Wednesday, April 26, 2017

Optimizing LSP Performance in the Artificial Intelligence Landscape

Artificial Intelligence and Machine Learning have been all over the news of late, and we see that all the internet giants are making huge investments in acquiring AI expertise and/or using "machine intelligence" which is the term being used when describing how these two areas come together in business applications. It is said that AI is attracting more venture capital investment than any other single area in VC at the moment, and there are now people regularly claiming that the AI guided machines will dominate or at least deeply influence and transform much of our lives in future perhaps dangerously so. This overview of AI and this detailed overview of  Neuralink is quite entertaining and gives one a sense of the velocity of learning, and knowledge acquisition and transmission that we are currently facing and IMO are worth skimming through at least.

However. machines learn from data and find ways to leverage patterns in data in innumerable ways. The value of this pattern learning can only be as good as the data used and however exciting all this technology seems we need to understand that our comprehension of how the brain (much less the mind) works, is still really is only in its infancy. “Machine learning” is a fancy way of saying “finding patterns in data”. The comparative learning capacity and neuroplasticity of any 2-year-old child will put pretty much put all of these "amazing and astounding" new AI technologies to shame. Computers can process huge amounts of data in seconds, and sometimes they can do this in VERY useful ways, and most AI is rarely if ever much more than this. If the data is not good, the patterns will be suspect and often wrong. And given what we know about how easy it is to screw data up this will continue to be a challenge. This will not change no matter how good the algorithms and technology are. The importance of data quality has been and will continue to be a major challenge for AI that learns from biased language data and several have pointed out how training on racially and gender biased data will propagate through machine learning.

The business translation industry is only just beginning to get into an understanding of the value of  "big data". Very few actually collect data at a level that they can really leverage these new options. I have seen that the industry has terrible and archaic practices with translation data organization, management and collection in my years working with SMT where LSPs struggle to find data to build engines. A huge amount of time is spent on gathering data in these cases, and little is known about its quality and value. However, as this post points out there is a huge opportunity in gathering all kinds of translation process related data, and leveraging it by properly cataloging and organizing it, so that most man-machine processes improve and continually raise productivity and quality.

This is a guest post by Danny de Wit about a cloud-based technology he has developed called Tolq. This is an AI infused translation management system with multiple translation processes that can be continually enhanced with AI-based techniques as process related data is acquired. The system is a combination of TMS, project management and CAT tool functionality that can continually be updated and enhanced to improve workflow ease and efficiency for both traditional and MT-related workflows. Thank you to Mila Golovine for introducing me to Danny.


Boom. There it was in the news: Google NMT makes all LSP's redundant with their Neural Machine Translation efforts! And if not now, then surely within a few years. Neural machine translation solves it when it comes to translation. Right?i

Translation industry insiders already know this is not the case, nor is it going to be the case soon. Translation is about understanding, and not just about figuring out patterns, which neural networks do so well. So there are still huge limitations to overcome: homonyms, misspelled source words, consistency, style and tone-of-voice, localisation, jargon/terminology, incorrect grammar in the source text and much more.

For now, artificial intelligence is a tool. Just like translation memory and cat tools are. But it's important to understand that this is a very powerful one and one of a different kind. This is a tool, which when applied correctly, gains strength every day and will open up more and more new opportunities.

Within a period of just a few years, the entire translation industry will be reshaped. Translation is not the only industry where the impact will be felt. The same will go for virtually any industry. The technology is that powerful. But it won’t be in the shape of zero-shot NMT.

Artificial intelligence is a wave of innovation that you have to jump on. And do so today.

Fortunately, it turns out that one key thing to help you do this is one that LSP's have been doing already for years: collecting high-quality data.

The Key is Data & Algorithms

All A.I. advancements are built upon models that are generated from data-sets. Without data, there is no model. Data is both the enabler, but also the limit. What's not in the data, is not available to be used.

LSP's have been collecting data for years and years. Using that data to power A.I. algorithms today and in the future is a key strategy to implement.

