Neural Machine Translation: New Approach to Communication and Translation
For decades, machine translation has been discussed to replace humans, but the goal has not yet been achieved. Neural Machine Translation (NMT) is a newer development in machine translation, and there are high expectations for it. The School of Languages and Translation Studies at the University of Turku organises courses on translation technology and machine translation with emphasis on practical approach to modern technology. Neural machine translation is marketed in the media as the solution to the problems between languages. Is that also the case beneath the surface?
– Neural machine translators function by learning to translate from already translated material between a specific language pair. It is an A.I. which learns to translate between two languages from the source material it has been given, by using algorithms. Humans cannot fully control the machine’s learning process, which creates different challenges for the machine’s development. It is possible that the machine learns to translate something wrong based on the source material, in which case even the machine’s developer is not able to effectively correct the mistakes, says teacher and researcher of translation technology, University Lecturer Maarit Koponen.
Recently, however, neural machine translation has been introduced to the public as a tool for solving practically every problem between different languages. Brands such as Facebook, Google and Microsoft use NMT and also market it strongly. Machine translation has improved greatly over the last few decades, but it still performs best in a regulated environment.
Better Translations with Room for Improvement
– For example, machines have been able to translate weather forecasts for decades, since the vocabulary and topics stay the same over the years. Neural machine translation has not advanced as much as it may seem based on advertisements, and this machine also performs best in a regulated environment, says Koponen.
In regard to technology, NMT could translate texts from multiple fields, but its learning material determines which texts are suitable for a single machine. The most reliable translations are produced when the learning material corresponds with the texts that need to be translated.
Reliability is an important factor in machine translation, especially with NMT which produces texts that resemble, sometimes almost perfectly, texts written by humans. According to Koponen, this is both the strength and the weakness of NMT. Even if the text produced by the machine is flawless, the subject matter might include critical mistakes. The role of a human translator is important in reviewing and post-editing the machine-translated texts, meaning that NMT-translated texts should be approached with a critical mind and examined beneath the surface.
– A neural machine translator needs great amounts of textual material from language pairs in order to learn to translate between them, and insufficient amount of material is one of the challenges in the machine’s development. At the moment, most of the translated texts are in English, so the other language of the pair needs to be English in order for the machine to have enough material for competent translations, adds Koponen.
Neural machine translation holds great potential, but it also has noticeable flaws. What could therefore be the future of NMT? The machine will not likely be able to translate entire books from one language to another, instead its application is more interactive. Koponen suggests that the machine would most likely function best as a tool to help human translators, or in information retrieval, as long as the user is able to evaluate the translations’ reliability. NMT could also be used in situations where the context is explicit and controlled, such as with communication problems in health care.
According to Koponen, the University’s machine translation studies aim for the future, as translator students will most likely need and use the technology in their working life. The course provides students with basic knowledge about machine translation, and they will also be able to evaluate the usefulness and applicability of machine translation in their work, as well as to communicate these things to their customers.
– It is difficult to predict the future, as machine translation has developed alongside other technologies, sometimes rapidly, and at other times slowly. There will be enough work for graduating translators in the future, but technology is going to play a larger role in their work. At the moment, it seems that neural machine translation will be strongly represented in future technology, concludes Koponen.
Jenni Maja