A computerised toddler called Hal is the first artificial intelligence program to 鈥渦nderstand鈥� everyday language, say its creators.
They believe that if they can give it the linguistic abilities of a five-year-old, we will soon be able to converse with our computers naturally and possibly hang up our keyboards for good.
The software-based toddler was developed by Artificial Intelligence Enterprises (Ai) of Tel Aviv, Israel. It is said to have fooled independent experts into believing they were reading conversations between an adult and a real 15-month-old child.
Advertisement
Ai鈥檚 chief scientist Jason Hutchens compares the potential impact of the technology to the invention of electricity. 鈥淥nce it exists there are millions of uses for it,鈥� he says.
He predicts that Hal could carry out commands issued without rigid syntax, and could also cope with confusing but similarly structured sentences such as 鈥淭ime flies like an arrow鈥� and 鈥淔ruit flies like a banana鈥�. It might even have a sense of humour.
Learning the language
Unlike the more traditional database approaches to natural language processing, which use statistical techniques to link vast lists of words to a pre-programmed approximation to grammatical rules, Hal attempts to learn language just like we do.
Armed with a collection of learning algorithms, Hal is taught language by a single 鈥渃arer鈥� who types in children鈥檚 stories and responds to its utterances like a parent. Hal鈥檚 only motivation is a built-in desire for positive reinforcement from the carer.
The program, which is small enough to run on a laptop, has no sensory input 鈥� just a stream of words coming in from a keyboard. So far, it is only capable of simple sentences of a few words. For example, if asked what game it wants to play in the park it might respond: 鈥淏all, mummy鈥� (it speaks English, but could learn any language).
In contrast to children, who can take years to learn the basics of language, Hal can be trained in just a few days. This is due both to the intensity of the training and because the algorithms have no distracting inputs, says Hutchens.
Turing test
After training, the carer evaluates how well Hal has responded to new algorithms and stimuli and advises the algorithm designers on what might be needed next. New algorithms are added and the process continues. Instead of telling it how to learn language, Hal figures this out for itself, says Hutchens. 鈥淭he whole point is that we don鈥檛 know how it鈥檚 doing it.鈥�
True natural language processing has been the holy grail of artificial intelligence research ever since British mathematician Alan Turing threw down the gauntlet in his eponymous test for intelligence.
He reasoned that if a person were unable to tell the difference between a machine and another human when conversing with it, then the machine could reasonably be described as intelligent. 鈥淎s long as it behaves like it has an understanding of the outside world then that is good enough,鈥� says Hutchens.
But what most people appear to have ignored, says Hutchens, is another suggestion Turing made. 鈥淗e said that the best way to pass the test is to build a baby machine and train it.鈥