Reassessing the Poverty of the Stimulus

Existing research on how people learn language has significantly impacted the long-standing nature vs. nurture debate. Evidence for innate linguistic knowledge often derives from certain grammatical properties that are invariant across languages. A stronger motivation, however, draws from the lack of linguistic data displaying these properties in a child’s input, a phenomenon known as poverty of the stimulus. This case is often explained by a grammatical parameter : a single dimension of variation in a language’s grammar that gives rise to correlated grammatical properties. The notion of grammatical parameters has been used to explain how English speakers can learn the impossibility of a sentence like (1) (here, * is used to signify a sentence that is ungrammatical). (1) *The man who you think that saw me just arrived. Holmberg and Roberts [14] argue that examples with the structure in (1) are too infrequent in a child’s linguistic input to be useful to a child during learning. Grammatical parameters account for this issue, since a child could learn the impossibility of sentences like (1) by determining the other grammatical properties that hold in her language. In this paper, I investigate the extent to which the impossibility of sentences like (1) can be explained by a more superficial alternative. Perhaps children understand that a complementizer (like that) followed by a finite verb (like saw) is dispreferred in English, and consequently judge sentences containing sequences such as that saw as ungrammatical. I use statistical models to test whether such a hypothesis could hold and I find that such models are able to succeed in learning that-trace contexts on the basis of their input. This finding brings into question widely held assumptions about the unlearnability of linguistic structures from primary data. If an understanding of grammar is to be motivated by questions of learnability, then these questions must themselves be subjected to serious investigation.