Since the time of the Tower of Babel, humans have been awaiting the frustrated dream of eliminating all language barriers from the planet. This is a wish that science fiction has also made its own. From the universal translator of the Star Trek saga, to Babelfish, the little yellow fish that slips into the ear for simultaneous translation, imagined by Douglas Adams in the 'Hitchhiker's Guide to the Galaxy'.
In this field, as in many others, reality quickly catches up with imagination. There are dozens of automatic translation programmes, but the best known and most widely used is undoubtedly Google Translate, which has been in operation for more than ten years and has 500 million users, 103 languages available and 100 billion words translated per day.
Translating, however, is like making love. It has a mechanical and a poetic side. Machine translator algorithms have so far been able to reproduce the mechanical and technical side, but they are rather weak when it comes to returning the emotional side. For this reason, Google Translate's work for many years has been synonymous with approximate, often useless, translations.
As Sabrina, a translator with twenty years of experience, recalls, “In the early 2000s, machine translation was more esoteric, sometimes poetic, than a working tool. I have never considered it as a competitor, and I have only been considering it as a possible aid for a couple of years now.”
In fact, at the end of 2016, Google's online translation service switched from the traditional phrase-based system, with which the software changed from looking for the best match with dictionary terms, to a new method called Google Machine Neural Translation (GMNT), based on artificial intelligence, which was capable of reducing the errors made with the previous algorithm by up to 80 percent.
The artificial neural network assumes the sentence, or paragraph, as the translation unit, thus considering the context around the words. In addition to statistical comparison between databases of translations already made, it also exploits a self-learning mechanism that allows it to deduce the rules independently, instead of receiving them from the programmers. The result is a text that is more fluent and closer to natural human speech.
Self-learning allows the neural network to develop its own interlanguage, thanks to which it can translate acceptably even between two formally unconnected languages. This means that the neural network knows how to form a conceptual-semantic representation that goes beyond particular languages and, therefore, can establish equivalences between words and phrases of different languages. In other words, it is an embryonic form of cognitive processing.
For language combinations where neural translation is already active, the quality of the service improves significantly. Google claims that today Translate scores 5.43 / 6 (where 6 is perfection) from English to Spanish, while the average for human translators is 5.50. From English to Chinese the average is 4.30, against 4.60 for humans. “The French-English or English-French combination gives amazing results to date, but with any other combination that does not include English, the result is very poor,” Sabrina warns.
In addition to being a great opportunity, automation also represents a challenge, if not a direct threat for many professional categories. According to a U.S. study, 41% of respondents fear losing their jobs in the coming years. Among the potential victims of technological disruption are translators and interpreters, who have been receiving machine-translated texts for years only to be proofread and edited, with some agencies already specializing in post-editing.
“The demand for translation has been growing steadily since the 1990s and there is every reason to believe that the trend will continue”, says David, the manager of a translation agency. “What we have noticed is a growing demand for language management solutions, where freelance translation sometimes has its place”, he notes.
On the other hand, while the interpreter is not yet completely dispensed with, their physical presence is becoming less necessary. This is the result of the advent of software such as Tywi, which reduces personnel costs by enabling remote simultaneous translation on videoconferencing platforms such as Skype, WebEx or Adobe Connect.
Google AI engineers have presented the results of a new system for translating directly from speech to speech, without using any intermediate text. The programme, called Translatotron, can even maintain the voice of the original speaker in the target language. The experiment is still in progress and the results are far from perfect, however, incredible potential can be glimpsed.
Introducing Translatotron: An End-to-End Speech-to-Speech Translation Model https://t.co/DBpuqYYwyz
Technological development almost always stems from military research, and translation is no exception. In 2011, Darpa, the Pentagon's technology agency, commissioned IBM to develop a project - Bolt (Broad Operational Language Translation) - that combines speech recognition, natural language understanding and machine translation technologies to enable the military to have conversations with people in the countries where they operate.
IBM committed to delivering it in 2022, hoping to achieve better results than the TransTac system, developed between 2005 and 2010. Tested in Iraq and Afghanistan, it was even counterproductive in terms of communication. As Adams writes, after all, translating does not always mean understanding: “the poor Babel fish, by effectively removing all barriers to communication between different races and cultures, has caused more and bloodier wars than anything else in the history of creation.”
Despite these advances, the goal of replacing human translators with neural robots still seems far off. In texts where the expressive and aesthetic function prevails, i.e. when an author's personal style is essential, or where the appeal to the addressee predominates (advertising, speeches) or in any text linked to a subculture, it is difficult to imagine a form of machine translation capable of completely replacing the human being.
The human factor is essential because it makes it possible to detect a whole series of characteristics that must be taken into account in order to produce an effective and accurate translation. This invisible and quantitatively undetectable subtext refers to a particular way of thinking and conceptualizing reality. “The professional translator brings the creativity, knowledge and sensitivity of the native speaker. A translator thinks, just like any human being”, Sabrina tells us.
David maintains an optimistic outlook for the sector: “The current situation is comparable to the state of autonomous driving: a car can move perfectly well autonomously in certain circumstances, but human intervention remains indispensable at key moments. I believe that we will see 100% autonomous vehicles before we see 100% reliable automatic translators.”
If this reassures literary translators in particular, on the one hand, it may also lead authors to write in a more original and less standardized way, so as not to be decoded and, in the long run, supplanted by the artificial brain. To avoid a future of robot novelists and essayists, it is necessary to experiment with language, to contaminate it in order to keep it alive and personal, even in defiance of the purists. One this for sure, automatic translation programmes will never be able carry out certified translations. For official uses, an accredited translator must take responsibility for their work. This is something that machines will never be able to do. If you're in need of a certified translation, or have any questions at all about the process, feel free to get in touch. We'll be delighted to help. As ever, thanks for reading!
Comments