At the beginning of last year, we asked the question whether
Artificial Intelligence can solve the translation challenge in Learning. Wouldn’t it be a breakthrough, if the traditional, highly manual – slow, cumbersome, and expensive translation processes could be complemented - and eventually be replaced – by machines who can do this on the click of a button? No more weeks of waiting time for “human” agencies to complete the job, only to wait for another couple of weeks for a human expert to review what’s coming out of the agency. That is: if you even find an expert in the first place, which is often a bottleneck. How about if instead translations of all assets could be processed instantly , as needed to provide learning assets in each language right away with the availability of the original language version? In times of Cloud solutions with release cycles of only 3 month or shorter, sure enough there is no time to waste! And how about if these translations would be of really high quality, not only language-wise, but also considering all your special company-lingo?
This vision is now becoming a reality: On December 2nd, we will go live with the first set of Neural Machine translated (NMT) assets on SAP Learning Hub!
This first set of learning assets includes topics like SAP HANA and SAP S/4HANA, translated to German, French, Spanish, Portuguese, Russian, Korean, Mandarin and Japanese.
Massively upscaling language availability of learning assets
At this point, machine translations will not replace the portfolio of human translations at SAP which we deliver with very extensive quality control for our top learning assets. While these will further be continued, the machine now complements the translation portfolio for many asset-language combinations that so far could be delivered in English only. We are therefore able to increase the coverage of language-specific learning assets from an elite set of top assets to the vast mainstream, and – ultimately, so the mid-term aspiration – ALL learning assets.
Starting on a high level today – and continuously improving translation quality over time
But how good is the quality really? Are deep and detailed learning assets, focused on complex IT and Business Software topics for professionals not too big a nut to crack for the machine? Stunningly, the review process of this first set of assets that will now be published on SAP Learning Hub has left our experts amazed about language quality coming out of the machine:
Is it perfect? By no means. Well, in fact the language quality and flow of sentences already comes close to a large extent. Often you can read several pages before you discover a first little mis-translation that indicates that this was not translated by a human native speaker.
We did face some bigger flaws with SAP-specific terminology, though, so we’ve set up quick-checks to catch and correct all big-fish terminology mistakes prior to publishing on SAP Learning Hub. The good news is: each of those corrections are not only applied once for what is now being published, but captured in our translation “memory”, which means that for all future translation of these terms – wherever they may come up - the engine now knows how to do it right.
Another challenge are texts in graphics: depending on the language, a translated text might be significantly longer or shorter in the target language, compared to the English original, which may lead to a graphic that includes texts not looking ideal in the target language. To date we don’t have an automatic mechanism to correct these, but for the most part this is not representing a true obstacle to learning in the target language.
The good news is, that this does not only represent, as we do believe, a very valuable starting point for our learners already today, but that we will see a continuous and unstoppable trend of further quality improvement, driven mainly by three elements:
- Ever-increasing Capability of NMT engines
At SAP, we use the best Neural Machine Translation engines available on the market. In lock-step with the steep learning curve of these engines, we will see further improvement of our translation qualities as well. Think about where machine translation quality was 5 years ago, and where it is now – and imagine where it will be in 5 years from now! We’ve build an open architecture that will allows to easily switch to whatever the best engine will be, whenever necessary.
- Growing “Memory”
As outlined before, every flaw that is found and corrected is captured in our translation “memory”, to ensure that for all future this term is being translated correctly.
- Using feedback from the Crowd
All machine translated language versions on SAP Learning Hub feature a one-click feedback process for everyone who has discovered a wrong or suboptimal translation to directly alert our experts. Paired with the “memory” logic, the more we feedback we get, the faster the learning curve of this memory!
Reality Feedback will drive further expansion of the approach
So, as the first larger set of machine-translated assets goes out, it is of utmost importance for us to collect feedback about these directly from the learner community on SAP Learning Hub. Based on this feedback, we will decide where to focus on further improvements, and where and how fast we will expand translation coverage to more - and ultimately all – learning assets.
By when will the machine deliver completely perfect results? Interested in your thoughts and opinions in the comments. I would bet it won’t be too far out in the future!
