Artificial Intelligence: And You, How Will You Raise Your AI?

This is the final post for the 2017 year, a guest post by Jean Senellart who has been a serious MT practitioner for around 40 years, with deep expertise in all the technology paradigms that have been used to do machine translation. SYSTRAN has recently been running tests building MT systems with different datasets and parameters to evaluate how data and parameter variation affect MT output quality. As Jean said:

" We are continuously feeding data to a collection of models with different parameters – and at each iteration, we change the parameters. We have systems that are being evaluated in this setup for about 2 months and we see that they continue to learn."

This is more of a vision statement about the future evolution of this (MT) technology, where they continue to learn and improve, rather than a direct reporting of experimental results, and I think is a fitting way to end the year in this blog.

It is very clear to most of us that deep learning based approaches are the way forward for continued MT technology evolution. However, skill with this technology will come with experimentation and understanding of data quality and control parameters. Babies learn by exploration and experimentation, and maybe we need to approach our continued learning, in the same way, learning from purposeful play. Is this not the way that intelligence evolves? Many experts say that AI is going to be driving learning and evolution in business practices in almost every sphere of business.

 ===================

Artificial Intelligence is the subject of all conversations nowadays. But do we really know what we are talking about? And if instead of looking at AI as a kind of software that is ready to use and potentially threatening for our jobs, what if it was thought to be an evolutionary digital entity, with an exceptional faculty of learning? Therefore, breaking with the current industrial scheme of the traditional software that requires code to be frozen until the next update of the system. AI could then disrupt not only technology applications but also economic models.

Artificial Intelligence does not have any difficulty to quickly handle exponential growing volumes of data, with exceptional precision and quality. It thus frees valuable time for employees to communicate internally with customers, and to invest in innovative projects. By allowing an analysis of all the information available for rapid decision-making, AI is truly the corollary of the Internet era, which is also the result of all threats, whether virtual or physical.

Deep Learning and Artificial Neural Networks: an AI that is constantly evolving

Deep Learning and artificial neural networks offer infinite potential and a unique ability to continually evolve in learning. By breaking with previous approaches such as the statistical data analysis approach which demonstrates a formidable, but trivial, memorization and calculation capacity, like databases and computers, the neuronal approach gives a new dimension to artificial intelligence. For example, in the field of automatic translation, artificial neural networks allow the "machine" to learn languages as we do when we are in an immersion program abroad. Thus, these neural networks have never finished learning, after their initial learning, they can then continue to evolve independently.

A Quasi "Genetic Selection"

Training a neural model is, therefore, more akin to a mechanism of genetic selection, such as those practiced in the agro-food industry, than to a deterministic programming process: in all the sectors used by neural networks, the AI is selected to keep only those learnings that progress best and the fastest or most adaptable to a given task. These AI techniques that are used for automatic translation are even customized according to customer needs - business, industry and specific vocabulary. Over time, some AI techniques will grow in use, and others will disappear because they will not demonstrate enough learning and will not be sufficiently efficient. DeepMind illustrates this ability to the extreme. It was at the origin of AlphaGo, the first algorithm to beat the human. AlphaGo had learned thousands of games played by human experts. The company then announced the birth of an even more powerful generation of AI. It managed to learn the game of Go without playing games played by humans, but "simply" by discovering, all alone and in practice, the strategies and subtleties of this game in a fraction of the time of the original process. Surprising, isn’t it?

 

Machines and Self-Learning Software

The next generation of neural translation engines will exploit this intrinsic ability of neural models to learn continuously. It will also build on the ability of two different networks to have unique pathways of progress. Specifically, from the same data, like two students, these models can improve by working together or competing against each other. This second generation of AI  is very different because not only does it have models taught from existing repositories (existing translations), but just like newborn ones, they also learn to… learn over time, placing them in a long perspective. This is the lifelong learning: once installed in production, for example in the customer information system, AI continue to learn and improve.

To each his own AI tomorrow?

