The history of machine translation stretches back nearly 90 years. Today, neural networks enable machine translation engines to convert entire sentences into a wide range of other languages in just a few seconds. It’s been a long and bumpy journey – particularly at the level of individual words.
1930 – 1950: In the beginning was the word-for-word translation
The first computers sparked a desire to understand and translate languages via machines, and initial attempts were made in France and Russia in 1933. However, these systems, which mechanically compared lexicons, only took into account the root form of each word. Towards the end of the Second World War, development gained both momentum and scientific expertise as British and US forces sought to translate German radio traffic mechanically. In 1950, US researcher Warren Weaver advanced the field further with machine-based code decryption, focusing on language processing above the word-for-word level for the first time. And in 1954, a pilot experiment by IBM and the University of Georgetown automatically translated 250 words from Russian to English.
The results led to euphoric predictions for the future, with both scientists and the general public hailing rule-based machine translation as the solution to all possible communication issues. “Rule-based” meant that the computer automatically adapted the syntax and word forms to manually defined rules for the target language. The catch: the translations sounded clumsy and didn’t even cover the 750 basic vocabulary items required for everyday life.
1960 – 1980: Stagnation and international advances
The US Department of Defense put a further brake on development in 1966, when it declared the technology to be unfeasible and a waste of resources. The result: research stagnated for 10 years. It was only in the 80s, with technological advances in Asia and research groups spread across various countries, that it gained new momentum.
All the research teams wanted the same thing: to generate high-quality translations using fully automated systems. This was an ambitious goal, given that machine output at the time was – at best – only suitable for gathering information or as a draft to be intensively post-edited by a human translator. It wasn’t until the end of the 80s that new example-based approaches made it possible to store and reuse segments of previous translations for the first time via translation management systems (TMS).
1990 – 2000: The great Google coup
Machine translation finally went mainstream in 1990. As PC-based systems and the internet spread, a broad range of people started using it as a communication aid. The 2000s saw the dawn of statistical machine translation, which is based on huge amounts of data gleaned from the analysis of mono- and multilingual text corpora. The best-known example: Google Translate. The technology hit the market in 2006, though inventor Franz-Josef Och, who later became a Google computer scientist, won a prize for its speed as early as 2003.
Statistical machine translation’s main advantage compared to previous systems was that it was considerably cheaper and less time-intensive, as manual data administration was no longer required. What’s more, most algorithms worked across languages and new languages could quickly be added by inputting more data. Meanwhile, hybrid solutions that combined rule-based and statistical programs provided qualitatively better results. They offered more than just two language combinations and fed linguistic insights back into the programming.
2010: Artificial intelligence as a translation aid
The popularity of Google’s translation service became widely apparent in 2012. Even back then, the number of words translated daily could easily have filled a million books.The quality of machine translation again improved noticeably in 2016 with the development of neural networks. Neural machine translation (NMT) was born. Today, programs such as DeepL, Yandex.Translate and Google Translate all make use of artificial intelligence to an astonishing extent. Training material is fed into an NMT system, which learns grammatical relationships from the examples and creates what are known as information vectors for every word and sentence.
These deep learning networks are based on the structure of the human brain, allowing them to store huge volumes of data and achieve extremely accurate translations. We don’t know the precise details of how the networks weight different clauses to produce a final text. However, the result is clear: translations generated using NMT sound much more natural, even for longer sentences.
And today?
Machine translation is currently capable of translating entire sentences into more languages than ever before, learning from them as it goes. Customized systems can be trained using client-specific data, taking into account a company’s terminology and corporate language. The final result is, however, highly dependent on both the engine type and the quality of the training data.
And plenty of challenges are yet to be solved – chief among them the need to produce translations that take into account the stylistic and cultural values of a text. It makes a difference whether a text for machine translation is intended for a US or a Chinese market, and whether it’s a set of operating instructions or a marketing blurb. In order to take such issues into account, the translation engine would have to depart from the word level, which isn’t possible with the current technology.
Whether or not to use machine translation is therefore a question of priorities: text quality versus volume. Check out our recent blogpost for an overview of the factors involved. And humans still have a vital role to play in machine translation: find out more about post-editing.
Cover image via Twenty20