Improving machine translation
The quality of machine translation is certainly improving (see ‘Neural machine translation), but it is not perfect yet. It can be improved by post-editing and adaptive machine translation. Time (and money) can be saved with technical translations by human post-editing of an imperfect machine translation, and most CAT tools support post-editing of machine translations. This is usually faster than translating from scratch, but the savings are difficult to predict. One study in 2010 reported time savings of about 40%; a more realistic Swiss estimate suggests that it is probably closer to 15-40%, but this is not always cost-effective.
Adaptive machine translation
Two companies have recently introduced a refinement of machine translation and post-editing, which they refer to as adaptive machine translation. The concept is simple, and involves adapting the machine translation according to the results of post-editing. According to SDL, “This is potentially a paradigm shift for the translation industry, as for the first time users can update a machine translation engine as they go along… Ultimately, with this shift, machine translation and translation memory technology start converging.”
A similar concept has been introduced by Lilt; this Silicon Valley startup developed their solution from scratch on the basis of research carried out at Stanford University, rather than as an addition to an existing CAT tool like SDL. In fact SDL is now suing Lilt, alleging patent infringement. It is an unequal match: SDL has been around for a long time and is the de facto standard as a CAT tool, while Lilt was only founded in 2015 and its market capitalisation is orders of magnitude less. However, Lilt’s claims are impressive. In a case study they carried out in conjunction with a translation and localisation business on a large travel website with 1.77 million words of content, human staff “increased translation speeds far beyond the industry average of 335 words per hour. … The client’s reaction was overwhelmingly positive. The number of errors was low compared to traditional machine translation solutions and the quality in line with standard human translations. Given a two-week window, the project was actually completed within 10 days.”
Whatever the relative benefits of the solutions offered by SDL and Lilt, adaptive translation seems to be a useful addition to the tools available to human translators.