Last month AMPLEXOR was honored to join DFKI to celebrate the center’s 30th anniversary in Berlin. During the event we had the opportunity to present the neural automatic translation systems (nMT) for the life sciences sector, developed in cooperation between AMPLEXOR and DFKI.
Subscribe to our blog
In Douglas Adams’ famous science fiction work “Hitchhiker’s Guide to the Galaxy,” there is a remarkable creature called the babel fish. The babel fish is small enough to be placed inside one’s ear, and once there, can instantly translate any language one hears by telepathically decoding the brain-wave matrix for speech. How absurd does that sound? Although placing a babel fish in one’s ear may be a bit of a stretch, enabling instant, exact translations for any language could become a reality sooner than you think.
We have already seen visions from many of our favorite science fiction series come true over the years, including instant messaging, interactive touchscreens, smart glasses and more. In fact, sometimes we hardly notice as these “far out” concepts become a reality. Often, these breakthroughs happen thanks to technology commonly referred to as Artificial Intelligence (AI).
For the past 30 years, the German Research Center for Artificial Intelligence (DFKI), the world's largest institute of its kind, has been dedicated to the development of AI technologies for humans. On October 17, researchers, politicians, journalists and sponsors of the DFKI gathered to celebrate the center’s 30th anniversary in Berlin. During the event’s open house, 16 of the DFKI’s current AI projects were presented, including the neural automatic translation systems (nMT) for the life sciences sector developed in cooperation between AMPLEXOR and DFKI.
For me, it was great opportunity to marvel at 15 completely different AI application areas, such as learning slacklining (yes, really!) or the medical supervision of people in need of care by means of technology woven into their clothing. Together with the DFKI’s experienced computer linguist Dr. Raphael Rubino, we also had the chance to answer guests’ questions about language technology when they stopped by our booth. The rush of visitors put our vocal chords to the test! Last but not least, the event was an ideal setting to discuss AI technology with the researchers in attendance.
The impressive advances in neural translation technology, especially over the past two to three years, have generated a lot of hype among researches and end users of AI. So much seems suddenly feasible, and all boundaries are shifting as human communication becomes more accurately deciphered in by machines. Are we about to put the ultimate babel fish in our ears? And how did this explosive surge in innovation come about in the first place? This event was the perfect opportunity to ask all those questions.
"We owe part of our success to the gaming industry," reveals Professor Stephan Busemann, Deputy Head of the Language Technology Division at DFKI. In addition to the quantities of speech data available today, it is above all the powerful graphics processors used by gaming companies that have provided the researchers with the basis for neuronal speech data processing. Before such technology, many AI theories for translation existed but were never realized.
In light of such rapid technological evolution, wouldn’t it be safe to assume the remaining problems with machine translation will be solved in the near future? How far away are we from a machine that will fully, autonomously and perfectly translate for people who communicate in different languages and thus revolutionize global communication?
Professor Busemann is used to hearing this question. “Just because the various recent successes in very different fields of application have impressed the public does not mean that the topic of artificial intelligence has now been solved.” He is convinced AI applications can and should support people in their work, yet not replace them.
"Machine translation was already a research topic during the Cold War," Dr. Rubino agrees. "Human language is very complicated and it is constantly changing. Translation is a non-deterministic process. One and the same text in one language can be translated correctly in many different ways. There are almost always several possible ways of saying something. That makes technical solutions very difficult."
According to Dr. Rubino, the development of machine translation systems to support professional translators poses further challenges. "Translators, due to their expertise and language skills, have very different system requirements from laypeople. They want to increase their productivity and cannot afford any mistranslations. It is not enough to convey the meaning approximatively." Consequently, the development of domain-specific nMT systems with AMPLEXOR focuses on technical precision. While general language-trained systems, such as Google Translate, DeepL, Microsoft Translator or Amazon Translate deliver good results on a wide range of texts, they often fail for specialized texts, where phrases and words often have completely different meanings than in everyday language.
Additionally, there are numerous specifications in the medical field regarding the correct naming of facts or objects. Words, such as rare, occasional or frequent, are not difficult to translate and are probably considered by very few to be technical terminology; however, on the package insert of a medication that contains descriptions of potential side effects, such terms are clearly defined, purposefully used and must not be replaced by similar adjectives.
The systems developed by AMPLEXOR and DFKI for the life sciences sector are fundamentally geared to this field. For corresponding medical texts, they perform better than general language systems, even if those systems were fed with the same training data.
"Training data in sufficient quantity and in high quality are essential for building high-quality, specialized systems, such as AMPLEXOR's," confirms Dr. Rubino. Nevertheless, machine translation remains a statistical guessing game based on probabilities it has learned from training data. Because it doesn’t translate meaning as a human would, machine translated content always contains a percentage of errors. However, it is amazing how accurate a well-trained machined translation engine can be.
As in the past, professional translators will continue to act as mediators between cultures and must adapt to new ways of working. For years, instead of translating texts from scratch, they have been increasingly post-editing machine inputs and correcting errors. That said, professional translators’ core competence – culturally contextualizing the meaning of content in another language – will continue to be critical.
Despite all the recent technological progress, there is still a lot of work to be done before machines can achieve the same results as professional translators. But, thanks to researchers at institutions, such as the DFKI, harnessing the power of AI to generate flawless machine translations may materialize. Perhaps a real babel fish will exist in the not-so-distant future . . .
About the author
Karina is Solution Manager Global Content Suite at AMPLEXOR. She is based in Berlin, Germany.