MT Technology Development Questions related to Rapid Adaptation

Dear friends,

 

As some of you know, I have been practising and exploring machine translation for some years now, and I'd like to share with you some thoughts on the future development of MT in the hope that you might add your reactions and suggestions.

 

(Most of you I have already corresponded with on MT matters; I have also - optimistically - added MT providers Memsource, DeepL and Lilt.)

 

It is my view that the current Neural MT, surprisingly good as it is, would benefit immensely from the addition of the following features:

 

- Instant adaptation when a new translation is added (already exists for some language pairs at SDL and of course Lilt; maybe also elsewhere)

 

- Combination with data from translation memories as provided by the translator

 

- The same with termbase data

 

My questions to you, then, are:

 

- Do you agree in principle?

 

- Are you aware of any attempts in these directions (apart of course from already existing instant adaptation)?

 

- Do you think that this might be realistically achieved within a few years?

 

- Have you any idea of whether these ideas (at least the first one) are considered by the Big Two providers, i.e. Google and Microsoft? (I have the feeling that their target groups are not quite the same as those of SDL, DeepL, Systran, etc., and that therefore they would be less interested in the use of translation memories and termbases, at least. Of course it would be very good to get their reactions as well, but I also have the feeling that it is almost impossible to get them to reply.)

 

My main reason for this investigation is to find out what the future might look like for freelance translators in particular and the translation business in general. If in let's say five years we will have substantially improved MT results this would of courses affect us (not counting translators of fiction) very much. And if this improvement to a significant extent would build on translation memories and termbases that are maintained by translators, that would also be an important part of the overall picture.

 

Thank you in advance for your help.

 

Kind regards,

 

Mats Linder (Author of award winning manuals on using TRADOS) 

  • Thank you Kirti for posting this! (But I want to point out that the addition in brackets after my name was put there by Kirti, not by me. Even if it's true.)
  • I think the notion of rapid or instant adaptation is a compelling one and is very much part of the vision of the technology at SDL.

    It is natural and likely that a robust TM technology will evolve and incorporate more of an assistive role using MT that goes beyond both the raw MT role and the string matching of pure TM. In a sense bringing them both closer together in a more dynamic and interactive way.

    Attempts to bring this closer together are underway at SDL and perhaps elsewhere too.

    This article points to some of the efforts in Korea on this: www.koreaherald.com/view.php

    "“Future translators, as I envision, would be freed from the mechanical aspects of translation -- the simple substitution of words or terms in another language, for example, which most find laborious(...) Translators feel rewarded by the more sophisticated parts of their job, such as analyzing the context, styling the prose and so on. BabelTop translators would be able to do only the latter.” To achieve this, Cho and her tech team are building a cloud tool that tracks the translation process, oversees the work itinerary and builds a glossary of terminology in order to facilitate the work of translators."
  • "Do you agree in principle?"

    I think at SDL we are very much in agreement and there are several initiatives underway that attempt to bring  (TM, NMT adaptation, Termbases) much closer together and working in tandem.

    We are also finding from feedback from the many translators that use NMT at SDL that the new SDL terminology controls are powerful enough that it minimizes teh need for an"Adaptive NMT" solution. This feature allows any and every translator to create their own preferred terminology lists.

    You will begin to see this unfold in the new Language Cloud solution this summer.

    There are many other assistive-AI things that are being envisioned around anticipating and suggesting useful support material when we see how many different users work.

    Kirti