How a Child Gets a Grammar

Somewhere around two and a half, a child who has been correctly saying went starts saying goed. It looks like going backwards. It is the single clearest piece of evidence in the field that something remarkable is happening.

Nobody says goed to a child. It cannot have been copied, and it cannot have been remembered, because there is nothing to remember. The only available explanation is that the child has extracted a rule (add -ed for the past) and the rule is temporarily overriding the irregular forms already stored. The sequence is even more telling than the error: correct irregular, then regularised error, then recovery. That U-shaped curve is very difficult to explain by imitation and entirely straightforward if a rule has been derived.

Jean Berko's wug test of 1958 made the point experimentally and with beautiful economy. Show a child a cartoon creature and say this is a wug. Show two and say now there are two ___. Children say wugs. The word was invented for the experiment, so retrieval from memory is impossible; the plural has to be generated by a rule. The entire argument rests on the word being fake, and the study is still the standard demonstration sixty-odd years later.

Two further facts constrain any explanation. The first is that explicit teaching does not occur: nobody sets out the rules of English morphology to a two-year-old, and it would not help if they did. The second is that correction barely happens and barely works. Parents correct the truth of what a child says far more than its grammar, and children who are corrected on grammar mostly continue unchanged. There are well-known transcripts of a child repeating an ungrammatical form through several patient rounds of correction. Whatever is producing the grammar is not running on error feedback, which rules out a large class of otherwise plausible mechanisms.

Then there is timing. Acquisition is close to guaranteed in early childhood and becomes progressively harder, with native-like grammar rare after adolescence. The strongest evidence comes from two sources: the tragic cases of children deprived of language entirely until late childhood, and the much larger and cleaner population of deaf signers who acquired sign language at different ages. Those exposed early reach native-like grammar; those exposed late generally do not, however many decades of fluent use follow. Nobody has a settled account of the mechanism, and the pattern itself is not in doubt.

The recent development that bears on all of this is the arrival of systems that learn language from text. The comparison cuts in more than one direction, which is why both sides of the old argument cite it. That usable syntax can be extracted from input at all weakens the strongest claims that no general learner could get there from the data. That it takes several orders of magnitude more data than a child receives, and yields a noticeably different error profile, suggests children are not doing the same thing at all.

A practical note, since it comes up constantly. A parent whose three-year-old has started saying mouses after previously saying mice is watching progress, not regression. The child has moved from storing words to running a rule, the rule is winning against the exceptions, and it resolves on its own. It is a useful example of a case where the surface behaviour and the underlying development point in opposite directions.