Potentially, tomorrow's computer systems may be built, like a seed planted on your computer, or in everyday objects. They will evolve to better meet your needs. Even if current technologies, especially software, are customized, these technologies remain 90% similar from one user to another because they are built into bundled and standard products. They cost very dearly because their cycles of development are long. The new AI, which tends to tailored solutions, is the opposite of this. In the end, it is the technology that will adapt to man and not the opposite, as it is at present. Each company will have the specific technical means, "hand-sew" and you may have an AI at home that you can raise yourself!

Towards a new industrial revolution?

What will be the impact of this evolution on software vendors or on IT services companies? And beyond, over the entire industry? Will businesses need to reinvent themselves to bring value to one of the stages of the AI process, whether it be adaptation or quality of service? Will we see the emergence of a new profession, the "AI breeders"?

In any case, until the AI is seen as a total paradigm shift, it will continue to be seen as software 2.0 or 3.0. A vision that hinders innovation and could make us miss all of its promises, especially to free ourselves from repetitive and repetitive tasks to restore meaning and pleasure to work.

Jean Senellart, CTO, SYSTRAN SA

The Most Popular Posts of 2017

2017 was an especially big year for the Neural MT momentum on multiple fronts. We saw DeepL and Amazon introduce new generic NMT product offerings, each with a unique twist of their own. They are both impressive introductions but are limited to a small set of language pairs, and both these companies have made a big deal about superior MT quality on using some limited tests to document this and attract attention. For those who define better quality by BLEU scores, these new offerings do indeed have slightly higher scores than other generic MT solutions. But for those who wish to deploy an industrial scale MT solution, there are other things that matter more e.g. the extent of customization possibilities and the range of steering possibilities available to update and tune the base engine, the overall build platform capabilities, and the ability to secure and maintain data privacy are particularly important. The Asian market also saw several NMT initiatives build momentum, especially with Baidu and Naver.

All of the "private" MT vendors have also expanded their NMT offerings, focusing much more on the customization aspects and two vendors (Lilt & SDL) have also introduced Adaptive NMT for a few languages. These Adaptive NMT offerings may finally get many more translators using MT on a regular basis as more people realize that this is an improvement over existing TM technology. We should expect that these offerings will grow in sophistication and capability over the coming year.

SDL Quality Measurements on Various Types of Customization on SMT & NMT

While industry surveys suggest that open source Moses is the most widely used MT solution in professional settings, I still maintain that many, if not most of these systems will be sub-optimal as the model building and data preparation processes are much more complicated than most LSP practitioners expect or understand. NMT currently has four open source toolkit alternatives, and some private ones thus the complexity for do-it-yourself practitioners escalates. However, in some ways, NMT is simpler to build if you have the computing power, but it is much harder to steer as was described in this post. The alternatives available thus far include:

• OpenNMT - Systran & Harvard
• Tensorflow - Google
• Nematus - University of Edinburgh
• Fairseq - FaceBook
• and now Sockeye from Amazon  which allows you to evaluate multiple options (for a price I am sure).

2018 looks like a year when MT deployment will continue to climb and build real momentum. Strategic Business advantage from MT only comes if you build very high-quality systems, remember that we now have great generic systems from DeepL, Google, Microsoft, and others. Translators and users can easily compare LSP systems with these public engines.

While the technology continues to advance relentlessly, the greatest successes still come from those who have the "best" data and understand how to implement optimal data preparation and review processes best for their deployments.

The most popular blog posts for the year (2017) are as follows:

1. A Closer Look at SDL's recent MT announcements : -- a detailed look at the SDL ETS system and their new NMT product announcements.  I initially set out with questions that were very focused on NMT, but the more I learned about SDL ETS, the more I felt that it was worth more attention. The SDL team were very forthcoming, and shared interesting material with me, thus allowing me to provide a clearer picture in this post of the substance behind this release, (which I think Enterprise Buyers, in particular, should take note of), and which I have summarized in this post.

2. Post Editing - What does it REALLY mean?  :-- This is a guest post by Mats Dannewitz Linder that carefully examines three very specific PEMT scenarios that a translator might face and view quite differently. There is an active discussion in the comments as well.

3. The Machine Translation Year in Review & Outlook for 2017  :-- A review of the previous year in MT (2016) and a prediction of the major trends in the coming year which turned out to be mostly true, except that Adaptive MT never really gained the momentum I had imagined it would with translators.

