Adaptive MT Just another hype?




  • Greater than 3 minutes, my friend!

    So first of all, for the record, I don’t have any interest in LILT. I just started to use the product in a very early stage and every time they announce something new, I try to test it as well.

    Recently SDL started to talk about Adaptive MT — this new technology they call “Transformational machine translation” will be part of Studio 2017. For those of you that don’t know what Adaptive MT is, test LILT. Some of you will like it, some probably won’t. In any case, I do like it.

    If you’re using a CAT today and you feel comfortable to post-edit MT output, you are using “static” MT: the whole sentence is translated by an engine and that is what you start to post-edit. Your work is stored in the memory and as long as you don’t retrain the engine, you will only get suggestions from your memory. The machine does not learn anything from your corrections until you or someone else retrains the engine. What is in the memory does not blend with the MT engine when you’re using static MT. Also, as long as the work isn’t done, you don’t know what is the quality of the MT output. You just edit what you get. And if you fix an error you may bump into the same error somewhere else in the text.

    With LILT that won’t happen: what you edit, is immediately stored in the MT engine; new sentences that require the same translation for the same subsegment, will get your own edit, and not the output of the engine. You will recognize what you translated before even when the MT engine is not re-trained. TM and MT are blended while you are translating. What you write now, influences the suggestion for the translation of the rest of the sentence. If you make mistakes, you will get more mistakes. If you do a good job, the job will get better: the further in the document, the more of your own work is already used as a translation suggestion. When you start using LILT you use a generic system, but the more you use the system, the more it becomes YOUR system, the more it uses your voice.

    That, in a nutshell is what Adaptive MT is.

    From a technical perspective it is pretty impressive: store your translation, don’t mix it with translations from others, re-use the subsegments you corrected before and fit them in the machine translation smoothly… it’s really high-tech. But the most impressive is, I think, the user experience. Creating a translation that is suggested on the fly, has more risk to get into conflict with the translation that the translator had in mind. Suggesting the pre-translation without frustrating a translator who wishes to be productive above all, is not easy. In my opinion, for those translators that can get used to this way of translating, LILT delivers. To me it feels more like translating than post-editing.

    MT creates translations of unknown quality. This may be one of the main reasons why some translators don’t want to use MT. Adaptive MT is still producing unknown quality, but as the tool adapts to your own editing at runtime, you have much more control 1. on how sentences are translated and 2. on all future suggestions. (For the record: Fair Trade Translation helps translators to cope with the risk of using static MT; it can’t measure the quality of translations done in LILT. Both products however strive to give more control on the job to the translator.)

    If an LSP (agency) pre-translates the files it outsources to freelance translators, usually not a lot of translators like the MT output in there, especially if the engine was not trained. Adaptive MT is probably not something LSPs are considering today. It is more something a translator can use, if he wishes and if he’s happy with this way of working. I feel much more in control of my own job with this new way of working.

    But also some LSPs could benefit from this new technology: if the LSP I’m working for makes all translators work together as a team, online & in real time, I can see how an adaptive MT system might boost our productivity and the consistency of the whole team more than the existing TM systems. An LSP that just cuts big jobs in small pieces and dispatches those to individual translators that don’t work together, adaptive MT won’t help them. But maybe this new technology finally gives them a good reason to change their approach.

    My feeling is that many existing tools will try to put Adaptive MT technology in their product in the next couple of years. They will have to compare their implementation to the one of the ground breaker. It will be different as their users should still recognize their tool. It will be hard to beat LILT — LILT created their solution from scratch, with Adaptive MT in the core of the product. They already tweaked their product based on feedback of loyal users. It won’t be easy for anyone to beat the new kid on the block.

    Looking forward to some constructive battles.

    Gert Van Assche

    About Gert Van Assche

    Skeptical technology lover, testing a lot of translation tools. Occasionally I translate and develop software myself as well. I help LSPs through Datamundi, translators through Fair Trade Translation.

    11 thoughts on “Adaptive MT Just another hype?

    1. Thanks for this great explanation Gert!
      I can’t wait to see it implemented on desktop tools, but I’m afraid it will be only possible with cloud-based tools like Lilt and LanguageCloud.

    2. For now, I guess so, Pieter. I think that there are many good, technical reasons why this is the case today. If PCs get more powerful processors, if the MT engine would use all cores in a multithreaded process, if Windows would allow local services to use more of the plugged in RAM,… I see no reason why on long term Tom could not do the same with Slate Desktop.

      But I think we should not underestimate how much work it is to create a tool like LILT. My feeling is only the big players are rich enough to fund this kind of development. I also know from my own experience you need a lot of customers to get the vital, valuable feedback to improve your product. Budget + customer base will reduce the number of players that can offer this.

    3. I’ve been using LILT for a few weeks now and I’m very impressed. The suggestions based on my previous translations (within the same project) really help my productivity, as they come in groups of words inside a sentence as opposed to the old ‘complete segment’ matching system found in the tools I had been using so far – Studio 2015 does really feel old and dumb now.
      I’ve mostly used it for projects that are from scratch, with no existing TM, I have yet to experiment with projects that rely heavily on large and old TMs that are quite often riddled with poor quality or out of context segments. But so far it’s been very positive and the UX is the best I’ve ever had with any CAT tool, it’s powerful, straightforward and you can see the benefits almost immediately.
      I have no interest in Lilt either!

