Séminaire de Recherche en Linguistique

Speakers of a language generalize their knowledge of syntax in a systematic way to constructions they have never encountered before. This observation has motivated the influential position in linguistics that humans are innately endowed with syntax-specific inductive biases. The applied success of deep learning systems that are not designed with such biases invites a reconsideration of this position. In this talk, I will review work that uses paradigms from psycholinguistics to examine the syntactic generalization capabilities of contemporary neural network architectures. Alongside some successes, this work suggests that human-like generalization requires stronger inductive biases than those expressed in standard neural network architectures.