4. Data Security Risks with Generic and Free Machine Translation:--   An examination of the specific and tacit legal agreements made for data re-se when using "free" public MT and  possibilities on how this can be avoided or determined to be a non-issue.

5. Private Equity, The Translation Industry & Lionbridge :-- A closer examination of the Private Equity rationale which seems to be grossly misunderstood by many in the industry. PE firms typically buy controlling shares of private or public firms, often funded by debt, with the hope of later taking them public or selling them to another company in order to turn a profit. Private equity is generally considered to be buyout or growth equity investments in mature companies.

6. The Problem with BLEU and Neural Machine Translation :--  The reasons for the sometimes excessive exuberance around NMT are largely based on BLEU (not BLUE) score improvements on test systems which are sometimes validated by human quality assessments. However it has been understood by some that BLEU, which is still the most widely used measure of quality improvement, can be misleading in its indications when it is used to compare some kinds of MT systems. This post describes how BLEU can sometimes under-represent performance of NMT systems.

7. From Reasoning to Storytelling - The Future of the Translation Industry   :-- This is a guest post by Luigi Muzii, on the future of translation, written in what some may say is an irreverent tone. For those participants who add very little value in any business production chain, the future is always foreboding and threatening, because there is often a sense that a reckoning is at hand.  It is easier to automate low-value work than it is to automate high-value work. Technology is often seen as a demon, but for those who learn to use it and leverage it, it is also a means to increase their own value addition possibilities in a known process, and increase one's standing in a professional setting. While translating literature may be an art, most business translation work I think is business production chain work. These are words translated to help you sell stuff, or help customers better use stuff that has been sold to them, or now increasingly it is a way to understand what customers who bought your stuff think, about the user and customer experience.

8. Never Stop Bullshitting: Or Being Popular in the Translation Industry  :--  This is a guest post by Luigi Muzii (and his unedited post title). Luigi likes to knock down false idols and speak plainly, sometimes with obscure (to most Americans anyway) Italian literary references. I would characterize this post as an opinion on the lack of honest self-assessment and self-review that pervades the industry (at all levels) and thus slows evolution and progress. Thus, we see that industry groups struggle to promote "the industry", but in fact, the industry is still one that "gets no respect" or is "misunderstood" as we hear at many industry forums.  While change can be uncomfortable, real evolution also results in a higher and better position for some in the internationalization and globalization arena. Efficiency is always valuable, even in the arts.

9. The Ongoing Neural Machine Translation Momentum :--   This is largely a guest post by Manuel Herranz of Pangeanic, slightly abbreviated and edited from the original, to make it more informational and less promotional. Last year we saw FaceBook announce that they were going to shift all their MT infrastructure to a Neural MT foundation as rapidly as possible, this was later followed by NMT announcements from SYSTRAN, Google, and Microsoft.

10. Creative Destruction Engulfs the Translation Industry: Move Upmarket Now or Risk Becoming Obsolete :--  This is a guest post by Kevin Hendzel whose previous post on The Translation Market was second only to the Post-editing Compensation post, in terms of long-term popularity and wide readership on this blog. This new post is reprinted with permission and is also available on Kevin's blog with more photos. I am always interested to hear different perspectives on issues that I look at regularly as I believe that is how learning happens.

Wishing You all a Happy, Healthy and 

Prosperous New Year

Post Editing - What does it REALLY mean?

While many people may consider that all post-editing is the same, there are definitely variations that are worth a closer look. This is a guest post by Mats Dannewitz Linder that digs into three very specific PEMT scenarios that a translator might view quite differently. Mats has a more translator-specific perspective and as the author of the Trados Studio manual, I think provides a greater sensitivity to the issues that do matter to translators. 

From my perspective as a technology guy, this post is quite enlightening as it provides real substance and insight on why there have been communication difficulties between MT developers and translator editors. PEMT can be quite a range of different editor experiences as Mats describes here, and if we now factor in the change that Adaptive MT can have, we now have even more variations on the final PEMT user experience.  

