History of Machine Translation

[0:16]Hello, and welcome to this video in which we will tell you a bit more about the history of machine translation. According to the homepage, Google Translate is used by over 500 million people every single day. And by 2021, the Google Translate app for mobile phones had been downloaded over a billion times. Most social media platforms like Facebook and Twitter offer automatic translation functions, as does the newest version of Microsoft Outlook. But have you ever wondered when all of this began? The field of machine translation is much older than you perhaps think. In fact, the technology dates back to the Second World War, when cryptography was all the rage. The German army had Enigma, a machine they used to send encrypted messages which were for a long time considered to be an unbreakable code, until Alan Turing built another machine and the puzzle was solved. While this makes for exciting movies and TV shows, it's important to realize that in those days computers looked nothing like the ones we have today, and there was no internet yet. The technology was in its infancy. Yet an American mathematician called Warren Weaver already envisioned that it should be possible for machines to translate text automatically. He wrote a memorandum in which he argued that translation could be considered similar to code-breaking, where a message in language A has to be encoded first into a transfer language and then decoded into language B. Weaver's memorandum sparked a lot of research on the first generation of machine translation systems, especially by the American and Russian governments, who were fighting the Cold War and desperately wanted to translate enemy messages without having to rely on a human translator, who could very well be a spy or a double agent. The Massachusetts Institute of Technology hired the first full-time researcher machine translation, Bar-Hillel, who organized the first international conference on machine translation in 1952, and led the Georgetown IBM experiment in 1954, when a Russian to English machine translation system successfully translated 60 sentences during a demo. And while the success of this first rule-based system was in fact small, it generated a lot of attention and attracted massive funding. It was believed that fully automatic, high-quality machine translation would soon be a reality. Unfortunately, this was not the case because rule-based systems were simply not very good at translating anything but highly controlled and unambiguous sentences. In 1966, the Automatic Language Processing Advisory Committee or ALPAC evaluated the current state of machine translation research and wrote a report in which they said that machine translation was simply too slow, too expensive and of poor quality. They proposed focusing on developing tools that would support human translators such as term-bases and translation memories, rather than continuing the search for fully automated translation. As a result, most of the funding dried up and very little happened for a very long time. While not much was happening in terms of machine translation research in the 1980s, computer technology was developing fast. Computers became much smaller and faster and much more powerful. In addition, the internet and the worldwide web were quickly becoming popular. As a result, machine translation engineers started thinking about machine translation in a different way, and decided that they should not try to have machines translate like human translators do, but to have the machines do what they're good at, namely finding patterns and running calculations, really, really fast. This led to the development of the first statistical machine translation systems, which used large electronic parallel corpora as their training data. Rather than using vocabularies and grammar rules, the systems relied on statistics to predict the right translation, and it worked. Because computers are so good at pattern matching and running calculations, and because they now had access to so much data, thanks to electronic corpora and the internet, the results of the statistical machine translation systems were much better than the old rule-based systems. As a result, the field found new energy and attracted new funding. Statistical machine translation systems remained dominant from the mid-1990s until the launch of Google Translate's neural machine translation system in 2016, which changed the field completely. Suddenly, everyone everywhere across the globe had free and unlimited access to high quality machine translation between over 150 languages. The switch from statistical to neural machine translation systems meant a giant leap forward in quality. In fact, the quality of current neural machine translation systems is deceptively good. It has become much harder to see the errors, and neural systems are much less predictable. Using free online machine translation systems is easy, but understanding how to use these systems effectively and ethically is not. This has launched the machine translation literacy movement which our videos, infographics and website are also part of. If you're interested in learning more about how those different systems, rule-based statistical and neural work, or if you want to know more about the legal issues and risks involved, we have more videos and infographics ready and waiting for you. To end with, we would like to recommend Lin Bowker's chapter on machine translation in her most recent book De-mystifying Translation, which you can download for free via the link provided. Thank you for watching this video.