It's more than 50 years later, and no foolproof Star Trek-style universal translator technology has yet materialized. The time is nevertheless ripe for such automated translation. The $5-billion-plus worldwide translation-services market is overburdened already, and demand is expected to grow to $7.6 billion by 2006 as the Internet becomes more pervasive.
In one of the latest efforts to crack the codes of language with machines, developers of a prototype translation technology hope to challenge the industry with a radically different technique. It essentially throws books in a blender to see how the comparative phrases in different languages stick back together again. Known as EliMT after its inventor Eli Abir of Meaningful Machines in New York City, the statistical technique may prove key not only to making machine translation, or MT, more accurate, but also in quickly rendering translations for languages that are currently neglected by the corporate world.
"The EliMT method is clearly the most promising and theoretically important MT development in the past several years, and probably since the advent of MT itself," claims machine translation expert Jaime Carbonell of Carnegie Mellon University.
The Trouble With Translations
Machine-translation services provided for free via Altavista's Babelfish or Google by industry leader Systran allow for so-called gist translation, where the translation provides the basic idea, with an error rate of 20 to 30 percent. For commercial applications, the extra time to polish out the inaccuracies in gist translations can prove costly: a professional human translator is paid some $20 per hour, and many are so busy that by the time they are available to take on the job, it may be too late to be useful in the cutthroat realm of international finance.
Most commercial MT systems work in much the same manner as how a person at a library might seek to translate a foreign language. First the systems analyze the unfamiliar text. Then they refer to the appropriate bilingual dictionaries and grammar guides. In a way, these "rule-based" schemes are similar to how someone would read a coded text, once that person knew the rules of the code.
However, after working under that assumption, in the 1950s, scientists quickly realized that natural languages are far more complex than artificial codes. This is due in large part to the problem of a how a word's meaning varies with context. The word "cool" used in regards to temperature, for instance, means something different when used by Fonzie. One apocryphal tale dating back to crude, early machine-translation attempts had the idiom "The spirit is willing but the flesh is weak" translated from English to Russian and back again only to yield "The vodka is good but the meat is rotten."
While rule-based MT has improved substantially since then, it's not foolproof. It can take a team years to develop and debug the algorithms to translate any two languages, and every language pair is a whole new endeavor--an English-to-Chinese system won't necessarily help translate Chinese to English or English to Swahili. Since roughly 20 to 30 languages are key economically, there are roughly 400 to 800 language pairs necessary for global finance. So far on Babelfish, only 19 language pairs are available, and other rule-based products do not offer many more options.
Statistics and Words
The EliMT technique works on a different strategy. Imagine a group of people going into a library, looking up the novel Crime and Punishment in the original Russian and then borrowing every English translation of Dostoevsky's work. If they compared how each sentence was translated, they could find statistically that certain phrases were often interpreted the same way. They could then stitch together a translation for a new sentence by recycling pieces of old translations, taking two halves of a sentence from different books. "Instead of translating from word to word, you're translating from sentence fragments to sentence fragments," says Steve Klein, Fluent Machines' chairman and CEO.
While the human brain could never hope to juggle the mental arithmetic involved, computers can. This technique goes through giant databases of translations and breaks the many sentences apart. It then looks for words with a tendency to cluster together. For instance, in a sample English-to-German text, it notes that the phrase "kids love" is linked to 223 occurrences of "kinder lieben," 201 recurrences of "kinder moegen" and 12 incidences of "kleine kinder." Since "kinder lieben" appears with the highest frequency, it will be the preferred translation, although EliMT also notes alternative translations if so desired. Matches between entire sentences and other long word clusters are preferred over shorter building blocks, since words in longer matches often are correctly translated in context.
Statistical MT techniques first emerged about 12 years ago, but since not every word cluster in the world is found in translations, this problem of incomplete databases meant that statistical MT relied on rule-based MT to fill in the blanks. Abir's new system eschews rule-based systems altogether and depends entirely on a statistical solution by looking for overlaps between sentence fragments. While the phrase "kids love chocolate" is not present in the sample text, for example, the segment "love chocolate" is; there are 256 incidences for "liebe schokolade" and 233 occurrences for "lieben schokolade." Even though the former occurs more often, its "liebe" doesn't overlap, so the system goes for the next most popular ranking with "lieben."
Carbonell says he believes EliMT could produce translations of higher accuracy than Systran in about 12 to 18 months--so much so that he applied for membership on Meaningful Machines' board after his assessment. Also, instead of waiting for decades to develop language-pair rules, by entering in translations from any language EliMT should be able to prepare a makeshift database quickly. "For 100 languages there are 9,900 language pairs, and while a shortcut for one pair is nice, shortcutting 9,900 pairs is essential," Carbonell says.
Another possible advantage of EliMT is that it could steadily refine itself in either a fully automated or a human-assisted manner as more data are entered, unlike rule-based systems that require meticulous tinkering with the rules. In addition, EliMT should recognize exactly where faults in its translation might lie to streamline the human editing process. "With other translations, all you know is that it's about 70 percent accurate, and you don't know which 70 percent that is," Abir says. "This system knows what it doesn't know."
Also, unlike other MT systems, results from other language-pairs could potentially help an EliMT translation by matching these segments--what Abir calls "blocks of meaning" or "the DNA of a language"--across different languages. The extent to which this actually might be of assistance, however, needs further testing.
Right now the EliMT system is still in preperatory stages, although the company hopes to field comparative tests soon. The core database may prove to be an unwieldy hundreds of gigabytes large and translation easily takes a great deal of computing power, so the company at this point plans on operating a server through which customers would process translations. Still, in the future Klein says he hopes to help allow real-time translation applications like e-mail, chat rooms and mobile devices. "Right now MT only accounts for 2 percent of the worldwide translation market, but we expect demand will go up once the supply--a near-human automated system--is finally there," he says.
Charles Choi is based in New York City.