Machine Translation Services: the use of computer software to translate text
Machine Translation differs from translation memory in that it is a completely automated process.
It turns out that getting machines to produce high-quality translations is very difficult. Traditionally machine translation and post-editing has produced a poor quality of output. In general therefore, most translation agencies, including TranslateMedia, would not work with translators who rely on machine translation. Recently, however, a number of highly regarded translators have gone on record to say that they find machine translation can be useful to stimulate ideas.
Customised Machine Translation
Customised MT engines can give better results than general machine translation. This involves training a specific engine to handle your work, using human translations. These are some of the criteria your work needs to meet to make this a realistic option:
We have incorporated tools that allow you to train your own machine translation engine into our STREAM Translation System. Doing this for a specific work stream within your organisation may well offer cost benefits over time, and by dictating a smaller domain for the machine translation to deal with results can be improved.
Rule-based & Statistical MT
There are two main approaches to machine translation: rule-based, where the computer program attempts to model the rules of language; and statistical, where the computer program attempts to learn from large amounts of previously translated text.
Statistical translation has improved substantially in recent years, and for some language combinations and types of text the quality is now reasonable. On very large projects, where it would not be economically feasible to translate everything by hand, machine translation can be useful. Post-editing by a professional human translator is normally done to improve the quality.
We test Google’s machine translation on a regular basis to see how the quality compares. The hypothesis is that one day Google’s quality will be sufficient that it can replace the translator in the translation process, and then be edited by a professional human translator to get it up to the required quality.
All of our testing so far has shown that it is more time-consuming correcting Google’s translations, but there are deviations from language pair to language pair, and according to the type of text.
In our tests Google Translate has consistently performed better when translating English into Spanish and Italian than it has translating into French and German. It also translates general texts – simple communications and simple grammatical phrases – far better than more complex specialised texts. However, even in the best cases we have so far found that the time required to correct the machine translated text means that it is more cost-effective to translate from scratch.