Muriel Vasconcellos
You need to keep several points in mind when deciding if MT is right for you. First, you must determine if you have an application for which MT is appropriate. It's important to pick your application and then decide on your system rather than vice versa.
Costs soar when the input isn't in machine-readable form, and an OCR (optical character recognition) device, while helpful, isn't a panacea. If your documents aren't in electronic form, you may want to think twice about using MT. In addition, there should be a large volume of material (e.g., 100,000 words per month) to be translated, with the expectation that more will be coming from the same source. In the beginning, there should be only one domain (i.e., subject matter); you can branch out later as you become more familiar with all the ins and outs.
Your decision to use MT will hinge, in large part, on the format, volume, and linguistic characteristics of the source-language text. The text should contain no ambiguities.
Hardware and Human Factors
You need a hardware platform that an MT system will run on. In a multiuser setting, you must be sure that there is good word processing support and that all the users have strong word processing skills. Multitasking workstations designed for translators are helpful.
Be sure to recruit people who have a positive attitude about using MT. This is especially important during the first few months while you are getting your system up and running. This stage involves customizing the dictionaries and gaining proficiency in postediting.
Your choice of a system will depend on the characteristics of your application, so it's important to identify criteria that are specific to your needs. And don't be tempted to buy software just because it's inexpensive. As with a house pet, the price you pay up front is a drop in the bucket compared to the cost of the care and feeding for the rest of your mutual lives. For example, a less expensive system might cost more in terms of support personnel and customer support (some MT companies charge you for it). Also, if your time frame for translations is tight and if your budget for human intervention is limited, it's crucial that you test the system's performance on randomly selected texts.
Despite decades of scientific study, the evaluation of translations is an uncertain exercise. The definition of an error will vary, depending on the purpose of the translation and the values of the end users. Errors in raw output are important mainly to the analyst, who knows the inner workings of MT systems and can classify the error types according to their causes. Such an analysis can tell something about the system's potential and the effort that will be required to fix and to maintain it. You should make sure that you compare outputs from different systems produced under the same conditions.
The value of a system depends on its potential to grow and to improve its performance, as well as how easy it is to use and to maintain. It's important to know the language combinations that have been developed for the system, the size of the dictionaries or knowledge bases, the ease with which you can add to the dictionaries, and the possibilities of extending the system to include the domain that you are interested in.
Muriel Vasconcellos is president of the Association for Machine Translation in the Americas and is a Washington, D.C.-based consultant in translation and machine translation. You can reach her on BIX c/o "editors" or on CompuServe at 71024,123.
Per comentaris i observacions sobre el servidor, poseu-vos en contacte amb l'Administrador WWW