In addition, LSP's have client specific data and add to that data on a daily basis. This means there's the opportunity to have client specific A.I. tools that gain in strength over time. Offering this to your clients is a big differentiator.

Having generic NMT plus different layers of client specific NMT and other client specific A.I. tools can provide you with workflows that were previously unachievable.

But how do you integrate this inside your operations? We all know that clients and therefore LSP's have a hard time managing data. The incentive to centralise was always there, but it has never been as clear or impactful as it is right now.


Unification of Workflow, Data Storage and Algorithms: the new holy trinity for LSP's

The interests of LSP's and clients are aligned where it comes to the unification of these elements.

Clients can take advantage of huge cost savings and new services (like generated multi-lingual content a.o.).

LSP's will operate more efficiently and gain huge workflow management advantages.

LSP's increase on their position as indispensable partners for clients, because of the complexity of the technology involved. Clients will not be able to implement anything like this themselves. In part due to the data + algorithms requirement to make this all work.

Simple workflow improvements, that companies like recently Lilt offer are a step forward, but due to their architecture can only add some efficiency gains inside an existing process, but not lift the LSP organisation to a new level taking advantage of all A.I. has to offer.

Instead, architecture that brings together the three key elements to take full advantage of the future is a much better alternative. Tolq calls this the "A.I. Backbone" which is present in all our operations. Unifying all workflows and data storage to a central, but still layered, structure. Then add to that different A.I. algorithms to optimise and expand the translation services process.

Architecture that makes your company stronger each day, with each translated word.


New Opportunities: Algorithms Galore!

What can other advances should we expect from A.I. in the near term?

LSP's can look forward to more advances for workflow optimisation and the possibility to introduce new services.

The generic engines will be one of the tools to take advantage of, but to get to the final product that clients require, LSP's that centralise their operations and data will be able to take advantage of new algorithms to offer clients new services. Some examples are shown in the diagram below. But many more will be made available as the A.I. wave increases in speed.


 Value Creation Strategies for LSP's

When it comes to creating value the core strategy for LSP's should be to start combining technology with data. Even the technology giants will be envious of that combination. We expect to see huge acquisitions in that space.

In addition, combining data with algorithms can provide powerful scalable profit centers due to their nature. It's data crunching vs. human workflows. 

Danny de Wit - CEO & Founder, Tolq

Danny de Wit founded Tolq in 2010. Danny’s drive to radically innovate the traditional world of translation was fuelled by the lack of satisfying solutions for website translation.

Danny de Wit studied Business Administration at the Erasmus University Rotterdam (Netherlands). After a career in sales, consultancy and interim management Danny’s primary focus and passion have been startups and technology.

Prior to founding Tolq Danny was involved as an entrepreneur in starting up several online businesses. Danny founded in 2000 en in 2009.

Danny’s innovative ideas, for virtualizing business organization using technology have been acknowledged by two European patents, two Dutch patents and one American patent pending. The intellectual property covers specific methods of distributing work over large populations of resources, whilst optimizing quality and efficiency of the work in real time. This technology is applied in

In 2012, Tolq was a Finalist at the Launch! Conference in San Francisco and was recognised by Forbes as a startup with very promising technology describing the platform as a faster, easier and cheaper way to translate any website.

In 2014 Tolq was selected as one of the top 15 companies at The Next Web Conference and mentioned by Forbes again as one of the leading startups in Holland. provides LSP's with technology architecture that puts A.I. at the core of all your LSP operations.

Thursday, April 20, 2017

LSP Perspective: MT Post-Editing Means a Drastic Reduction in Translation Cost

This is a short guest post by @translationguy also known as Ken Clark.  

These initial preamble comments in italics are mine. 

Today, many LSPs and Enterprises are working with MT and there is enough evidence that MT works even when you don't really know what you are doing. Unfortunately, many agencies still try to do it themselves with Moses and most of these DIY experiments either completely fail or produce systems that are not as good as the public systems produced by Microsoft and Google, which defeats the whole point of doing it. MT as a technology only provides business leverage if you have a superior MT system and have aligned processes to take advantage of this. 

