This article will not make jokes about funny translations you can obtain from free internet translation engines, such as Google Translate or Bing Translator. Just one small remark though: did you also notice that it is getting much harder these days to produce those funny junk translations?
Slowly but surely, machine translation is shedding off its image of eternal promise. It is gaining credibility, not only with the average internet user but also with an increasing number of translation companies and translation buyers. Whereas early adopters used to be very discrete about using machine translation, they now openly advertize it.
Technical integration of MT has become commonplace. Any modern CAT tool offers some way of adding MT to the translation process, but more often than not the choice is limited to the engines that are freely available on the internet.
One can wonder why anyone would take the risk of using those engines for professional translation, considering all confidentiality and data protection warnings. Reality is, quite a few people do use them, at least judging by research carried out by Common Sense Advisory and others. Including – against all expectations - professional translators.
A few years ago, Microsoft narrowed the gap between the free online machine translation and professional translation by introducing its Translation Hub. The Hub allowed users to enhance the engine using their own content and glossaries, which addressed one of the main criticisms of the professional translation community.
The same approach is now followed by a number of relative newcomers who took advantage of the development of statistical MT engines and in particular of an EU-funded project called Moses, to bring user-trained machine translation within reach of any translation company or translation department. Providers such as KantanMT and Tilde (LetsMT) are offering cloud-based MT services specifically targeted at translation providers. Gone are the days that you had to have your own team of MT gurus and a fat wallet of investment money to make the technology work for you. Training an MT engine has become as easy as uploading a bilingual corpus and a terminology list, or so it seems.
So, if it is available to us all, why aren't we all using it? Well, let's say that the translation industry itself is not very adventurous, to put it mildly. When CAT tools made their entry in the '80s, they were not welcomed by translators or translation companies, but translation buyers, who quickly saw the savings and productivity potential of the technology, dragged us kicking and screaming into that brave new world.
MT is the next step in that process and we are witnessing the same reluctance, but things are changing already. The latest EUATC survey on expectations and concerns of the European translation companies showed that an increasing number of them – and this time not just the large corporations but also small outfits - are implementing or planning to implement machine translation technology.
None of us really expects MT to replace human translation any day soon, but we do realize that it can be an interesting new tool in our toolbox. One that we will have to learn to handle properly. Because even though statistical MT has made it possible to improve the engine without prohibitive linguistic efforts, the returns are not all that obvious – at least compared to translation memories.
For one, the return on investment of translation memories increases over time, in particular in business use where text content is often – at least partially – repetitive in nature. Machine translation, on the other hand, can only deliver returns on new text, which gradually decreases since the growing repetitive part of the content is retrieved from the translation memories it is integrated with.
There is also the question of content suitability. Whereas CAT tools can be easily used on any content, except maybe on the most creative types, machine translation is an entirely different matter. As soon as a decent level of editorial effort is required, time savings compared to standard human translation quickly drop to near zero level. After all, as Arle Lommel said: "Machine Translation will only replace those humans that translate like machines".
And finally there is the fact that, no matter how well you train the engine, you will never reach the reliability level that you are getting from you CAT system. Granted, trained statistical MT may produce target sentences that read wel, but you should never lower your guard! The translation can hit you with a negation that was not there in the original, or with an unexpected change of content.
Potential users, be it individual translators, translation companies or translation buyers, therefore need to weigh the potential returns against the required investments – licence or subscription fees, development of appropriate terminology and bilingual corpora, the inevitable editing stage, as well as the learning curve and initial resistance to change of their own staff and postediting resources. The lack of speed of MT adoption may therefore not be a bad thing. It allows us to learn from the earlier adopters to avoid the pitfalls and to adopt best practices.
There is no doubt in our mind that MT will ultimately follow the same path as CAT tools. In a decade from now, translation companies not using MT at all will be a minority. But we do expect that the technology will be used in a much more selective and reasoned way than translation memories.
You may have noticed that this article is not speaking about using MT for so-called information only or lower quality translation. That type of MT usage will be discussed in an article focusing on differentiation in translation quality, which you will also find in the Hot Topics area.
In the meantime, we invite you to participate in the discussion forum on machine translation, here on the EUATC website.