.

In the rest of this chapter we will describe the most common evaluation methods that have been used to date and discuss their advantages and disadvantages.

next up previous contents index
Next: Intelligibility Up: Evaluating MT Systems Previous: Some Central Issues

Arnold D J
Thu Dec 21 10:52:49 GMT 1995