Compositional Structure Facilitates Systematic Generalization in Deep Neural Networks and Humans

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.