Deep Learning and Large Language Models in Linguistic Research

Catalog description

The goal of this course is to give students the concepts and skills they need to apply methods from Deep Learning and Large Language Models in research on speech, language, and communication. We will survey the interesting past, the exciting present, and the promising but uncertain future of these technologies, focusing on their limitations as well as their capabilities. A key idea is moving from local to contextual features, "situated" relative to patterns learned from large bodies of training material. Instead of words in isolation, we look at patterns of words in text; instead of 10-to-30-msec audio windows in isolation, we look at patterns of signal-derived features in phrase-sized contexts. Used as inputs to a wide variety of prediction and classification systems, these approaches lead to large improvements in performance. We will also survey the wide range of architectures and training methods, including the spectrum from supervised to lightly-supervised, self-supervised, and unsupervised methods; and also the integration of these systems with other old and new structures from physics, mathematics, and linguistics. We will explore the relevance of the learning models to the central issues in the science of language and how language is situated in the human cognitive system. The details of coverage will depend in part on the interests of participants. No specific prerequisites are required, though obviously participants will need at least basic programming skills.