Jay Marciano is Lengoo’s Director of MT Outreach & Strategy and has been recently re-elected to the board of directors of AMTA (Association for Machine Translation in the Americas). Marciano has over 20 years of experience in the development and application of MT. He talks to us about "augmented translation" at Lengoo and about how advancing AI technology is changing the translation process.

Neural Machine Translation (NMT) - an application of machine learning - delivers natural-sounding translations. Will machines replace translators?

Jay Marciano: NMT is very real and it will have a bearing on how translators work. However, the goal is not to get rid of them. Digitalization is changing everybody’s job and life. The fact that we’re conducting this interview by video call right now, for example, would’ve been unimaginable 20 years ago. Without a doubt, the day-to-day work of a translator is changing due to the influence of technology.

What is so radical about the change that comes with NMT?

Marciano: Every major technological advancement in the way information is stored, transported, and retrieved has had an impact on how translators work: writing on paper instead of etching in stone, using printing and fax machines, storing translations electronically, using Translation Memories. However, we’re now grappling with the fact that computers are increasingly able to do things that were thought to require human intelligence. In this case, the translation itself. That’s what is so radical about the change that comes with artificial intelligence and machine learning. Trained properly, NMT produces pretty damn good translations.

How is NMT changing the way translators work?

Marciano: Let’s distinguish between the work that translators currently do and the work that someone with the same fundamental skill set could do. Computers are already playing a bigger and bigger role in the translation process. You could compare the effect of NMT on translators with that of calculators on the work of engineers. People wondered what engineers would actually spend their time on if calculators did all the math for them. Well, they are still around and have the time to do more interesting things. As computers take on more and more of the basic work of translation, a person with truly refined language skills and subject area expertise could for example move into more technical work, along the lines of data science and data curation that makes these NMT systems possible. “Language engineer” could be a good job title for that kind of work.

At Lengoo AI and translators are working together. Could you call the duo “augmented intelligence”? The idea is that AI-technology frees humans from boring and repetitive tasks to make them more productive and in the end to make them happier.

Marciano: In a way, yes. What you usually mean by this is that a translator makes a change to the machine translation output, the machine keeps learning from the human, and in return, the machine translation becomes better and better. Translators can focus on highly nuanced parts of language. Ensuring that they are correctly rendered in the target language is a highly cognitive task. NMT gives them the opportunity to actually focus on what they are truly interested in: the meaning of language. NMT often provides surprisingly fluent translations, but we cannot be confident yet that those translations are actually accurate. There still needs to be an expert revision. It’s very necessary to have translators who understand all the semantic and stylistic nuances of source content and ensure that it is present in the translation. That remains and will remain a critical skill.

Apart from “augmented intelligence” the expression “augmented translation” seems to have become popular in the field. What is meant by that?

CSA Research coined the term “augmented translation” in 2017 to describe their vision of the translation process of the future. It’s based, in turn, on the concept of “augmented reality”, in which the human ability to perceive and interact with the world is greatly enhanced by a tight integration of computer systems. That might sound scary, but imagine being at your local farmers market and having your augmented reality goggles identify a mysterious vegetable, list its nutritional information, and provide a few recipes that you could make with stuff that’s already in your refrigerator at home.

Augmented translation: "Human is at the centerpiece with access to technologies."

In translation there are many interesting and useful technologies that are currently suboptimally integrated into the translator’s work. The point of augmented translation is that the human being is the centerpiece and has access to technologies like automated content enrichment (ACE), terminology management, translation memories, and fully automated project management, all of which, of course, need to interoperate smoothly with each other. The translator remains in control and uses these technologies to deliver the best possible value and quality.

Is that what Lengoo provides?

No one has yet fulfilled the vision of augmented translation, but Lengoo is well on its way. We already have many of the pieces in place to achieve this new paradigm. We have a matching algorithm to find the best translator with subject-area expertise for the job, and a translation platform that allows us to put all kinds of helpful tools right at the translator’s fingertips. And, of course, we continue to develop our custom NMT systems to provide draft translations that require fewer and fewer changes by the translator. At Lengoo, we also like to think about “augmented translation” from the perspective of human intelligence augmenting artificial intelligence. For example, translators use CAT-tools (editor’s note: Computer Assisted Translation), but it is just as important to provide what I would call a “TAC-tool” as in “Translator Assisted Computer”. Optimally, the augmentation goes both ways.

Why is “augmented translation” the future?

Marciano: There are about 300,000 translators on planet earth, a number that has remained fairly steady over the last several years. In contrast, the amount of content that we are creating is skyrocketing. Every year we produce 40 percent more data than we did the year before. So the amount of translatable content is increasing much faster than the volume that professional translators can handle. Just to stay even with the percentage of content that we are translating each year, we have to increase our translation capacity radically every year. Without technology, this would be impossible. But equally important, if we want to address this huge market opportunity we need to change the ROI model for translation.

What’s wrong with the ROI model?

It has to become economically feasible to translate much more content, and we need innovative ways of measuring the economic value of translation. Today a massive amount of content doesn’t get traditionally translated because we consider the process too expensive or time-consuming. If we can radically change the costing model of translation by optimizing the interplay of technology and language professionals, then the demand for translation will increase massively.

Which insights of last year will have an impact in the field of machine translation?

Marciano: The next big thing is “Quality Estimation” (QE). This will be happening on a wide scale within the next two to three years. NMT-systems produce a translation and then another neural network will initially assess the quality. One neural network is, practically speaking, supervising the other, with the subject-area language expert reserved for working on those translations that the NMT and QE system are unable to handle.

Why will this have such a great impact on the translation process?

Marciano: If the Quality Estimation system could predict with high confidence that a given machine translated sentence would not be changed by a proficient human reviewer, the computer will be able to very confidently say: Hey, translator, don’t worry about this one; it can go straight into the final quality review at the very end of the process. This might not sound like terribly much, but it will allow translators to work much faster by freeing them up to concentrate on the sentences that will benefit most from their expertise.