After the rush to embrace neural machine translation (NMT), it has finally arrived at our -- the translators' -- doorstep with adaptive models. And at this point, not just one has been released but three, two just in these past couple of weeks. They really are three different kinds of models because they each use a very different approach to adapting the MT output -- so different, in fact, that I wonder whether they should all be called by the same "adaptive" moniker.
Let's start with the first one, KantanMT's neural system, which has been in use since March of this year.
As I talked with KantanMT's CEO Tony O'Dowd about this, it became increasingly clear to me how disadvantaged an MT provider is who does not also offer a full-fledged translation environment. To put it differently: There are equally important fields of responsibility on the side of the translation environment provider and machine translation provider to make machine translation work smoothly and productively in a translator's workflow.
Let's back up first, though. Here's how KantanMT deals with neural machine translation. As for any neural machine translation system, KantanMT's initial processing requirements to train neural engines are very high. Tony mentioned that for the first training pass, the processing of a 100-million-word bilingual corpus takes three to four days. It's overwhelming to imagine that this would have to be done again and again for an adaptive system. The nice thing about neural machine translation, however, is that the ongoing training for adapting the translation model does not require a complete retraining (as it did in many cases with statistical machine translation, with the notable exception of Lilt -- see below).
KantanMT has set up the incremental training of the neural MT engines by collecting data in a TM of sorts (which first is just kept in the temporary cache of the server but is converted into a file once the user logs out). At certain intervals (but not every time a segment is finalized), this TM is fed into the neural engine for adaptive purposes. In the meantime (and after the passing of the data), the TM (in KantanMT-speak "Total Recall") is used as a translation memory that acts very much like most TMs in combination with MT: perfect matches and any match with a fuzzy match rate of 85% or higher is preferred to MT suggestions and entered automatically to be confirmed or edited by the translator. Otherwise, the neural machine translation engine suggests translations. Since the engine in most cases is solely trained on the data that the client has provided and is being adapted and improved as the projects are being processed, there is a relatively good chance that the suggestions are of reasonably good quality (at least as far as terminology is concerned).
You might have the same question I had: How does this process using a front-end TM differ from any other translation environment tool? It doesn't really, if not for the fact that both the MT and the TM are cloud-based and therefore available to everyone in a real-time translation team. And while this is also available in other server-based workflows offered by the various translation environment tools, more often than not, this is not set up for projects that you and I work on.
But here is the hitch: While many translation environment tools support the use of KantanMT, only a handful allow its use upstream (i.e., sending data back to the KantanMT's cloud so it can then be used across translators). The ones that do include Memsource, Trados Studio, and (in a limited fashion) Across. Tony's (unsurprising) prediction for 2018? "Next year adaptive MT will become the big thing that will be seen in CAT tools," meaning that tools like KantanMT will have more complete access.
The main clientele for KantanMT is clearly the translation buyer. Early on there were plans to also offer products to other stakeholders, but its easy access to large amounts of focused data made the translation buyer the primary target. Still, additional features that would be helpful for translators would be welcomed, such as more interactivity between the machine translation engine and the "Total Recall" TM engine to "fix" TM matches (with MT) or MT suggestions (with TM).
The other engine that has just gone live (within Translated's translation environment tool MateCat and the paid Pro edition of the MyMemory app for SDL Trados Studio) is ModernMT. ModernMT is a three-year EU-funded project with a number of partners, including Translated, much like MateCat itself was a few years ago. If you remember, MateCat's original purpose was to "investigate the integration of MT into the CAT workflow [with] self-tuning MT that adapts MT to specific domains or translation projects and user-adaptive MT that quickly adapts from user corrections and feedback" (Source: Proceedings of the 17th Annual Conference of the European Association for Machine Translation). While the adaptive system worked reasonably well, that part was unceremoniously and frustratingly dropped from MateCat, and the EU agreed to confer another three-year contract. This time the adaptive MT is here to stay, according to Translated's Alessandro Cattelan, whom I spoke to for this report.
You gotta feel for all these MT developers working so hard on the cutting edge of technology when, just like that, a new and improved technology enters the field and, alas, all their prior work is ready for the trash heaps of technology history. I'm not completely sure this is entirely how it went with ModernMT, but I would guess it's pretty close, since only this summer they turned their attention to neural machine translation after having spent more than two years on statistical, adaptive MT. Amazingly enough, they were still able to present a result just a few months later.
The adaptive part of the technology is fundamentally different than other adaptive engines because there are actually no changes in the baseline engine happening at any time. Instead, the system uses a technology called "instance-based adaptive NMT." Similar to KantanMT but for a different purpose, this consists of the translation request first being sent to a TM layer (which can consist even of a relatively small TM as long as it's highly-tuned). With similar segments found in that TM layer, the NMT engine's "hyperparameters" are adapted on-the-fly so that a more suitable suggestion is generated. This concept is based on this paper by the Fondazio Bruno Kessler, which is part of the consortium working on ModernMT.
The benefit -- in theory -- is that you don't ever need to actually train a specific MT engine, but you can instead use a large generic engine whose suggestions are specialized by having the query parameters adapted as the translation is happening.
I tried to work with the engine in MateCat after I talked with Alessandro, but I wasn't able to get any suggestions other than the ones from Google Translate (you can't actually choose the engine because the system selects it for you).
