ALPHA BLOG
How good is Machine Translation?
October 17, 2018
The development and optimisation of machine translation (MT) has made great strides over the past two decades and has become therefore an ever-relevant topic for our industry. This partly due in fact to accelerating globalisation and the need for quick (ad-hoc) translations.
How good is Machine Translation? To answer this question, it is important to first discuss what MT can currently do, what it cannot “yet” do, and what areas of application are being considered now and in the future.
In short, the latest MT methods work like this: as many already translated texts in a language combination as possible are stored as a language pair. New texts to be translated are then compared with these existing translations using complex algorithms in a program (“engine”). The new translation is generated from a combination of the existing translation out of the most similar sentence pairs.
If the MT “engines” are extensively fed and “trained” with many existing (parallel) texts for a very specific subject area, the results are sometimes surprisingly good—but sometimes still miss the mark entirely.
Machine Translation Limitations
Even if the results are quite appealing, however, freedom from errors cannot be guaranteed. Texts for which exactness and precision is essential (e.g., legal texts or instructions for medical equipment) will not be translated via MT in the future (i.e., without human verification). Even texts from the field of marketing, where it is particularly important to evoke certain emotions in the target language, will probably never achieve their desired effect with machine translation.
What MT can already do
MT can, among other things, translate web pages on-the-fly into the target language. This is especially useful for sites whose content changes constantly, such as forums or social networks. Mistakes or even incomprehensible content is forgiven here: the benefit is greater than the harm.
Another area of application is the “rough” or “informative” translation, a.k.a. so-called “gisting“. Here, the goal is to ascertain the core theme and the general meaning of the text. The linguistic correctness and individual details are not the main focus here.
Free tools from Google, DeepL or Microsoft are often used for the purpose of “gisting”, but caution is advised: not only because of the often-high number of errors, but also because of data security. Almost any text that is entered will be saved and reused by the providers, which most users are not aware of.
Another possible application is the specialised use of specially trained “engines”. As many specific texts as possible for a given subject area are stored so that they can serve as samples, with the success that similar texts can be translated automatically with a very effective result. This is already successfully practised by large companies (which have their own servers and often their own department for this purpose). These texts are used with or without post-editing (a human correction), depending on the field of application. For example, in the USA, Ford translated internal instructions for car mechanics from English into Spanish and Portuguese ten years ago. The fact that the employees had the instructions in their language at all was more important in that case than individual small errors—in context, they basically understood what had to be done.
In the case of post-editing, it must always be carefully weighed against to what extent MT is worthwhile: if the results are poor, sometimes human “re-translation” is easier, faster and more accurate.
In summary, MT is best suited for additional or optional translations of large amounts of text that, if MT did not exist, probably would not have been translated at all.
MT has long since left its niche existence and is becoming an significant and helpful tool for certain human translations in the coming years to make the translation processes faster and more efficient. Especially with the ever improving “Custom MT” and new “Machine Learning Systems” the accuracy of translation results can only increase in the future.
At Alpha Translation Service GmbH, we do not see MT as a competitor, rather as an opportunity to be able to manage the enormous and ever-increasing volume of translation demand in the future. Upon request and in cooperation with some of our customers, we are already successfully carrying out MT projects and post-editing.
If you have projects for which MT might be an option, please do not hesitate to contact us!
Kategorien: know-how, translation industry
Tags: AI, machine translation, MT
