New research suggests that the sudden ability of young children to understand and form complex sentences is comparable to a physical phase transition. Using principles from statistical mechanics, Eric DeGiuli at the École Normale Supérieure in Paris has explained the abrupt transition by comparing a child’s learning of language with the freezing of water.
A key stage of a child’s development is the point at which they switch from seeing language as a random jumble of words to a highly-structured system for conveying information. After this transition – which happens remarkably quickly – they can for the first time use grammatical structures to construct meaningful new sentences, even if they have never heard them before. The cause of this sharp transition has mystified linguists for many years, and in this study DeGiuli tackled the problem from the angle of statistical physics.
DeGiuli took his inspiration from the hierarchical tree structure that describes how sentences are formed in almost all human languages. In this model, the sentence forms the “trunk” of the tree, with the branches representing smaller elements of the sentence, such as noun or preposition phrases. The tree continues to branch for successively smaller subdivisions of the sentence, eventually to the point where individual words form the “leaves” of the tree.
As very young children listen to people talking around them in fully-formed sentences, they are exposed to the “surface” of this tree-like network. They start to identify and learn individual words – the leaves of the tree – but cannot yet discern the deep branching structures underneath the leaves.
At this stage, proposes DeGiuli, all possible arrangements of the individual words are equally likely – even nonsensical arrangements that don’t convey any valuable information. This, he says, is similar to the microstates in statistical mechanics, the set of all possible particle configurations in a physical system.
Reducing the possibilities
As children learn, they instinctively reduce the number of possible word arrangements. They assign “weights” to different branches, continually adjusting them in response to the sentences they hear. Eventually, nonsensical branches acquire small weights compared to information-rich branches, effectively “pruning” the tree to discard random word arrangements while retaining those with meaningful structure.
This pruning process can reduce both the number of branches both near the tree’s surface, and those deeper down. According to DeGiuli’s analysis, this plays a similar role to lowering the temperature of a physical system, which at a certain point leads to a phase transition. When water freezes, for example, the reduction in temperature at both the water’s surface and in its interior reduces the number of possible microstates, causing the system to transition rapidly to a more ordered solid.
The theory appears to be consistent with psychologists’ observations of language acquisition in young children. DeGiuli hopes that his results could help to inform neurological studies, potentially enabling researchers to discover how language learning is inhibited in children with learning disabilities.
The research is described in Physical Review Letters.
