next up previous contents index
Next: Evaluation of Engine Up: Evaluating MT Systems Previous: Introduction

Some Central Issues

The evaluation of MT systems is a complex task. This is not only because many different factors are involved, but because measuring translation performance is itself difficult. The first important step for a potential buyer is to determine the translational needs of her organization. Therefore she needs to draw up a complete overview of the translational process, in all its different aspects. This involves establishing the size of the translation task, the text type  of the material and its form (is it machine readable and if so, according to which standards). It also involves considering organizational issues, e.g. the tasks of each member of staff concerned in some way with translation. With that information at hand she can start to investigate what the consequences of the purchase of an MT system would be. These are some of the factors to keep in mind:

Organizational Changes
Incorporating an MT system into the translation process will impact upon both the process and the personnel involved. There will be consequences for system administrators and support staff, but above all for the translators themselves, whose tasks will change significantly. Whereas before they will probably have spent the major part of their time actually translating or editing human translations, they will now find themselves spending a lot of time updating the system's dictionaries and post-editing  the results of machine translation. There may also be a need to build automatic termbanks . Translators will need to receive training in order to perform these new tasks adequately.

It is important that the personnel support the changeover to MT. They may not always be aware of the fact that MT can lead to more job satisfaction among translators since MT systems are particularly efficient at tedious, repetitive tasks whereas more challenging translation work often still needs to be done by the human translators. If translators in an organization have decided for some reason or other that they do not want to work with MT, imposing it on them is guaranteed to produce poor results.

Technical environment
We have emphasised right from the start that success depends in part on MT being effectively incorporated as part of a wider document preparation process inside an organization. Smooth handling of text throughout the whole process will prevent unnecessary delays. The MT engine and the document system may well come from different suppliers but they must adhere to the same standards and formats for textual material.

Bear in mind that good document preparation facilities in themselves can improve translator productivity. A decade or so ago much of the productivity increase claimed by some vendors of smaller MT systems could be attributed to their providing rather good multi-lingual word processing facilities, at a time when many translators used only an electric typewriter. Some MT vendors still supply a whole MT system package where the engine is inextricably wrapped up with some specialised word processing and text-handling tool unique to that particular system. This is undesirable on two counts: first, if you are already familiar with a good multi-lingual word processor, little is gained by having to learn another which does much the same things; second, it is likely that an MT vendor's home-grown text-processing facilities will be inferior to the best independent products, because most of the effort will have gone into developing the translation engine.

Status of Vendor
Buying an MT system is a considerable investment, and the stability and future solvency of the vendor is an important consideration. After all, contact with the vendor is ideally not just limited to the initial purchase of the system. A solvent vendor can provide installation support and training in the early stages, and general support and updates later, which may improve performance considerably (e.g. specialized dictionaries, or new language pairs which can be integrated into the existing MT set-up).

Engine Performance: Speed
In some circumstances, the speed at which the
engine churns out raw translated text won't actually be crucial. If the system requires interaction with the translator whilst it is translating, then of course it should not amble along so slowly as to to keep the translator waiting all the time. But if it is functioning without direct interaction, it can proceed at its own pace in the background whilst the translator gets on with other jobs such as post-editing  or hand translation of difficult material. This aspect also depends on the user's translational needs: if the user's material requires 15 hours daily on a fast MT system and 20 on a slower one, no one will notice the difference if the system is running overnight. Of course, there are situations where the quick delivery of translation output is essential. (The agronomist in Chapter gif, who wants to process very large quantities of material to a low level may be an example.) But in general, slow speed is the one component of MT performance of which upgrading is relatively easy: by buying some faster hardware for it to run on.
Engine Performance: Quality
  This is a major determinant of success. Current general purpose commercial MT systems cannot translate all texts reliably. Output can sometimes be of very poor quality indeed. We have already mentioned that the post-editing  task (and with it the cost) increases as translation quality gets poorer. In the worst case, using MT could actually increase translation costs by tying up translators in editing and maintenance tasks, ultimately taking up more time than would have been required to produce translations entirely by hand. Because of its enormous influence on the overall translation cost, translation quality  is a major aspect in MT evaluation.

next up previous contents index
Next: Evaluation of Engine Up: Evaluating MT Systems Previous: Introduction

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