Rethinking Discourse Processing in the Era of Large Language Models

Representations of discourse structure were once known to be useful for many downstream tasks, from classification tasks like sentiment analysis to generation tasks like summarization. In today's world, however, LLMs are capable of generating coherent text with excellent structure, and effectively "solve" many long-standing (generation) tasks in NLP. So what challenges remain? In this talk, I reflect on this question from two angles. First, can we leverage the new generative capabilities that we have, and rethink what discourse analysis could look like? We show that LLMs, with design and training, enable the analysis of discourse via Questions Under Discussion, resurfacing curiosity-driven questions and grounding their elicitation and answers in text. Second, we leverage the implicit questions to discover new applications that are more user-oriented and challenging, especially in the domain of text simplification: new tools to generate explanations and address information loss.

Short bio: Junyi Jessy Li is an Associate Professor in the Linguistics Department at the University of Texas at Austin. She earned her Ph.D. (2017) in Computational Linguistics from the Department of Computer and Information Science at the University of Pennsylvania. Her research interests are in computational linguistics and natural language processing, specifically discourse and pragmatics, natural language generation, and computational social science. She received an NSF CAREER Award, a Google Faculty Research Award, an ACL Outstanding Paper Award, and an ACM SIGSOFT Distinguished Paper Award, among various other honors. Jessy is the Secretary of the North American Chapter of the Association for Computational Linguistics (NAACL).