Language by the Numbers
By tapping the increasingly sophisticated "intelligence" of computers, Hopkins researchers aim to make possible the translation of nearly every written language in the world.
By Joanne Cavanaugh
British mathematician Alan Turing, who designed a protocomputer to break the German Enigma code during World War II, also proposed a test in 1950 that he believed would demonstrate when computers reached this level of artificial intelligence, or AI.
In the Turing Test, as it has become known, an observer or "judge" initiates a question-and-answer session via a keyboard with two entities: one a computer, one a person. If the judge can't tell the difference in the majority of cases, the machine could be described as effectively "thinking."
Despite Turing's prediction that by the year 2000 a human judge would not have more than a 70 percent chance of making the correct identification -- and even with a number of contests, including one offering $100,000 in prize money -- no program has met the Turing Test to any degree of general acceptance.
Sure, by the mid-1990s, IBM's Deep Blue computer defeated chess world champion Garry Kasparov in 19 moves, and a computer at the Argonne National Laboratory in Illinois developed a proof for the Robbins Conjecture, a difficult problem that had stumped human mathematicians for more than 60 years. But the question remains: When exactly does a computer understand? What is "understanding" or intelligence anyway?
Here at Hopkins, unique research into the areas of language and computer programming has been probing such questions. As part of the Center for Language and Speech Processing (CLSP) at the Whiting School of Engineering, researchers are training computer programs to "understand," translate, and cull information from texts in Chinese, Basque, Tagalog, Czech, and dozens of other sometimes obscure languages around the world.
David Yarowsky, associate professor of computer science, co-leads the Natural Language Processing, or NLP, research group. "A lot of people in computer science don't worry about whether computers think, or what qualifies as intelligence," says Yarowsky. "That is a philosophical question in the realm of Sartre or Kierkegaard, up there with the question of 'What is the meaning of life?' After a while, what does it matter? If the computer gets so good at something that it looks like it's intelligence, maybe you can just call it that.
"Computers now play chess so well, and some of the questions answered by a machine can be quite sophisticated," Yarowsky adds. "Maybe the computer is just doing sophisticated pattern matching. But if you get back the right answer, does it matter if the computer understood?"
Yarowsky is sitting on a worn brown couch in his department's stripped-down lounge in Homewood's New Engineering Building. He has long been intrigued by foreign languages -- having lived abroad, he speaks Spanish, Japanese, Nepali, and Ladakhi, a Tibetan dialect. And he touts the potential for computer translation of human languages, also known as machine translation, in an ever-shrinking world where what's whispered in a mountain cave in Afghanistan is of interest to the U.S. Pentagon.
Automatic translation technology is useful outside national security circles as well. With vast and ever-growing information sources worldwide, today's scholars and researchers, for example, can't access all the archived texts or published papers -- especially in foreign languages they don't understand. So the ability to use computers to scan texts in various languages for a piece of information, a trend, or a link between disease symptoms, for example, would be invaluable. There are potential benefits, too, to international commerce, where e-mail and Web sites could be more accurately translated, as could manuals, legal documents, and even phone conversations. "The goal is the universality of information," Yarowsky notes.
To help accomplish this goal, NLP researchers are tapping the vast memory capability, processing power, and increasingly sophisticated "intelligence" of computers to make machine translation, as well as information extraction, possible for nearly every written language in the world. As Yarowsky explains: "We want to make humans able to understand foreign languages, and computers able to understand any human language."
|"The notion of a Universal Translator is a very real concept," says Yarowsky. "I believe that in my lifetime we will have computers that can roughly translate all the written languages in the world."||
There's that word again. Understanding.
A computer program named Brutus can now translate simple Latin into English, helping students learn the ancient Roman language. An IBM scientist and world traveler recently used a digital camera and cell phone to send pictures of Chinese grocery signs to a server, where software translated the text and flashed the words in English on his cell phone screen.
American soldiers in Afghanistan this year used a hand-held device called the Phraselator to translate up to 1,000 phrases, including, "I am here to help you" and "Show me your identification," into Pashto, Urdu, Arabic, or Dari. And in Croatia last year, conversation between Croatian and English speakers became possible using a portable computer translator and speech synthesizer.
