Item-Based Language Learning in Children and Connectionist Networks

Much of children's early language is grammatical. This fact has been used as a support for the idea that babies are born with abstract grammatical categories and rules. However, some recent investigations suggest that the grammaticality of early utterances can be attributed largely to imitation of adults. That is, children's early language is item-based or context-bound and based on specific linguistic items, expressions, or phrases they comprehend and produce, often in specific semantic contexts. Here, a connectionist network is presented that exhibits similar behavior while performing a comprehension task. When trained on a small number of words in a few contexts, the network exhibits poor generalization and signs of item-based representations. With increasing vocabulary, the network's generalization improves dramatically, indicating a move to more context-free, categorical representations. This suggests a simple account of the item-based nature of early language, and calls into question the need to propose elaborate innate apparatus. Some evidence that children's input also becomes gradually more complex, and this helps them form better representations, is also reviewed