I think a case can be made for both major cases of PEMT that I see from my vantage post, the batch chunk mode and the interactive TU inside the CAT mode.  Batch approaches can make it easier to do multiple corrections in a single search and replace action, but interactive CAT interfaces may be preferred by many editors who have very developed skills in a preferred CAT tool. Adaptive MT, I think, is a blend of both and thus I continue to feel that it is especially well suited for any PEMT scenario as described in this post. The kind of linguistic work done for very large data sets is quite different and focuses on correcting high-frequency word patterns in bulk data, described in this post: The Evolution in Corpus Analysis Tools. This is not PEMT as we describe here, but it is linguistic work that would be considered high value for eCommerce, customer support and service content and any kind of customer review data that has become the mainstay of MT implementations today. 

For those in the US, I wish you a Happy Thanksgiving holiday this week, and I hope that you enjoy your family time. I have pointed out previously, however, that for the indigenous people of the Americas, Thanksgiving is hardly a reason to celebrate.“Thanksgiving” has become a time of mourning for many Native People, hopefully, this changes, but it can only change when at least a few recognize the historical reality and strive to alter it in small and sincere ways.

The emphasis and images below are all my doing so please do not blame Mats for them.

==========

I have read – and also listened to – many articles and presentations and even dissertations on post-editing of machine translation (PEMT), and strangely, very few of them have made a clear distinction between the editing of a complete, pre-translated document and the editing of machine-translated segments during interactive translation in a CAT tool. In fact, in many of them, it seems as if the authors are primarily thinking of the latter. Furthermore, most descriptions or definitions of “post-editing” do not even seem to take into account any such distinction. All the more reason, then, to welcome the following definition in ISO 17100, Translation services – Requirements for translation services:

      post-edit

      edit and correct machine translation output

Note: This definition means that the post-editor will edit output automatically generated by a machine translation engine. It does not refer to a situation where a translator sees and uses a suggestion from a machine translation engine within a CAT (computer-aided translation) tool.

And yet… in ISO 18587, Translation services – Post-editing of machine translation output – Requirements, we are once again back in the uncertain state: the above note has been removed, and there are no clues as to whether the standard makes any difference between the two ways of producing the target text to be edited.

This may be reasonable in view of the fact that the requirements on the “post-editor” arguably are the same in both cases. Still, that does not mean that the situation and conditions for the translator are the same, nor that the client – in most cases a translation agency, or language service provider (LSP) – see them as the same. In fact, when I ask translation agencies whether they see the work done during interactive translation using MT as being post-editing, they tell me that it’s not.

But why should this matter, you may ask. And it really may not, as witness the point of view taken by the authors of ISO 18587 – that is, it may not matter to the quality of the work performed or the results achieved. But it matters a great deal to the translator doing the work. Basically, there are three possible job scenarios:

1. Scenario A:- The job consists of editing (“post-editing”) a complete document which has been machine-translated; the source document is attached. The editor (usually an experienced translator) can reasonably assess the quality of the translation and based on that make an offer. The assessment includes the time s/he believes it will take, including any necessary adaptation of the source and target texts for handling in a CAT tool.
2. Scenario B:- The job is very much like a normal translation in a CAT tool except that in addition to, or instead of, an accompanying TM the translator is assigned an MT engine by the client (normally a translation agency). Usually, a pre-analysis showing the possible MT (and TM) matches is also provided. The translator is furthermore told that the compensation will be based on a post-analysis of the edited file and depend on how much use has been made of the MT (and, as the case may be, the TM) suggestions. Still, it is not possible for the translator either to assess the time required or the final payment. Also, s/he does not know how the post-analysis is made so the final compensation will be based on trust.
3. Scenario C:- The job is completely like a normal translation in a CAT tool, and the compensation is based on the translator’s offer (word price or packet price); a TM and a customary TM matches analysis may be involved (with the common adjustment of segment prices). However, the translator can also – on his or her own accord – use MT; depending on the need for confidentiality it may be an in-house engine using only the translator’s own TMs; or it may be online engines with confidentiality guaranteed; or it may be less (but still reasonably) confidential online engines. Whatever the case, the translator stands to win some time thanks to the MT resources without having to lower his or her pricing.