Such computer-aided language translation seems like science fiction. And Universal Translators like those used by Star Trek's Captain Kirk and crew won't be on Circuit City store shelves any time soon. Nevertheless, says Yarowsky, "the notion of a Universal Translator is a very real concept. I believe that in my lifetime we will have computers that can roughly translate all the written languages in the world."
Yarowsky, who earned his PhD in computer and information science from the University of Pennsylvania, says he was drawn to this field after Harvard University computer science professors showed him how computers could analyze language. "Human languages have so many different interesting properties and complexities," says Yarowsky. An adventurer by nature, he did volunteer work through a Rockefeller fellowship in Nepal and Tibet in the late 1980s, after completing his undergraduate degree at Harvard. In the early 1990s, he worked with speech synthesizers and language analysis at Bell Labs.
About eight years ago, Yarowsky decided to take the academic approach to improving computer-based speech understanding and translation by joining Hopkins' interdisciplinary CLSP, of which the NLP group is the computer science wing. CLSP, which was set up at Hopkins in 1992 with support from the National Science Foundation (NSF), the Department of Defense, and other federal agencies, brings together researchers from six Hopkins departments, including Biomedical Engineering, Cognitive Science, and Computer Science. Through collaborations, researchers focus on such areas as language modeling and acoustic processing (how humans hear language), as well as on how language is acquired. The center, considered one of the best in the world, draws top guest lecturers in the field and hosts an annual international research workshop.
In one project at CLSP, for example, researchers are working on speech recognition technology to help transcribe more than 117,000 hours of interviews with Holocaust survivors videotaped by the Shoah Visual History Foundation. With that much material, it's dauntingly time-consuming -- and exorbitantly expensive -- to have humans transcribe or index every tape. So, as part of a $7.5 million NSF research grant, Hopkins computer scientists are developing software to recognize several languages, including Czech, Russian, and Polish. It's a challenging endeavor. As Bill Byrne, associate research professor in the Department of Electrical and Computer Engineering, has noted, such speech is heavily accented and highly charged. "When people get emotional, the [speech] recognizers have a hard time. But that is the sort of spontaneous speech we want to record."
Yarowsky has long been intrigued by foreign languages.
Having lived abroad, he speaks Spanish, Japanese, Nepali,
and Ladakhi, a Tibetan dialect.
Photo by Jefferson Jackson Steele
Various universities have built large research groups
dedicated to computers and language -- including Carnegie
Mellon, the University of Pennsylvania, and the University
of Southern California's Information Sciences Institute.
Hopkins, through the NLP lab, has found a cutting-edge
niche, specializing in what's known as statistically based
machine translation and text analysis.
Currently, most machine translation technology, including consumer-oriented programs such as Systran's Babel Fish, have been "taught" the rules of language, such as verb tenses and when to use parts of speech. Programmers painstakingly hand-build systems based on such rules. "The computer is told, if you see this thing in Russian, replace it with this thing in English," explains Yarowsky.
While somewhat effective, such systems are time-consuming to build (consider how long it takes most humans to learn a language and all its rules), and resulting translations are still marred by grammatical and other errors. Those that do work fairly well usually tackle popular Western languages, such as French, German, and Spanish; there are few translation programs developed for other important tongues, such as Chinese, Turkish, or Arabic, let alone for more obscure languages like Tajik.
To tackle a broader range of the world's languages, and to improve on the quality of machine translation, Yarowsky and his Hopkins colleagues are developing computer programs that can be trained to figure out any language using statistical analysis, i.e., looking at the probabilities of language patterns. In what's known as automatic knowledge acquisition, the computer could "learn" Serbian well enough to translate future documents or conversation, or at the least pick out pertinent words like "bomb."
As Yarowsky explains: "Say you want to teach a computer how to translate Chinese: You give the computer 100,000 sentences in English and the same 100,000 sentences in Chinese and run a program that can figure out which words go to which words. If in 2,000 sentences you have the word Washington, and in about the same number of sentences you have the word Huashengdun, and they occur in the same place in the sentence, these words are likely translations.
