How poor is the stimulus? Evaluating hierarchical generalization in neural networks trained on child-directed speech
This addresses the problem of understanding language acquisition biases for linguists and AI researchers, showing that standard neural networks fail to replicate human-like generalization from limited data.
The study trained LSTMs and Transformers on child-directed speech data to test if they learn hierarchical rules for English yes/no questions, finding that both models generalized incorrectly using linear rules instead, despite good perplexity scores.
When acquiring syntax, children consistently choose hierarchical rules over competing non-hierarchical possibilities. Is this preference due to a learning bias for hierarchical structure, or due to more general biases that interact with hierarchical cues in children's linguistic input? We explore these possibilities by training LSTMs and Transformers - two types of neural networks without a hierarchical bias - on data similar in quantity and content to children's linguistic input: text from the CHILDES corpus. We then evaluate what these models have learned about English yes/no questions, a phenomenon for which hierarchical structure is crucial. We find that, though they perform well at capturing the surface statistics of child-directed speech (as measured by perplexity), both model types generalize in a way more consistent with an incorrect linear rule than the correct hierarchical rule. These results suggest that human-like generalization from text alone requires stronger biases than the general sequence-processing biases of standard neural network architectures.