Blowfish12: Baby robot learns first words from human teacher

Saturday, June 16, 2012

Baby robot learns first words from human teacher

http://c.brightcove.com/services/viewer/federated_f9?isVid=1

AT FIRST it's just noise: a stream of incoherent sounds, burbling away. But, after a few minutes, a fully formed word suddenly emerges: red. Then another: box. In this way, a babbling robot has learned to speak its first real words, just by chatting with a human.

Seeing this developmental leap in a machine may lead to robots that speak in a more natural, human-like way, and help uncover how children first start to make sense of language. Between the ages of 6 and 14 months children move from babbling strings of syllables to uttering actual words. It is a necessary step en route to acquiring full language. Once a few "anchor" words have been established, they provide clues as to where words may start and finish and so it becomes easier for a child to learn to speak.

Inspired by this process, a team led by computer scientist Caroline Lyon at the University of Hertfordshire, UK, programmed their humanoid robot iCub – called DeeChee – with almost all the syllables that exist in English – around 40,000 in total. This allowed it to babble rather like a baby, by arbitrarily stringing syllables together.

The researchers also enlisted 34 people to act as teachers, who were told to treat DeeChee as if it were a child. DeeChee took part in an 8-minute dialogue with each teacher but, between each session, its memory was saved, and then wiped and reset, so that with each teacher, the experiment started anew. At the outset of the dialogue, each of the syllables in DeeChee's lexicon had an identical score.

Lexicon score

All that started to change once the lesson began. Programmed to take turns listening and then speaking, DeeChee turned the teacher's speech into syllables, totting up the number of instances of each one. It then used this information to update the scores in its own lexicon, giving extra points to syllables the teacher had used.

When it next spoke, it would be more likely to repeat the syllables the teacher had uttered because these now had higher scores.

Lyon says this is reminiscent of human infants. "When they hear frequent sounds, they become sensitive to them," says Lyon. "They prefer what's familiar."

This learning by imitation was then reinforced, as teachers made encouraging comments when DeeChee spoke a recognisable word. DeeChee was programmed to detect these comments and give extra points to the syllables that preceded the teacher's approval. Inevitably some nonsense syllables would get extra points too. But as this process was repeated, only those syllables that made up words would keep showing up in strings that gained approval.

Though the robot was still uttering nonsense streams of syllables, towards the end of the 8 minutes, real words kept popping up more often than if DeeChee were still selecting syllables at random.

That words can emerge from babble using a statistical learning process not specific to language demonstrates that this stage of language acquisition does not require hard-wired grammar faculties, says Lyon.

Paul Vogt, a cognitive scientist at Tilburg University in the Netherlands is impressed: "It's a very interesting first step towards having robots that can help us study language acquisition."

Right now, DeeChee's speech is a far cry from full-blown language, but starting with babbling could be the best way to create robots that speak naturally. "If you want the robot to work with natural speech, then you might need to teach it from the very beginning," says Lyon

Blowfish12@2012 blowfish12.tk Author: Sudharsun. P. R.