"It's all just observation," Yarowsky adds. "Children do the same thing, but they also do it through visual stimulation and feedback. They see a book and hear the word 'book,' and eventually they learn that it's a book. They see a bird with its wings flapping around and learn that is called a bird. It's the same with machines, only they have much better memories. Computers could remember exactly when and where they saw the words bird and book."
So, instead of telling a computer how to do something -- conjugate the verb 'to be' in Spanish, for example (I am = soy) -- researchers give it tens of thousands of examples and program the computer to find repeated patterns that the computer can use to conjugate new verbs. Trained this way, the program could potentially "learn" phrase structure and the rules of translation.
As Yarowsky notes in his 100,000-sentence example, one way to accomplish automatic knowledge acquisition is to use bilingual or parallel text. The program "reads" a document in English and then a version in a second language. Such texts used by Hopkins researchers include the Bible, which is available on the Web in more than 60 languages, the Book of Mormon (over 60 languages), and the United Nations Declaration of Human Rights (240 languages).
Aiding the computer is the fact that the English version of such texts can be annotated by hand or using another computer program -- essentially marked up to show, for example, that Jesus is a noun and pray is a verb. The translation program-in-training needs such information because it cannot translate future text just by substituting individual words in each language; it must also be able to analyze how sentences work. To do so, the computer program uses pattern recognition templates and other tools to understand sentences on a syntactic level. Simply put, the program is essentially given clues to know what to look for, notes Yarowsky: "It should figure out the subject, figure out the object, and other elements of sentence structure."
Other tools used by Hopkins researchers to train computer programs to translate languages include bilingual dictionaries or lexicons that can be fed into the program, as well as WordNet, a thesaurus of sorts that shows links between words like pain, headache, and migraine. The end result: A computer program will be "trained" to translate Pashto or Basque or Hindi into English, even though it doesn't actually understand them. Or does it?
"It sort of understands," says Yarowsky. "It partially understands some of the ambiguities, some of the meanings when words can mean multiple things. It can understand a lot of the structures of language, but it won't understand deeper subtleties. Some languages, for example Chinese, don't distinguish the male and female pronoun. He or she is the same word, so it can be ambiguous who something refers to. And sometimes there's a subtle metaphor."
So far, statistically based translation is faster to develop and more flexible, though often more plagued by grammatical or translation subtlety errors than the rule-based approach. Hopkins researchers have trained a program for Chinese, as well as one for Czech and French, that could roughly translate nearly any text. They are pursuing other projects with data from 240 languages. "It's intense work," Yarowsky notes. In some languages, like Turkish, a whole sentence can be represented by a single word and with Chinese, there are no spaces between words. A Chinese translation program created by Yarowsky and his colleagues already has outperformed current commercially available programs at recent machine translation competitions. "It's much more accurate on news text, which is what it was trained on, but it probably won't do very well on poetry," Yarowsky says. "Its accuracy depends on how many training sentences it has seen."
Solving different pieces of the puzzle: (l to r)
Eisner, Florian, and Schafer
Photo by Jefferson Jackson Steele
A famous anecdote in the machine translation field centers
on the biblical saying "The spirit is willing, but the
flesh is weak." When the phrase was translated into
Russian by an early computer translation program in the
1950s, the story goes, the answer came back: "The whiskey
is strong, but the meat is rotten." Over the years, that
story has been debunked as myth.
Yet enter the same phrase into Babel Fish Translation online today and translate it into, say, Spanish, and the answer comes back, "The alcohol is ready, but the meat is weak." For some real fun, translate that back into English. The resulting phrase harkens to that game known as "Telephone" where a phrase is passed down the line and misinterpreted along the way. The next Spanish-to-English version reads: "The ready alcohol this, but the meat is debil." And that's for two of the most commonly spoken and computer-translated human languages.
In the 1950s, during the infancy of machine translation, hopes were high that systems would soon be developed to rival high-quality human translation. The United States government poured millions of dollars into projects, fueled by an interest in Cold War-era translations and language analysis of Russian documents and radio transmissions.
