Franz Och: The man behind Google Translate
By Philip Bethge
A German scientist has developed one of the first translation programs suitable for everyday use. Sheer computing power gives the Google software surprisingly good results — perhaps the best yet seen created by a machine.
It’s a good sign when the creator of a piece of software ends up using it. On a recent trip to Japan, Franz Och, who doesn’t speak Japanese, was able to decipher restaurant menus and even read local news — using his mobile phone, which provided him with the translations within seconds.
Och spent the last six years developing Google Translate, a translation program, at Google headquarters in Mountain View, California, “and so far I’ve never really used it myself,” Och admits. But then the 38-year-old research scientist has a change of heart and adds, “I am very happy with what we have achieved.”
Och, a German citizen, is the behind-the-scenes star of a segment of the software industry that has taken on a challenge no less daunting than tearing down global language barriers. In his job at Google, Och wrestles with multi-clause sentences, the subjunctive and auxiliary verbs, to produce a result that is an affront for any linguist. His machine translation program is based on sheer computing power, not linguistic know-how.
The system already commands 52 languages, and the databases for 296 other languages are in development. They include such exotic tongues as Sardinian, West Frisian and Zulu.
Google Translate translates entire websites, theses and even love letters in next to no time, often delivering surprisingly useable results. For Google, the benefits are obvious: With such a useful application, which also happens to be free, even more Web surfers can be lured to the company’s website.
“Machine translation has reached a new quality level,” Och enthuses, “it is much more heavily used all over the place; the software now has an impact in the real world.”
“What Google is doing here is very impressive,” says Alon Lavie of Carnegie Mellon University in Pittsburgh. The computer scientist sees the entire industry in motion. The market for translation software is growing rapidly, says Lavie. “These are extremely exciting times.”
The age of machine translation has begun. Programs like Google Translate are pointing the way to a future in which anyone will be able to speak in foreign tongues at the press of a button. The ultimate goal of scientists who develop translation programs is an electronic version of the Babel fish, the fictitious species British author Douglas Adams concocted in his science fiction classic “The Hitchhiker’s Guide to the Galaxy.” In the book, he describes a leech-like creature, which simultaneously translates any language when it is inserted into a person’s ear. Arthur Dent, the novel’s protagonist, can even understand the crude poetry of the Vogons.
Developers haven’t come that far yet in real life. However, there are already iPhone apps like “Jibbigo,” which translates spoken English into Spanish at lightning speed. Alex Waibel, a computer scientist at the University of Karlsruhe in southwestern Germany, and also at Carnegie Mellon University, created the software. Waibel already uses computers to simultaneously translate many of his lectures, and he has also tested the technology with parliamentary debates.
A translator straight out of a computer lab was long seen as an audacious dream. How was the machine to know that in English, for example, the expression “breaking records” doesn’t usually mean destroying vinyl LPs? Or that the German sentence “wir treffen uns im Schloss,” means “we meet in the castle,” and not, “we meet in the lock” (the German word “Schloss” means both castle and lock).
For a long time, computer scientists tried to cram the necessary world knowledge into the programs using a complex system of rules. But even with straightforward texts, the software often produced complete nonsense. According to Swamy Viswanathan of the US company Language Weaver, the attempt to force the English language, with all of its nuances, into a set of rules is a “nightmare.” “Words often have several meanings, and the number of combinations is endless,” says Viswanathan.
This prompted the experts at Language Weaver to pursue a different concept early on. They fed countless texts from the Internet that already existed in multiple languages into their systems. The specialists’ reasoning was that almost every sentence and every phrase has already been translated many times over, and that pure statistics would suffice to decipher a linguistic construct.
For example, to figure out the German sentence “wir treffen uns im Schloss,” the program searches its database for texts in which the words “treffen” (“meet”) and “Schloss” (“castle” or “lock”) appear in close proximity to one another. Then it goes through the translations of these texts, where it frequently finds the word “castle.” As a result, the computer spits out the phrase “we meet in the castle” and not “we meet in the lock.”
Rosetta Stone of the Digital Age
Och has now perfected this statistical process for Google. During his doctoral work Och, who is from northern Bavaria, specialized in language recognition. Then he went to the University of Southern California. The Pentagon soon began to show an interest in his work. After 9/11, the US intelligence services wanted to be able to monitor Arab newspapers, chat rooms and websites more closely.
But in 2004, Google convinced the language tamer to come to Mountain View, where Och could have the Internet giant’s massive computing power at his disposal. Och isn’t willing to mention any numbers. However, the Google databases contain billions of entries for many language pairs. Important resources for the word archive include, for example, the Bible, which has been translated into many languages, United Nations transcripts and European Union documents, which are available in 23 languages.
Such “parallel texts” are something of a Rosetta stone of the digital age. The ancient prototype bears the same inscription in Greek, Demotic and hieroglyphs.
And now Och’s software is doing exactly the same thing. One of the strengths of the system is that one and the same source code works for all languages. The only catch is that enough translated text has to exist.
A letter-cruncher as a universal interpreter? Many linguists say that such statistical tricks are rubbish. “Statistical translation will quickly reach its limits,” says linguist Martin Kay of Stanford University. “The approach ignores the complex structure of language.” For example, the technology fails when it comes to the positioning of the main verb and auxiliary verb commonly used in German. According to Kay, it also has trouble distinguishing between subject and object.
“For really good results we have to look somewhat deeper in the language,” says Hassan Sawaf, chief developer with the US software maker Apptek, which uses a hybrid approach. In addition to statistical algorithms, Sawaf also applies classic rules of grammar. “This makes the system so much better and considerably improves sentence structure and clarity.”
Sawaf is also critical of the fact that Och’s system only works online. “Anyone who works offline can forget about Google Translate.” German computer scientist Alex Waibel is also skeptical. “Imagine you’re in a foreign country and you want to converse with a salesperson. First you have to find a network, and on top of that, you’ll also be paying high roaming fees. It isn’t practical.”
The fact that Google Translate only works on the Internet is one of its greatest weaknesses. Nevertheless, the California-based company remains undeterred. Its scientists are already developing a special version of the program with integrated voice recognition for Google’s Android mobile phone operating system. The ability to have text on photos translated in no time is also just around the corner. It would enable someone traveling in China, for example, to take a picture of a sign written in Chinese characters, and promptly learn that he is on his way to Beijing.
Another moneymaker for the Internet giant seems to be in the works. But Och demurs. Like many Google employees, he prefers to see himself as part of a campaign for freedom and equality on the Internet. “Someone who doesn’t speak English can only use a fraction of the Internet,” he says. His goal, he claims, is to make the richness of the Internet available to everyone.
There is at least one indication of the programmer’s noble intentions. Och and his team have developed a special program that allows interpreters to feed translations into the system on their own, including the translations of extremely exotic idioms in the Bantu language Xhosa, the language spoken by members of the Ainu ethnic group in Japan and the Inuit language, Inuktitut. The software developers hope that the program will give a voice to languages that are in danger of being forgotten. Te Taka Keegan, a computer engineer at the University of Waikato in New Zealand, has already tested the program with the language spoken by the Maori people. Keegan recently spent six months at Google to figure out whether the digital language miracle from Mountain View could protect the idiomatic expressions of New Zealand’s indigenous people from extinction. His experiences have been consistently positive.
“The quantity and quality of Maori translations is growing constantly with the help of this tool,” Keegan reports. According to Keegan, a digital archive is being developed that will give the language a significant boost.
“The digital world is our children’s future,” says Keegan. “The language will only survive if we manage to make Maori part of this world.”
Translated from the German by Christopher Sultan.
See Related: SOCIAL DIARY