Rule generalization from impoverished input

点击翻译按钮获取中文摘要

My NSERC program aims at determining the mechanisms underlying children's amazing ability to learn language. My past NSERC projects studied how infants learn & use functional items (functors) for early acquisition. Functors, which are highly frequent words & affixes (e.g., articles, tense endings), have crucial grammatical functions. They are, however, absent in children's early speech. Thus, the classic view is that children have no grammatical knowledge in the first 2 years of life. I challenged this view, by testing the prosody-functor bootstrapping model. Using subtle, sensitive measures of perception & comprehension, we discovered that infants have knowledge of functors from the first year of life, long before they can talk, and that functors help them break into the grammar, starting around the first birthday. Our findings changed the established view about the early state of linguistic knowledge. Our new research will focus on rule generalization involving functors. Specifically, we will test how infants acquire abstract grammatical rules from inconsistent input. Infants do not receive instruction on rules from parents; they rely only on hearing rule output to derive the rules & to generalize to novel instances. But children's input is known to be impoverished, with partial rule output & exceptions. These observations motivated Chomsky to posit that the grammar cannot be induced from the input, and the child is guided by her innate linguistic knowledge. However, many others believe that input is rich in distributional cues and that the child is capable of inducing a grammar from the cues. My new NSERC studies will be the first to test these important theoretical positions. We will first test the hypothesis that the language-universal frequency property of functors (i.e., a tiny set but highly frequent relative to a massive set of infrequent lexical items such as nouns/verbs) is crucial for grammatical rules to be learnable. Infants will be trained & tested in artificial languages, and the hypothesized learnable & unlearnable input groups will be compared. The training will be fully rule-consistent. Next, we will use artificial grammars with the learnable property of functors, but will expose infant groups to varying impoverished training input (e.g., including exceptions). The goal is to simulate the natural environment and test the power/limit of induction & possible effects of innate knowledge. Finally, we will use a natural language, testing the hypothesis that innate structural knowledge guides infants to attend to certain relations between functors & lexical items, but not to others, even though the input supports both equally. These new studies address the question of nature versus nurture, one of the most fundamental debates in psychology, biology, neural science & linguistics, one that impacts many areas of research (e.g., speech sci., language disorder/therapy, childhood education, language policy, etc) and our future society.