With all the early limitations in hardware, software, and computer memory, the first machine translation researchers relied almost solely on bilingual dictionaries, and word-for-word translation. But researchers quickly realized that "perfect translation" was more difficult than they imagined. A federally commissioned report by the Automatic Language Processing Advisory Committee (ALPAC) found that machine translation had failed to reach its goal of adequate-quality translation by the 1960s, and likely would never be cost-effective. Generous funding sources soon dried up.
The Holy Grail question then and now remains: Will a computer ever be as good as a human translator? In many ways, not even close, at least until AI reaches the level of Star Trek's android character, Data. That's because language, in its many forms, is complicated and nuanced, ambiguous and contradictory, illogical and artistic -- much like humans themselves. "Language is an incredibly complex, multifaceted puzzle, too big for any one person to solve," Yarowsky says.
Nonetheless, advances are being made today. And researchers are finding that machine translation doesn't need to be "perfect" to be useful. Computers, in some cases, can do much of the heavy lifting in translation, with post-editing being done by humans. Partly to minimize such clean-up measures, Hopkins NLP researchers are tackling theoretical research in language acquisition and creating practical tools to improve translation.
Gideon Mann, now starting the fourth year of his PhD in computer science at Hopkins, says he was a fan of science fiction who hoped someday to converse with Asimov-style robots: "When I grew up, I was really upset that there weren't any computers I could talk to, so I thought, 'I guess I'll have to build them.'"
|"When I grew up, I was really upset that there weren't any computers I could talk to, so I thought, 'I guess I'll just have to build them,'" says Mann.||
So far, Mann is developing software that can answer simple
questions by analyzing sentences. Say, for example, that
one has the question: "When did Hitler's armies invade
France?" Mann's programs can search the Internet, looking
for Web pages where a date or year is found near words
from the question (i.e., invade, France, Hitler). In this
case, "1940" would be the program's most confident answer
based on statistical analysis relative to the syntactic
context. In general, "the Web has a nearly limitless
supply of information, and the more we understand about
language structure, the more effectively we can harness
this information," says Mann.
While such approaches are incremental and highly specific, these are the building blocks on which language "understanding" works -- for humans as well as computers. Yarowsky, and the other researchers in his lab, are, in a way, engineers and architects and general contractors figuring out how to make each piece of the computer-language edifice fit together.
Linguistics is at the cornerstone of their endeavors.
The NLP lab's co-leader, Jason Eisner, assistant professor in computer science, uses a familiar computer science tool known as "finite-state machines" to program computers to analyze sentences on a highly syntactic level known as parsing -- much like how English students look at the logical structure of sentences when diagramming parts of speech.
Richard Wicentowski, a linguist and computer scientist who has just finished his PhD at Hopkins, has been working with morphology, or the study and description of word formation. "Basically, it's the way that new words are built up from old words," Wicentowski says.
To provide a clearer picture of this linguistics-computer science link, Wicentowski explains how he trains computer programs to discern whether one word is related to another, such as drink and drank. "What you are trying to do is find ways for the computer to automatically discover the relationship between drink and drank," Wicentowski says. One way is to recognize that the words are nearly the same, except for one letter. Or, the program could scan nearby words, such as Coke or milk, for clues.
In a unique demonstration of how this technique could be used in any language, Wicentowski trained a program to conjugate Klingon, a language made up by particularly avid Star Trek fans. "It turns out Klingon is a very easy language for computers to learn because although it is complex morphologically, it was designed very consistently by one person," he says. Though Yarowsky's office boasts a copy of Shakespeare's Hamlet translated into Klingon by the Klingon Language Institute, neither he nor Wicentowski speaks Klingon. (The obvious question? If you want to ask, "To be or not to be?" in Klingon, simply utter "taH pagh taHbe'!".)