Ken differentiates between light and full post-editing in his view of post-editing, and I would like to add another dimension to this discussion. It is my experience that full post-editing is done with smaller (in MT terms) projects, or when the information translated is very critical to get right. Thus, in a knowledge base project context, content related to security, privacy, and legal terms may be sent for full post-editing, and other content may just get a lighter post-edit. Also, when one is involved with very large MT projects like the team at eBay is, where hundreds of millions of words are involved, it is not possible to do a full post-edit on all the data so a light post-edit is done, or maybe nothing beyond the very specific linguistic work on high-frequency n-grams and important patterns that Silvio Picinini describes in this post. Unfortunately, it’s hard for translators and clients to agree on when we’re done with “light” post-editing, so it’s a headache to manage as editors often cannot tell when to stop.

Thus, as agencies really get involved with "real MT " projects they will do corpus profiling work and focus their attention on critical patterns as Juan Rowda has described in this post.

To me, real competence with MT in an agency or enterprise is demonstrated when there is some expertise with as many of the following core functions as possible:

  • Understanding the Data - Corpus Analysis
  • Focusing Linguistic Work on High-Frequency Patterns
  • Working with Expert MT Systems Developers in a pro-active way
  • Understanding MT Output Quality
  • Driving MT quality higher with specific linguistic feedback
  •  Managing Post-Editing Processes and Compensation
TAUS provides an excellent overview of the larger perspective in this post on best practices in MT

As the MT technology evolves I think we will see that strategies that made great sense with phrase-based SMT may not always make sense with the new Neural MT technology. I am talking to SYSTRAN about the realities in the NMT paradigm and hope to produce a post on this soon.



Machine translation has improved by leaps and bounds. What was once considered machine-produced gibberish is increasingly giving human translators a run for their money, particularly for predictable texts like weather reports.

While machine translation (MT) is also more economical than human translation, it’s not a true alternative yet. In most cases, machine translation can’t be used as is. And that’s where the expertise of machine translation post-editors comes in. Machine translation post-editors are the human editors that work to improve the output of machine translation. They combine the MT output with their linguistic expertise to provide a better reading experience to human audiences.

Besides the cost savings, it is estimated that machine translation plus post-editing is 40% more efficient than human translation alone. But what exactly do machine translation post-editors do, and how do they do it?

Types of Machine Translation Post-editing

Machine translation post-editing comes in two flavors: light post-editing and full post-editing.
Light post-editing suggests a lighter touch, only asking the human editors to ensure that the MT output is accurate in meaning and understandable to the reading audience. However, this means that style is not taken into account, grammar and syntax may be awkward, and the text may sound as if it were produced by a computer. It’s the most economical option, but for reasons of quality, light post-editing is typically only used when a translation is needed urgently and/or for an organization’s internal purposes.

Full post-editing, on the other hand, calls for a higher level of involvement by the post-editor. (This makes it more expensive than light post-editing, but still less expensive than full human translation.) In addition to making sure that the MT output is accurate in meaning and understandable to the reading audience, full post-editing addresses the text’s grammar, syntax, and punctuation, ensuring they are correct and appropriate. The result is similar in quality to a human translation, although it may not yet match the style of a native-speaking translator. Full post-editing is typically used when a machine-translated text is intended to be published, or widely disseminated inside or outside an organization.

MT Post-editing Strategies

How do they do it? Let’s examine some of the things that post-editors watch out for.

Light post-editors use the machine translation output as much as possible. However, they take special care that information has not been inadvertently added in or left out. They also edit anything they have identified as offensive or culturally unacceptable.

In addition to the above, full post-editors correct any grammatical and syntactical errors. They pay particular attention to terminology, making sure that the terms have been translated in the appropriate way (or left untranslated per the client’s wishes). They also ensure that the spelling and punctuation, as well as formatting, are correct.


