Language learning from fragmentary input

A model of vocabulary and grammar acquisition is presented. Two agents are involved in the simulation, a mother and a child. The mother is equipped from the outset with a substantial knowledge of language, in the form of two sets of rules. Her lexical rules map atomic meanings (‘concepts’) onto words. Her grammatical rules state generalizations about mappings between complex meanings and strings of words. The mother’s rules allow her to utter strings of words expressing any meaning drawn at random from a large set. At the outset, the child has no such rules. The child does, however, share his mother’s capacity for semantic representation; he has access to the same set of propositional representations, composed of the same atomic concepts. The mother utters wordstrings, which the child hears in full, but for each string, the child is made aware of only a small fragment of the mother’s original meaning. From this exposure to wordstrings and small fragments of meaning, the child acquires a set of rules functionally equivalent to his mother’s, and is capable of expressing the whole range of meanings with the same wordstrings as her. The child has fully acquired his mother’s language, from data that is semantically highly degenerate. Early lexical acquisition is bootstrapped from observed correlations in the child’s input. Grammar acquisition depends on some earlier acquired vocabulary. Later lexical acquisition takes advantage of acquired grammar rules. The whole process is informed by the learner trying to make sense of the data.