Letting math do the talking
Philipp Koehn, Daniel Marcu, Kevin Knight and William Wong, inventors of an advanced method for automated computer translations
Today’s automatic translation software may not be as useful as having C-3PO, the trusty drone from Star Wars fluent in 6 million languages, at your side. However, thanks to German computer scientist Philipp Koehn, dreams of mastering translation via artificial intelligence are no longer restricted to a galaxy far, far away.
Statistical machine translations, or conversions of text from one language to another carried out instantaneously by a computer, were born of the Cold War. At the time, American scientists were under pressure to decipher Russian texts quickly so they could always remain one step ahead of their enemy.
But their problem was speed. The computers were no faster than human interpreters, and their output was often impossible to understand because each grammatical rule had to be coded by hand.
A breakthrough in machine translation wouldn’t come until the late 1980s, when a group of researchers at IBM realised it was much harder to teach a computer the vocabulary and bewildering grammar rules of a new language than it was to simply teach the computer how to teach itself.
A household name
Instead of analysing sentences word by word, the way children learn to do in school, their new technique finds the best statistical match for words based on how often they appeared in pairs of existing translations.
It was as if they had just provided pupils with the answers to an English test rather than handing them dictionaries and teaching them syntax.
“With all the problems in natural language processing, it’s the simple things that work best,” said Philipp Koehn, who works as a machine translation expert at the University of Edinburgh in Scotland.
Having worked on machine translation since 1997, Koehn would eventually go on to become a household name in computer-based translation. Today, if you give the man a computer, he can translate chunks of text into any of 64 languages within seconds.
Babylonian problem, modern solution
Until Philipp Koehn came along, machine translation models had not come close to reaching their full potential. Cumbersome algorithms looked for parallels in original texts one word at a time and simply were not as fast as they could be.
“I’ve always been fascinated with teaching computers to learn. It’s a special challenge because they have to actually understand how we use language,” Koehn said.
After receiving his PhD from the University of Southern California in 2003, Koehn, together with his professors Daniel Marcu and Kevin Knight, filed for a basic patent on statistical machine translation.
He called his method ‘phrase-based’ machine translation and would eventually go on to co-publish a paper describing how phrase-based statistical machine translation models outperform word-based ones because they provide more efficient and accurate results.
Help for anyone with an IP address
The aim of phrase-based translation is to reduce the restrictions under the old word-based system by translating whole sequences of words rather than one word at a time. A word may have several potential meanings (computers unfortunately cannot understand context), but phrases usually only have one.
This idea attracted wide attention in the machine-translation community, and today most of the big names in Internet translation, including Google and Microsoft, have integrated the new model.
Although Koehn has never worked directly with Google, much of his research is implemented in the tech giant’s online translator. “There is a huge world market for ‘imperfect’ translations, where people need a quick translation to understand the general meaning,” he said.
From his office at the University of Edinburgh, Koehn maintains the open-source machine-translation platform Moses, which is essentially a scaled-down version of the system in place at Google.
In contrast to general-purpose translation services, such as Google and Bing, systems built on Moses can be trained on user-specific and domain-specific data, resulting in much better translation quality.
In a single month, the Moses website gets 20,000 page views, and it is used by big companies like Sony and Adobe for document and website localisation, so customers all over the world can read their instruction manuals or properly install a computer programme.
Thanks to Koehn, applications like multilingual news monitoring are also possible. Services such as the European Union’s Europe Media Monitor translate hundreds of thousands of articles every day into 50 languages for analysis and reporting.
Speaking everyone’s language
A lucrative market has already emerged around Koehn’s invention, worth nearly $600 million in 2010. He sees this market continuing to expand and his technology ultimately helping translation specialists in the field.
“It’s all about helping humans – that is professional translators – with their work,” Koehn said. “And in the coming years when the translation quality improves and makes fewer mistakes, we will be able to reduce the costs and the effort of human translation significantly.”
Although it is impossible to gauge the total economic impact of this breakthrough, hundreds of millions of dollars exchange hands each year in language-translation-software markets ($575.5 million in 2010 to $3 billion by 2017), and machine translation is used by export-oriented multinational corporations and international organisations alike.
With their revolutionary method, Koehn and his team opened up a multilingual world of meaning to anyone with an IP address.
How it works
A word-based translation model, the predecessor to phrase-based ones, starts by chopping up an input sentence into individual words and translating each one into the target language. Everything is then reordered so that the target sentence at least appears to have coherent syntax.
Reordering is achieved by learning which words align when translating from a to b and vice versa. Memorising patterns, as children in grammar school do, is helpful here. By comparing these two sets of data, the computer can make an educated guess about the order of the final translation.
One downside: The number of output words is always equal to the number of input words, which often results in superfluous verbal clutter.
Phrase-based translation streamlines this process. Rather than asking the computer to translate each word individually and then splice together a sentence, Philipp Koehn programmed his computer to take groups of words (i.e. phrases) and look for translations that way.
For example, a word-based model would translate the German sentence “Maria geht mit mir einkaufen” by dividing it up into individual words (“Maria”, “goes”, “with”, “me”, “shopping”) and reordering them into a target sentence, “Maria goes shopping with me”.
A phrase-based model would take multiple words, such as “Maria geht” and “mit mir einkaufen”, and translate them as whole phrases. This reduces the number of steps necessary to for a final result, which doesn’t only save time but also leads to cleaner translations.
How do you say ‘innovative’ in Estonian?
The European Patent Office has joined a growing list of multinational organisations, such as the European Union and the World Trade Organisation, which rely on phrased-based translation to connect large multilingual communities.
The EPO’s Patent Translate service, launched through a partnership with Google in early 2012, enables entrepreneurs and inventors to research whether a possible new innovation is truly unique or has already been dreamed up by another inventor who speaks another language.
The online service simplifies what once involved time-consuming and costly research. It currently offers translation between seven European languages and covers about 90% of all patents issued in Europe. By the end of 2014, this list will grow to all 28 languages of the EPO member states, as well as Chinese, Japanese, Korean and Russian.
The “learning” nature of Google’s computer translation technology enables it to adjust to the highly technical language typical of patents and improve translation accuracy as more patents and languages are added.
Machine translation in a store near you
The commoditisation of translation technology is upon us, and recent months have seen a flurry of excitement over highly anticipated simultaneous-translation solutions that seem like they were taken right out of a science fiction novel.
The first is a system by London inventor Will Powell that acts as an interpreter between English and Spanish speakers. Each party dons a hands-free headset attached to a mobile phone and translations appear as subtitles in the lenses of specially designed goggles.
Second is a service from NTT DoCoMo, a major mobile phone operator in Japan, that translates phone conversations between Japanese and three other languages. Introduced last November, it involves recording one side of the conversation, quickly processing the data and then communicating the result to the other end in either a male or female voice.
The third innovation debuted last October, as Microsoft’s chief research officer, Rick Rashid, attended a conference in China. Speaking in English, Rashid’s words were instantaneously translated into Mandarin and displayed as subtitles on an overhead screen. This was promptly followed by a computerised voice reading his Rashid’s words aloud, inflection and all.