How do children learn grammar rules nobody ever teaches them?

Children end up knowing grammar rules that nobody taught them and that the speech around them may never clearly show, and linguists still argue about how. The puzzle is called the poverty of the stimulus. Noam Chomsky, who coined the term in 1980, argued that what children hear is too thin to learn every feature of a language without built-in, language-specific biases, a 'universal grammar'. Critics answer that ordinary learning is more powerful than that argument assumes.

Here is the kind of case Chomsky's side points to. In 'While he was dancing, the Ninja Turtle ate pizza', 'he' can be the Ninja Turtle. In 'He ate pizza while the Ninja Turtle was dancing', it can't. Every English speaker knows this. Yet children only ever hear what people do say, never a correction about what can't be meant, and their input is full of slips, false starts and non-native speech.

The other side has striking evidence too. In a classic study, 8-month-old babies listened to just two minutes of a monotone stream of nonsense syllables and then reacted differently to its 'words' than to the same syllables in a new order: they had picked up the statistics. More recently, some neural networks have learned hierarchical sentence structure without being given that constraint, which suggests some poverty-of-the-stimulus arguments may have been mistaken.

Chomsky calls the deeper question 'Plato's problem', after the dialogue in which Socrates draws knowledge of geometry out of a slave who was never taught it: how do we know so much from so little?