Predicting Age of Acquisition in Early Word Learning Using Recurrent Neural Networks

AbstractVocabulary growth and syntactic development are known to be highly correlated in early child language. What determines when words are acquired and how can this help us understand what drives early language development? We train an LSTM language model, known to detect syntactic regularities that are relevant for predicting the difficulty of words, on child-directed speech. We use the average surprisal of words for the model, which encodes sequential predictability, as a predictor for the age of acquisition of words in early child language. We compare this predictor to word frequency and others and find that average surprisal is a good predictor for the age of acquisition of function words and predicates beyond frequency, but not for nouns. Our approach provides insight into what makes a good model of early word learning, especially for words whose meanings rely heavily on linguistic context.


Return to previous page