I have to be honest and say that I don't completely understand the concept of this technology, but if it is indeed able to produce better results, then all power to it. I am pretty sure, though, that "adaptive" is a bit of a misnomer. Not because the machine translation suggestion is not being adapted (it is!), but because the process is different from what is typically understood as adaptive MT. Maybe a term like "responsive," "reactive," or even "ad-hoc-tuned" might be better.
ModernMT is completely open-source (in fact, you can download everything from its website), but that does not include the data, of course. This is where Translated's hyper-TM MyMemory comes in and where the company sees a possibility for itself to market this solution -- to translators and LSPs as a paid service through MateCat, Trados, and possible other translation environments and to large translation buyers as an in-house solution. (According to Alessandro, at least one large Silicon Valley company has already done some extensive testing and found the results better than those from Microsoft Translator Hub, the trainable machine translation engine Microsoft offers.)
I also asked Alessandro how the non-adapted baseline engine compares with the Google Translate NMT engine, and he was honest enough to say that it produces worse results -- unless the adaptive process is taking place and then it's better.
Lilt has also released a neural MT engine (for now only between English and German and English and Chinese and only for select users -- see more on this below). Since the team at Lilt consists of remarkably young and energetic folks, they not only introduced their new engine but also completely converted the user interface, oh, and rebranded (see below as well).
And, yes, they also settled the silly lawsuit SDL had brought against them (see edition 273 of the Tool Box Journal). (The very non-communicative and, it seems, legally prescribed official statement from Lilt: "Lilt is pleased with the resolution of this dispute. The settlement was mutually agreeable. We will continue to focus on products and services that democratize access to information.")
Anyway, I was really intrigued and -- I admit it -- at first puzzled about Lilt's move toward neural machine translation. Here's why: Lilt essentially broke down all the borders that existed between the different assets translators use by creating one single database for all the data used by the (at that point, statistical) machine translation, the translation memory engine, and the automated term extraction/termbase engine. This was made possible because the data was sitting in an open text-based table that could be utilized for any of these options. Even though I'm not super-technical, I also knew that wasn't possible when it came to neural machine translation, where the translation model really consists of numbers rather than language data. So my confusion had to do with me wondering whether Lilt had thrown all of its basic concepts overboard.
Turns out it hasn't, at least not completely. As in the previous incarnation, there is still a massive text-based table containing data that the TM and the termbase features are based on and that is being used to train the neural machine translation engine initially and continually. (The difference involves a different level of immediacy of the data in the data table and how it's used for machine translation purposes.)
Lilt's NMT engine is trained on the same bilingual data as the previous SMT engine (that's at least true for the four existing language combinations) minus the monolingual data needed for SMT but not for NMT.
The engine interacts with the user as he or she translates through the same process as previously: Every time a word is entered, the machine translation engine calculates a new suggestion for the rest of the segment to match the entered data. (Lilt's Carmen Heger, whom I talked to about this, mentioned a barely noticeable increased delay in the response. She was right, it's hardly noticeable; in fact, I didn't notice it.)
Also like before, every time you confirm the translation of a segment, the machine translation data model (which now runs parallel to the data table) is immediately adapted -- in fact, much more efficiently than before. And just as in the case of KantanMT, while the initial training and so-called batch updates (like when you upload a whole translation memory) are very resource- and processing-intensive, incrementally adding to the neural engine is not, allowing it to be done on regular CPU servers (vs. the more powerful GPU servers).
So, what about the quality of the suggestions? One important consideration when answering this question is that with Lilt it does not primarily have to be the initial suggestion for the whole segment to be useful as long as the constantly evolving suggestions for each segment are useful. In my very limited and unscientific testing in EN>DE, the initial suggestions seemed OK but tended to be less accurate than Google Translate or DeepL (for test purposes I usually copied and pasted the same segment in the other interfaces). But when I diverged from the original suggestion they became ever more helpful, and tended to be more helpful than in the earlier SMT-based Lilt.
(Here is a bit more scientific data: Lilt ran a case study with one of their corporate partners comparing the statistical vs. neural quality by professional translators. The results: NMT was chosen 44% of the time, 29% of the time translators chose SMT, and in 27% of the time it was about equal.)
Since we're talking about corporate clients, you might have seen that Lilt has also done some rebranding with the slogan, "The New Engine for Enterprise Translation Workflows." What does this mean for translation professionals? Not much as far as the availability of the tool, but it does reveal that we haven't been as eager to switch to a new technology paradigm as some might have expected. And that in turn may point to something more profound (forgive the over-generalization): Traditionally translation professionals have been very slow to accept technology. But once they accepted translation technology as a normal part of their work environment, they were (and are) just as slow to change their chosen technology (which was so hard to accept in the first place). Maybe it's just the human condition, but I think what we might not completely understand is that being technologically adept by definition means a constant willingness to change the technology that becomes available as it develops.
Now, I'm not saying this because I think Lilt is the one and only kind of technology out there. It isn't. There are plenty of interesting new tools and technologies. And if you honestly look at yourself, are you completely open to continuously moving on? (I have to admit that I'm not, and I also know that I would be a better translator if I were.)
To finish this up, as mentioned above, Lilt also has a complete new interface as you can read in this document. I like that there's no separate editing interface, it's now very easy to switch between a vertical and a horizontal display, and there is a better sense of accessibility -- even as far as accessing available options and keyboard shortcuts.
I did not like the "on-demand tag editor" which, while a good idea in general, did not work as seamlessly as it should have.
Oh, and I forgot to mention that the neural mode is at this point available only for corporate clients (and LSP clients on demand), but it will be available for all sometime in the new year with all language combinations enabled.