Questions1. In which ways are you already supporting working with MT that go beyond the classic post-editing?
2. And what are your plans to do that in the near future?
Answer (by Daniel Brockmann)
Trados Studio at last count offers access to 60+ providers or machine translation. Like you often say, the app store/open platform approach is something that we have always been keen on, and this is a nice proof point. At the same time, it also shows how fragmented the MT market has become, not least after the rise of NMT which has provided the quality progress that was needed for MT to become a realistic productivity helper for professional translators.
We have also decided to provide free capped NMT with Language Weaver (former SDL Machine Translation) and see thousands of users using it every day — it's actually the most used MT in Studio today. DeepL and other providers are also very prominent obviously — it's very important to have choice to have the optimum engine for the job at hand.
Why have so many providers gravitated to our platform over the years? We have the platform that is probably easiest to plug in — see previous point. We are doing the same in the cloud where we have the beginnings of an app store as well with well-known MT plug-ins available also.
I think we are providing all the features you mentioned:
— Interactive MT usage including segment by segment or fragment by fragment
— I would say that fragment/AutoSuggest was more meaningful for SMT than NMT, as NMT tends to be more fluent so it's less important to be able to delete a bad translation and just use the useful fragments — but users can still do this of course if it's better for their particular scenario.
— Many options around applying MT/TM interactively — leaving segment empty by default and use fragments, or start with a segment and then use productivity tools for efficient review
— Batch MT usage (classic PEMT use case)
— Fuzzy match repair (MT fragments correcting fuzzy matches)
— Using several engines at the same time
— Using apps to evaluate several MT engines and pick the most appropriate one — such as with
this tool— Some of the apps also go very far in supporting NMT beyond the basics — this tool is an interesting example
Our plans include these:
With Trados in the cloud, we have introduced the concept of "translation engine" where you have just one 'engine' that combines NMT, TM, terminology, rather than having to specify three separately.
This is an evolution away from looking at the three main productivity tools separately with a somewhat fragmented UX and to help users to configure the best 'engine' for a job at hand by combining the best NMT model(s), TM(s) and termbase(s) in one go.
This also holds interesting potential for the future when it comes to always providing the 'best possible match' for every segment and boosting productivity a bit more than today.
MT quality estimation is also an interesting area to explore — as this might help the translator/reviewer in the loop to focus most on those segments which might need deeper review than others.
Adaptive Machine Translation (based on NMT rather than SMT) is also a key area that we are exploring in Language Weaver + Trados as a result. It's obviously already possible to train models, but shifting this to becoming more real-time will be interesting.
Having said all this — and I think this is also key — it's important to understand that while review times may be lower than translation-from-scratch times, it is not good to squeeze rates even further by assuming kind of "quasi-zero review time." It is very key to be aware of the fact that translators need time to review, and this needs to be reflected in the rates as best as possible.
Overall, with NMT becoming better, translation work will keep shifting away from translation from scratch to reviewing more-or-less 'reasonable' initial MT suggestions. I hesitate to call this PEMT, as that term for me is quite wedded to SMT, while with NMT the cognitive task is shifting a bit more to reviewing fluent translations — which often makes it harder to spot errors such as terminology, faux sense, etc. So while overall, things are becoming more efficient, it's key to grant enough time (and payment!) for this shifting task!
— The 331st Tool Box Journal