More compositional and structured linguistic input leads to better memorization, more systematic generalization, and greater similarity to human learning in both deep neural networks and large language models.
