Interactions Between Syntax and Semantics in Language Models

Abstract

How do syntax and semantics interact in neural language models? Is the intuitive dichotomy between semantics and syntax useful as a mental model of their behavior? In this thesis, I systematically investigate how models handle the interactions between sentence corruptions of each kind. I develop a setup for generating “syntactic” and “semantic” corruptions and combining them. Using this setup, I develop a general framework for understanding how classes of corruptions interact with each other, arguing that this can be understood in terms of a single “entanglement” parameter. This allows me to quantify entanglement between syntax and semantics in a variety of situations. The result is that semantics is more entangled with syntax than it is with itself, and that syntax is highly self-entangled.