Tuesday, April 11, 2017

The Problem with BLEU and Neural Machine Translation

There has been a great deal of public attention and publicity given to the subject of Neural Machine Translation in 2016. While experimentation with Neural Machine Translation (NMT) has been going on for the last several years, 2016 has proven to be the year that NMT broke through and became a big deal, and became more widely understood to be of great merit outside of the academic and research community, where it was already understood that NMT has great promise for some years now.

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.

The basis for the NMT optimism is related both to the very slow progress in recent years with improving phrase-based SMT quality, and also the striking BLEU score improvements that were seen coming from neural net based machine learning approaches. Much has been written about the flaws of BLEU but it still remains the most easily implementable measurement metric, and also really the only one where there are long-term longitudinal data available. While we all love to bash on BLEU, there is clear evidence that there is a strong correlation between BLEU scores and human judgments of the same MT output. The research community and the translation industry have not been able to come up with a better metric that can be widely implemented to enable ongoing test and evaluation of MT output so it remains as the primary metric.The alternatives are too cumbersome, expensive or impractical to use as widely and as frequently as BLEU is used.

However, there is also evidence that BLEU tends to score SMT systems more favorably than RBMT and NMT systems, both of which may produce very accurate and fluent translations to a human perspective, but differ greatly from the reference translations that are used in calculating the BLEU score. To a great extent the BLEU score is based on very simplistic "text string matches". Very roughly, the larger the cluster of words that you can match exactly, the higher the BLEU score.

To illustrate this, lets take a very simple example, say a reference translation is: "The guests walked into the living room and seated themselves on the couch." and an NMT system produces something like: "The visitors entered the lounge and sat down on the sofa." This would result in a very low BLEU score for the NMT segment, even though many human evaluators might say it is quite an acceptable and accurate translation, and as valid as the reference sentence.

If you want a quick refresher on BLEU you can check this out:

The Need for Automated Translation Quality Measurement in SMT: BLEU

Some of the optimism around NMT is related to its ability to produce a large number of sentences that look very natural, fluent and astonishingly human. Thus, much of the early results with NMT output show that it is considered to be clearly better to human evaluators, even though BLEU scores may show only 5% to 15% improvement (which is also significant). The improvements are most noticeable when considering fluency and word order issues with machine translation output. NMT is also working much more effectively in what were considered difficult languages for SMT and Rule Based MT, e.g. Japanese and Korean.

And here are some examples provided by SYSTRAN from their investigations where the NMT seems to make linguistically informed decisions and changes the sentence structure away from the source to produce a better translation. But again these would not necessarily score much better in terms of BLEU scores even though humans might rate them as significant improvements in MT output quality and naturalness.

But we have seen that in spite of this there are still many cases where NMT BLEU scores significantly outpace the phrase-based SMT systems. These are described in the following posts in this blog:

A Deep Dive into SYSTRANs Neural Machine Translation (NMT) Technology


An Examination of the Strengths and Weaknesses of Neural Machine Translation


Real and Honest Quality Evaluation Data on Neural Machine Translation 


and this is even true to some extent in the exaggerated over-the-top claims made by Google when they claimed that Google NMT was “Nearly Indistinguishable From Human Translation” and “GNMT reduces translation errors by more than 55%-85% on several major language pairs" which is described below.

The Google Neural Machine Translation Marketing Deception


The KantanMT NMT vs PB-SMT Evaluation Results

I had an interesting conversation with Tony O'Dowd at KantanMT about his experience with his own initial NMT experiments.While Kantan does plan to publish their results in full detail in the near future, here are some highlights Tony provided from their experiments, that certainly raises some fundamental questions. (Emphasis below is mine.)

