Topics in Computational Linguistics: Bridging Symbols, Vectors, and Probabilities

Undergraduate

This course focuses on potential ways to reconcile three types of computational models: symbolic systems, neural networks, and Bayesian models. These three types of models are the most popular approaches in computational linguistics (and computational cognitive science more broadly). Each one captures important properties of the human mind: symbolic systems capture how the mind can combine familiar components in novel ways, neural networks capture how the mind can handle complex, naturalistic data, and Bayesian models capture how the mind can make rich inferences from incomplete evidence. Despite their independent successes, these types of models appear qualitatively different from each other. In this class we will study proposals for how these modeling traditions might be unified to realize their complementary strengths. Proposals that we will discuss include meta-learning, neurosymbolic architectures, tensor product representations, and Bayesian interpretations of large language models. Prerequisites: LING 2270 or other coursework in neural networks and Bayesian models; Python programming experience QR , SO T 4pm-5:55pm