Children and language learning

Paris, August 30, 2002

A team of researchers from S.I.S.S.A* (Trieste) and from the University of Ferrara, working in collaboration with researchers from CNRS, have performed a series of experiments on learning of artificial languages, aimed at gaining a better understanding of the mechanisms of language learning in children. Their results show that the human brain has at least two distinct learning modes that are used on speech: a statistical mode which, in a continuous speech stream, analyses transitions between syllables and identifies the successions of syllables that make up words; and a rule-learning mode which is triggered only once the words have already been identified, and which is used in particular in learning rules of syntax. These results fit into a broad debate between those who think that learning a language can be explained exclusively by very general mechanisms extracting statistical regularities, and those who think that it also requires formal rules to be learnt, in particular in the field of syntax. The work of the Franco-Italian team would tend to support the latter. This research work was published in Science on August 29, 2002.

To acquire a command of a language, it is necessary to build up a vocabulary and a system of rules for generating words and sentences. These two tasks are not simple, in particular since spoken language (unlike written language) does not contain any physical signals marking the beginning and the end of a word. The way in which children rapidly develop their lexical and grammatical knowledge remains a mystery. Research published in Science in 1996 revealed that children and adults have unsuspected capacities for performing complex statistical computations that could help them to find the words in a continuous speech signal. The impact of that work was very considerable because several researchers thought that the statistical capacities discovered by the authors could suffice to explain learning of all of the aspects of a language, including its grammatical properties.

The team of researchers from S.I.S.S.A (Trieste) and from the University of Ferrara, in collaboration with researchers from CNRS, performed an experiment essentially putting adults under the conditions of a baby who must find words or rules within a continuous speech stream coming from a language that it does not know. For this purpose, the team invented small nonexistent “languages” and used a computer to synthesize continuous speech segments of these languages, while eliminating any prosodic information. The task of the subjects was to find the “words” of the imaginary language after having listened to the speech stream for a period varying from 30 minutes to 2 minutes, depending on the experiment.

The researchers discovered three basic facts. Firstly, they showed that adults can perform statistical computations that are even more complex than those discovered previously. Secondly, the researchers showed that, in spite of the subjects’ statistical skills, they were unable to discover the structural properties contained in these imaginary languages, whereas they could have done so easily on the basis of the statistical computations that they performed. Finally, the researchers concealed 25-millisecond pauses in the continuous speech stream. These gaps indicated the frontiers between the words but were not consciously picked up by the subjects. They also showed that, when the speech stream contained these subliminal segmentation signals, the subjects rapidly captured the same structural property that previously escaped them (with exposure of only one fifth of the previous exposure).

Theses three results would suggest that the mind is richer than we thought. Its statistical capacities are greater, and yet they are insufficient to capture simple generalizations that are necessary to learn a language.

*SISSA: Scuola Italiana di Studi Superiori Avanzati (Italian School of Advanced Higher Studies)

Reference: Signal-driven computations in speech processing, Marcela Peña, Luca L. Bonatti, Marina Nespor, Jacques Mehler, Science, August 29, 2002.

Researcher contact:
Jacques Mehler
Laboratoire de sciences cognitives et psycholinguistique (CNRS-EHESS) - SISSA
e-mail: mehler@lscp.ehess.fr
Luca Bonatti
Laboratoire de cognition et activités finalisées (CNRS – Université de Paris 8)
e-mail: lucabonatti@mac.com

CNRS press contact :
Carine Noël
Tel: +33 1 44 96 49 88
e-mail: carine.noel@cnrs-dir.fr