  1. Scope of Test - We built identical systems for SMT and NMT in the following language combinations - en-es, en-de, en-zh-cn, en-ja, en-it. Identical training data sets and test reference materials were used throughout the development phase of these engines. This ensured that our subsequent testing would be of identical engines, only differing in the approach to build the models. The engines were trained with an average of 5 million parallel segments ranging from 44 - 110 million words of training data.
  2. BLEU Scores - In all cases, the BLEU scores of NMT output was lower than SMT. 
  3. Human Evaluation:  We deployed a minimum of 3 evaluators for each language group and used KantanLQR to run the evaluation. We used the A/B Testing feature of KantanLQR. Sample A was from SMT, Sample B was from NMT. We randomized the presentation of the translations to ensure evaluators did not know what was NMT and SMT - this was done to remove any bias for one approach or the other. We sampled 200 translations for each language set.
  4. In all cases NMT scored higher in our A/B Testing than SMT. On average NMT was chosen twice as often as SMT in our controlled A/B testing.
  5. For low scoring BLEU NMT segments, we found a high correlation to these segments being the preferred translation by our [human] evaluators - this pretty much proves that BLEU is not a useful and meaningful score for use with NMT systems.

Clearly, this shows that BLEU is of limited value when the human vs. automated metric results are so completely different and even diametrically opposed. The whole point of BLEU is that should provide a quick and simple way to get an estimate of what a human might think of sample machine translated output. So going forward it looks like we are going to need better metrics that can map more closely to human assessments. BLEU is not a linguistically informed measure and thus the problem. This is easy to say but not so easy to do.  A recent study pointed out the following key findings:

  • Translations produced by NMT are considerably different than those produced by phrase-based systems. In addition, there is higher inter-system variability in NMT, i.e. outputs by pairs of NMT systems are more different between them than outputs by pairs of phrase-based systems.
  • NMT outputs are more fluent. We corroborate the results of the manual evaluation of fluency at WMT16, which was conducted only for language directions into English, and we show evidence that this finding is true also for directions out of English.
  • NMT systems do more reordering than pure phrase-based ones but less than hierarchical systems. However, NMT re-orderings are better than those of both types of phrase-based systems.
  • NMT performs better in terms of inflection and reordering. We confirm that the findings of Bentivogli et al. (2016) apply to a wide range of language directions. Differences regarding lexical errors are negligible. A summary of these findings can be seen in the next figure, which shows the reduction of error percentages by NMT over PBMT. The percentages shown are the averages over the 9 language directions covered.

 Reduction of errors by NMT averaged over the 9 language directions covered

Given that there are currently no real practical alternatives to BLEU, there is perhaps an opportunity for an organization like TAUS to develop an easy to apply variant from their overall DQF framework, that can focus on these key elemental differences and can be done quickly and easily. NMT systems will gain in popularity and better measures will be sought. The need for an automated metric will also not go away as developers need some kind of measure to guide system tuning while they are in the development phase. Perhaps there is some research underway that I am not aware of that might address this, but I have seen that SYSTRAN uses several alternatives but everybody still comes back to BLEU.

Comparative BLEU score-based MT system evaluations are particularly problematic as I pointed out in my critique of the Lilt Labs evaluation, which I maintain is deeply flawed, and will result in erroneous conclusions if you take the reported results at face value. Common Sense Advisory also wrote recently about how BLEU scores can be manipulated to make outlandish claims by those with vested interests and also point out that BLEU scores naturally improve as you add multiple references.

"However, CSA Research and leading MT experts have pointed out for over a decade that these metrics are artificial and irrelevant for production environments. One of the biggest reasons is that the scores are relative to particular references. Changes that improve performance against one human translation might degrade it with respect to another. "
Common Sense Advisory, April, 2017

There is really a need for two kinds of measures, one for general developer research that can be used everyday like BLEU today, and one for business translation production use which indicate quality from that different perspective. So as we head into the next phase of MT, driven by machine learning and neural networks, it would be good for us all to think of ways to better measure what we are doing.  Hopefully some readers or some in the research community might have some ideas on new approaches to do this but this is an issue that is something worth keeping an eye on. And if you come up with better a way to do this, who knows, they might even name it after you. I noticed that Renato Beninatto has been talking about NMT recently, and who knows he could come up with something, I know we would all love to talk about our Renato scores instead of those old BLEU scores!