Scientists teach computer to speak

Scientists in Israel are treating their computer like an unruly child, correcting its mistakes and punishing it for errors.

If you've ever felt the urge to reprimand your PC after it accidentally deletes a day's work then you might sympathise with this behaviour, but there is more to this than meets the eye. These researchers are, in fact, conducting pioneering work into the development of personal computer that they hope could eventually carry out a meaningful relationship with its owner.

The company behind this work, Dutch-based firm Artificial Intelligence (AI), caused a stir in technology circles recently by claiming to have developed a computer that had learnt language to the level of a 15 month-old child. At a research facility in Tel-Aviv, Israel, Artificial Intelligence (AI) demonstrated its computerised toddler -- named HAL -- for the first time and revealed that it is betting million pounds on the scheme.

For such ambitious goals, HAL's design is surprisingly simple. Running on top of Windows 2000, HAL is little more than handful of simple learning algorithms. But from these simple building blocks, HAL's creators hope to generate a computer capable of holding its own in conversation.

Past attempts have typically involved programming computers with basic rules of language, a technique that has always fallen short of providing stimulating and, at times, even comprehensible conversation. This may be because, according to linguistic theorists, the rules of language are virtually impossible to pin down, existing only in a state of flux.

In contrast to these previous attempts, HAL is designed to learn language for itself. Simple learning algorithms allow HAL to mimic different patterns of text characters. The real innovation occurring at AI, however, is that a child psychologist interacts with HAL, correcting mistakes and nurturing correct use of words and sentences.

At the heart of the learning algorithm is the interplay between the computer's prior experience and its ability to predict what output will be rewarded. The approach is taken from Bayesian statistical analysis; an area of mathematics that involves calculating probabilities according to newly accumulated knowledge. Andrew Blake, a mathematics and electronics researcher at Cambridge University, describes this as "a very strong field in AI at the moment".

Jason Hutchens, chief scientist and programmer on the HAL project, says that this basic probabilistic approach enable the computer to form its own "atoms" of language, which turn out to be coherent words and sentences. Hutchens has a growing reputation in the field of artificial intelligence. In 1996, he won the Loebner prize, awarded by the Cambridge Centre for Behavioural Studies in Australia to the most human-like conversational program entered, with a simulator called Hex. He also helped design the artificial intelligence engine in the popular and impressive new computer game Black and White.

"Hutchens is a very clever guy," says Professor Yorick Wilks, head of the Department of Computer Science at the University of Sheffield, who's own research has been entered into the Loebner prize in the past. "He's proved that he can create a convincing conversational machine already."

However, as Hutchens says, he is not concerned with creating a machine that is genuinely intelligent, assuming such an abstract thing as intelligence can ever be measured. Instead, he and the rest of AI simply hope to create a computer with the appearance of intelligence.

According to some measurements, this is all it takes. In 1950, British mathematician Alan Turing proposed a test for intelligence whereby any machine capable of convincing a human that it was talking to another human should be considered intelligent. This test still used by some artificial intelligence researchers today.

Ironically, because HAL is not programmed with hard and fast rules, just basic learning ability, Hutchens cannot know how exactly HAL processes language. "I have an understanding of the algorithms and I have my expectations but HAL is constantly surprising me," he says. "Hal knows that cat and dog are associated through a set of symbols. It is working at the moment and working beautifully."

The HAL project is clearly more than an exercise in software engineering, however. Dr Anat Treister-Goren is the linguistics expert and child psychologist charged with tutoring HAL and she says that her young pupil is coming along well. "HAL went through a babbling stage," she says. "First he only produced word-like utterances. He's definitely in the area of an 18 month now."

Testing HAL according to a Child Language Analysis Program (CHAN), Treister-Goren says that her "child-computer" currently passes for a 18 month old child and has a 50 or 60 word vocabulary. By the end of 2003, AI expects to have a version of HAL capable of talking like a three-year-old and by 2005 hopes it that that it will have the conversational skills of an adult.

Steve Grand, author of Creation: Life and how to make it, says that learning is undeniably crucial to creating machine intelligence. "You can't have language unless you learn to use it," he says. But you have to talk to it an awful lot to make it

Whatever the case, even at its early stages, HAL has a certain charm. Goren admits to having an emotional attachment to HAL. "Sometimes I don't have the heart to tell him that he's wrong. He is so good and so creative," she says. "One time I went home and a funny thing happened: My kid said, 'You love HAL more than us.'"

Simulating capabilities of intelligent humans, especially language, with computers is perhaps the most sought after and most illusive goals in science. Researchers have wrestled with ways of making computers more human for decades, with different schools of thought, including expert systems, fuzzy logic and neural networks successively promising a revolution that will lead to artificial intelligence

Such a revolution might change the way computers are used altogether. Jack Dunietz, president of AI and currently the project's sole investor, certainly has high hopes that his baby will bring about a revolution in personal computing. "It will be the end of the graphical user interface," he claims. "It's a quantum leap. A paradigm shift." Dunietz believes that within ten years AI will introduce a computer that can learn about and converse convincingly with its user and that AI will just one of many start-ups to usher in an era of computing dominated by artificial intelligence. "I've never seen such a big wave as that of AI," he says.

Few could be more suited to pursuing such an ambitious goal. Dunietz made his fortune by investing in computer companies in the late 70s and early 80s but in 1992 he returned to university to study linguistics. He has so far put "a few million" dollars of his own money into the project but admits that it is likely to take another $50m dollars or so to see it through. He claims that investors are knocking on the door. "There is no question of raising money for this project," he says. "The question is over the profile of the people we want to invest." Dunietz says that AI will announce third party investment this year.

Dunietz, Hutchens and Treister-Goren all concede that a number of uncertainties hang over the HAL project. Not least of these is the fact that HAL currently only interacts with the outside world in the form of text, even though AI hopes to introduce voice recognition by 2005. Hutchens acknowledges that a lack of other sensory input restricts the computer's ability to make links between different words and language contexts.

But once the basic learning algorithms have been perfected with text, Hutchens says, HAL will then be able to learn using to the phonetic rather than text characters. "The Helen Keller analogy is a good one," he says. "Eventually we could use phonemes rather than characters to give the depth of human interaction."

Another possible stumbling block is the amount of processing power that may be needed to bring HAL to full functionality. HAL currently runs on modest PC hardware but Hutchens says that performing complicated statistical predictions based on an ever-growing memory is increasingly power intensive and is already looking at moving HAL to a server system. "We are hitting performance issues," says Duneitz. "But we have a team engineers working on efficiency alone."

Andrew Blake at Cambridge University cautions that such a simplistic approach as AI's has yet to be proved at more advanced level. "Its true that simple things often work, but the idea the idea that you can capture huge slices of human intelligence so simply remains to be seen," he says.

At the University of Sheffield Wilks takes a more optimistic tack and suggests that AI's daring start-up attitude could eventually be the secret of its success. "They're ambitious, have money and really want to do it and it's always been that way with AI research," he says. "It never quite measures up to what you expect but often it takes you further and, who knows, they might just pull it off."