Human and Machine Inductive Biases for Compositional Linguistic Generalization

Explore the intricacies of compositional linguistic generalization in both humans and artificial neural networks (ANNs) through this insightful talk by Najoung Kim from Boston University. Delve into the concept of compositionality as a fundamental property of human language and its role in enabling generalization. Examine several semantic parsing tests designed to evaluate compositional linguistic generalization in ANNs, comparing the results to human generalization patterns. Discover how models perform in cases of lexical substitution versus structurally novel generalizations. Gain insights into the challenges of testing generalization in the current modeling landscape, particularly without open access to training data. Consider the opportunities for better understanding structural generalization in humans and its implications for artificial intelligence research.