# How poor is the stimulus? Evaluating hierarchical generalization in neural networks trained on child-directed speech

**Authors:** Aditya Yedetore, Tal Linzen, Robert Frank, R. Thomas McCoy  
**arXiv:** [2301.11462](https://arxiv.org/abs/2301.11462)  
**Date:** 2023-01-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2301.11462v1/

## Key points

- Study explores whether children's preference for hierarchical rules in language acquisition is due to a learning bias or more general biases
- Researchers train LSTM and Transformer neural networks without a hierarchical bias on text from the CHILDES corpus
- Models evaluated on how well they learn English yes/no questions, which require hierarchical structure
- Results show models capture surface statistics of child-directed speech accurately but generalize with an incorrect linear rule rather than the correct hierarchical rule
- Stronger biases needed for human-like generalization from text alone compared to standard neural network architectures

## Summary

The study titled "How poor is the stimulus? Evaluating hierarchical generalization in neural networks trained on child-directed speech" explores whether children's preference for hierarchical rules in language acquisition is due to a learning bias or more general biases that interact with hierarchical cues. To investigate this question, the researchers train Long Short-Term Memory (LSTM) and Transformer neural networks without a hierarchical bias on text from the CHILDES corpus, which is similar to children's linguistic input. They then evaluate how well these models learn English yes/no questions, which require hierarchical structure. The results show that while the models capture the surface statistics of child-directed speech accurately, they generalize in a way that aligns more with an incorrect linear rule rather than the correct hierarchical rule. This suggests that stronger biases are needed for human-like generalization from text alone compared to the sequence-processing biases of standard neural network architectures.