Scientists Create Artificial System Capable Of Learning Human Language

An artificial system of nerons called ANNABELL has learned to pick up language by interacting with humans. Aysezgicmeli/Shutterstock

The ability to learn and creatively use language is among the most impressively complex cognitive processes, and continues to set humans apart from even the most advanced machines. However, a team of scientists has now created an artificial system of neurons that is capable of learning words, phrases and syntax with no prior programming, thereby sustaining a dialog using processes that resemble mental actions.

The program, called ANNABELL (Artificial Neural Network with Adaptive Behavior Exploited for Language Learning), is made up of 2.1 million “neurons,” linked via 33 billion virtual connections – and the source code of the software can be downloaded here. While this may sound like a large and complicated system, researchers say the sophistication of its language use was comparable to that of a four-year-old child. This included the ability to “learn nouns, verbs, adjectives, pronouns and other word classes, and to use them in expressive language,” according to a paper published by the team in the journal PLOS ONE.

What distinguishes ANNABELL from previous human-machine dialogue programs is the fact that it began from a state of tabula rasa, meaning it had no a priori knowledge of the words or phrases it eventually came to use. Rather, it was able to learn, organize and select its words purely by picking them up from interactions with a human.

This was achieved thanks to two of ANNABELL’s key properties. The first of these is synaptic plasticity, which refers to the process whereby a connection between two particular neurons – a synapse – becomes increasingly efficient the more it is activated. The other vital mechanism is neural gating, relating to the ability of certain neurons called bistable neurons to act as switches that can activate or deactivate the transmission of signals around the brain, thereby controlling the flow of information.

Using a text-based interface, the program was fed 1,587 input sentences, from which it was able to generate 521 responses. For instance, when asked the question “how many games did you play?”, it answered: “I played Space Invaders, one; Pac Man, two; Donkey Kong, three; three games.”

The researchers, from the University of Sassari in Italy and the University of Plymouth in the U.K., insist that their intention was not to engineer a “solution to the human-machine dialogue problem,” but to shed light on the cognitive mechanisms by which language is learned in the human brain. As a result of their work, they propose the hypothesis that the elaboration of verbal information is controlled by a central executive, meaning a supervisory system of neurons responsible for coordinating the flow of information, regulated by neural gating mechanisms.


If you liked this story, you'll love these

This website uses cookies

This website uses cookies to improve user experience. By continuing to use our website you consent to all cookies in accordance with our cookie policy.