Wicentowski says that using Klingon in translation and language research emphasizes how a computer program doesn't, in his opinion, actually "understand" the text: "The computer couldn't possibly understand what it is doing because I'm the one who told it what to do, and I don't understand."
|The ambiguous meaning of words remains a primary hurdle. The word "plant," for example, could refer to a biological organism, a factory, a police "plant," or a ringer in the audience. How's a computer program to know?||
For researchers like Wicentowski, it's the ambiguous
meaning of words that remains -- as was shown by the
spirit-is-willing example -- one of the primary hurdles.
The word "plant," for example, could refer to a biological
organism, a factory, a police "plant," or a ringer in the
audience. How's a computer program to know? The process to
clarify the meaning of such words in various languages is
known as "word sense disambiguation."
Radu Florian, also finishing his PhD, has been working on algorithms, or sequences of instructions, that teach computer programs to assign a specific sense to a word by giving it a large number of examples for when each meaning of the word is used. Through statistics, the program will know there's a 70 percent chance that when it sees the word worker near plant, plant will likely refer to a factory. "The program is given different parameters for different words," Florian says. "If the word leaf is near the plant, it would know that it's a living plant, not a manufacturing plant."
Yarowsky envisions how advances like those being pursued by himself, Eisner, Florian, and others will inevitably propel statistical machine translation to the next plateau. "With each [researcher] tackling a different piece of this puzzle, he says, "they can help provide an end-to-end solution."
A database residing at the NLP lab holds two terabytes of memory -- that's 2,000 billion bytes or characters of text. And lab researchers have filled most of that memory up with stored text from over 100 languages, mostly news stories pulled off the Web. On a daily basis, a computer robot that acts like a super search engine accesses the Internet and automatically visits many of the newspapers and news sites in those 100 languages and downloads information.
"It takes the pages and strips the images and the ads and what's left is a news story about the events of the day. We try to line those stories up across languages," Yarowsky says. "If there is an earthquake in Chile, for example, a story on the earthquake might run in Poland, and China, and in Bangladesh."
Though the stories won't necessarily match word for word, much of the content, including the use of the word "earthquake" in various languages, will be similar. Through a process known as "iterative alignment," a computer program, given enough text, will start to pick out such similarities and translate key words.
Before the advent of the Web, and the subsequent explosion of sites in hundreds of languages, the availability of bilingual text was limited -- especially in such languages as Azeri, Icelandic, or Punjab. Today's researchers, however, can in most cases find the comparable documents they need to train translation programs, whatever the language.
Hopkins graduate student Charles Schafer does research in information projection across languages. He uses bilingual texts to take NLP programs that analyze English and automatically develop the same analysis skill for a different language.
"Say you have a program that reads English sentences and identifies where people are claiming responsibility for bombings -- people have spent lots of effort creating this capability over the years," Schafer says. "We can then run our existing programs on the English text, and use statistical techniques to figure out what kinds of clues in the Arabic translation indicate people claiming responsibility for a bombing. So you get the Arabic NLP program for free -- as long as you can find the translated texts you need for this technique."
Schafer, now in the fifth year of his PhD, also was drawn to this focus in computer science because of a fascination with language, but in his case it was the history of English and the origins of words. He doesn't own an Oxford English Dictionary, though he points to a tattered expanded Random House sitting on a shelf by his desk. "The OED is on my wish list," says the graduate student.
Schafer's wish list includes perfecting the area of science he intends to make his career. It's a long shot, he knows. "We can make estimates that in several decades we will have one million times the processing capability," he says. "For the time being, we can improve. But it won't be human-quality anytime soon."
But that doesn't mean computer scientists can't dream about what Turing himself envisioned as the spark of nonbiological intelligence that could someday lead to a deeper level of understanding, perhaps even surpassing that of humans.
Hopkins PhD student Florian tells a well-known joke about scientists building a computer as large as a planet. Once they build it, they try to figure out what to ask it. Eventually, they decide on the most central question plaguing humankind since the dawn of civilization: Is there a God?
The answer: "There is now."
Joanne Cavanaugh Simpson is a senior writer at Johns Hopkins Magazine.
The Johns Hopkins Magazine | The Johns Hopkins University |
3003 North Charles Street |
Suite 100 | Baltimore, Maryland 21218 | Phone 410.516.7645 | Fax 410.516.5251