    4. Great article, Gert. Thanks for mentioning me and Slate Desktop in your comments. Regrettably, I can’t comment about first-hand Lilt experience because I’m not a translator. I’ll say I think they have a great team that created an inspirational and novel service.

      Re adaptive MT hype, I’m taken by tool vendors sudden attention to serving the translator. It seems the Lilt and SD teams were working simultaneously in our skunk works. Soon after our near parallel announcements (late summer 2015), SDL, MateCat, et al started popping up with their offerings. As they say, success is relative… the more success you have, the more relatives come out of the woodwork. For now, I’ll let others judge the boundaries of hype while I look for some time to post my own blog.

      Re doing something similar with SD, I’m monitoring the opportunities. Like many things in SMT, dynamic adaptation was pioneered in speech recognition. Dynamic tuning for dictation engines premiered with Dragon 9 in 2006. Based on my speech background, I have some ideas where this technology could go.

      The challenge isn’t related to the PC’s power. Lilt’s local components run in a browser mostly with JavaScript. It’s not exactly an optimized environment (although JavaScript is much better than it used to be). There’s plenty of horsepower on most all desktops.

      SD’s challenge is our strategy to serve the largest possible translator population via the CATs that they already use. Lilt chose a different strategy to require translator to abandon CATs they already use in favor of Lilt’s new CAT online service.

      Our strategy puts us inside each CAT’s MT connector API, i.e. if the CAT vendor publishes one. We rely on the CAT vendor if they don’t. From inside the connector, we’re limited to the app events and API that the host CAT gives us. Trados Studio is the only CAT (I’m aware of) that exposes enough of its API to control and respond to the user’s interaction from an MT plugin. There’s already an open source Trados plugin implementation of MT4CAT (predictive typing with Moses). If when we support that from SD, we don’t have to start from scratch. That said, I’m happy to work with any CAT vendor who wants to jointly develop predictive adaptations for their products.

      I’ve said this in other forums, I predict Amazon will release an online CAT service with predictive translation late this year or early next year. Let’s see if my crystal ball is tuned to the cosmos 🙂

    5. I’m very much looking forward to testing SDL’s version of Adaptivity, in Studio 2017. The only problem is that it will only work with SDL’s own Language Cloud engines, which are nowhere near as good as vanilla Google Translate (or Microsoft Translator) via the API. I haven’t yet properly tested the engine quality of Lilt, but am curious how it compares to GT.

    6. @Gert Thanks a lot for the thorough review of adaptive MT. As a former freelance translator and now developer of translation tools (amongst other things), I am wondering what potential users (translators and LSPs) here think about switching to a system like Lilt. What does it take to change tools, especially moving to a cloud-based one? What are the key considerations? Privacy? Security? Configurability? Anything else?

    7. Martin, it would be good if other translators would answer your question. For me privacy is not an issue as my customers know I play with tools and that I may always opt to do a job in the cloud. Just like Michael I’m really curious to see how SDL is going to do it. I can imagine some translators are reluctant to do stuff or put stuff in the cloud, but as I explained I don’t see how Adaptive MT can be done without sending any data to the internet. In the end you will always need an MT engine to translate part of a sentence for you. As long as you don’t have a local MT server (on your own network or on your own PC running as a service, and accessible by your CAT), MT can’t help you in any case. An omelet & egg story.

    8. Martin, I can’t address considerations to “change tools” because I’m not a tool user, but I’ll share the main reasons we created Slate Desktop for the desktop not the cloud. You might find it surprising that it has nothing to do with security or privacy. Each translator’s privacy and security requirements are different. I’m preparing a separate blog post about these issues.

      Early in my many years working with and supporting translators, I learned that the “workstation” experience is very personal and intimate to translators. The physical keyboard, mouse, (dual) monitor, editor software, and even today’s dictation software… each translator has unique wants, needs and comfort zone. I can’t count the number of times I heard, “Don’t mess with my workstation” (keyboard, mouse, software, etc).

      We created SD to augment a translator in his environment rather than force him to abandon his current tools. That single strategy was the primary force behind our a gargantuan effort to make Moses run on Windows. Most translators use Windows and won’t change to Linux. All other Moses implementations are in the cloud because it’s the only way to deliver Linux-bound technologies to Windows users. If a translator uses a CAT (regardless of OS), he’s probably using the CAT he likes, or at least already knows how to use. Why force him to abandon what he knows to learn yet another tool?

      While we were debugging cross-platform C++ code, the Lilt team was working on a different and just as enormous effort to orchestrate all the technologies that inter-operate behind the scenes in Lilt. As Gert stated, “we should not underestimate how much work it is” to create these tools. I applaud the Lilt team for their effort and results.

      After all’s said and done, I think translators settle in for the long haul with a CAT they like, where “like” ranges from learning curve to features. For short-term requirements, they’ll use other CATs when a customer requires. The strengths and weaknesses associated with the
      cloud vs desktop option just become a few more feature bullet points and I’ll list them in my pending blog.

    Leave a Reply

    The Open Mic

    Where translators share their stories and where clients find professional translators.

    Find Translators OR Register as a translator