In addition to this, there are differences between scenarios A and B in how the work is done. For instance, in A you can use Find & replace to make changes in all target segments; not so in B (unless you start by pre-translating the whole text using MT) – but there you may have some assistance by various other functions offered by the CAT tool and also by using Regular expressions. And if it’s a big job, it might be worthwhile, in scenario A, to create a TM based on the texts and then redo the translation using that TM plus any suitable CAT tool features (and regex).

Theoretically possibly, but practically not

There is also the difference between “full” and “light” post-editing: Briefly, the former means that the resulting text is comprehensible and accurate, but the editor need not – in fact, should not – strive for a much “better” text than that, and should use as much of the raw MT version as possible. The purpose is to produce a reasonably adequate text with relatively little effort. The latter situation means that the result should be of “human” translation quality. (Interestingly, though, there are conflicting views on this: some sources say that stylistic perfection is not expected and that clients actually do not expect the result to be comparable to “human” translation.) Of course these categories are only the end-points on a continuous scale; it is difficult to objectively test that a PEMT text fulfils the criteria of one or the other (is the light version really not above the target level? is the full version really up to the requirements?), even if such criteria are defined in ISO 18587 (and other places).

Furthermore, all jobs involving “light-edit” quality is likely to be avoided by most translators 

Source: Common Sense Advisory

These categories mainly come into play in scenario A; I don’t believe any translation agency will be asking for anything but the “full” quality in scenario B. Furthermore, all jobs involving “light” quality is likely to be avoided by most translators. Not only does it go against the grain of everything a translator finds joy in doing, i.e. the best job possible; experience also shows that all the many decisions that have to be made regarding which changes need to be made and which not often take so much time that the total effort with “light” quality editing is not much less than that with “full” quality.

Furthermore, there are some interesting research results as to the efforts involved, insights which may be of help to the would-be editor. It seems that editing medium quality MT (in all scenarios) takes more effort than editing poor ones – this is cognitively more demanding than discarding and rewriting the text. Also, the amount of effort needed to detect an error and decide how to correct it may be greater than the rewriting itself and reordering words and correcting mistranslated words takes the longest time of all. Furthermore, it seems that post-editors differ more in terms of actual PE time than in the number of edits they make. Interestingly, it also seems that translators leave more errors in TM-matched segments than in MT-matched ones. And the mistakes are of different kinds.

These facts, plus the fact that MT quality today is taking great steps forward (not least thanks to the fast development of neural MT, even taking into account the hype factor), are likely to speed up the current trend, which according to Arle Lommel, senior analyst at CSA Research and an expert in the field, can be described thus:

"A major shift right now is that post-editing is being replaced by “augmented translation.” In this view, language professionals don't correct MT, but instead, use it as a resource alongside TM and terminology. This means that buyers will increasingly just look for translation, rather than distinguishing between machine and human translation. They will just buy “translation” and the expectation will be that MT will be used if it makes sense. The MT component of this approach is already visible in tools from Lilt, SDL, and others, but we're still in the early days of this change."

In addition, this will probably mean that we can do away with the “post-editing” misnomer – editing is editing, regardless of whether the suggestion presented in the CAT tool interface comes from a TM or an MT engine. Therefore, the term “post-editing” should be reserved only for the very specific case in scenario A, otherwise, the concept will be meaningless. This view is taken in, for instance, the contributions by a post-editor educator and an experienced post-editor in the recently published book Machine Translation – What Language Professionals Need to Know (edited by Jörg Porsiel and published by BDÜ Fachverlag).

Thus it seems that eventually we will be left with mainly scenarios B and C – which leaves the matter, for translators, of how to come to grips with B. This is a new situation which is likely to take time and discussions to arrive at a solution (or solutions) palatable to everyone involved. Meanwhile, we translators should aim to make the best possible use of scenario C. MT is here and will not go away even if some people would wish it to.

-------------

Mats Dannewitz Linder has been a freelance translator, writer and editor for the last 40 years alongside other occupations, IT standardization among others. He has degrees in computer science and languages and is currently studying national economics and political science. He is the author of the acclaimed Trados Studio Manual and for the last few years has been studying machine translation from the translator’s point of view, an endeavour which has resulted in several articles for the Swedish Association of Translators as well as an overview of Trados Studio apps/plugins for machine translation. He is self-employed at Nattskift Konsult.