Abstract: NLP researchers have typically treated text summarization as a problem with a singular solution: we need to define what an ideal summary looks like, collect many examples of such summaries for our domain of interest, and train a neural network model on this data. But for real-world applications, different users are interested in different aspects of information, sometimes only having a broad intent or vague idea of what they're seeking. In this talk, I will describe our work on aspect-oriented summarization, which aims to summarize documents given a collection keywords (under)specifying a user's intent. What the user is looking for may not even be present in their data. Building a system that can handle this setting reliably is hard, however, and we observe that popular approaches often fail to produce factual summaries. I will further describe a thrust of our work attempting to ensure the factuality of generated text in a fine-grained way, as well as an ongoing project analyzing how we should decide when these models should return empty inputs or ignore parts of the user's request.

Speaker Bio: Greg Durrett is an assistant professor of Computer Science at UT Austin. His current research focuses on making natural language processing systems more interpretable, controllable, and generalizable, spanning application domains including question answering, textual reasoning, summarization, and information extraction. His work is funded by a 2022 NSF CAREER award and other grants from agencies including the NSF, DARPA, Salesforce, Amazon, and Walmart. He completed his Ph.D. at UC Berkeley in 2016, where he was advised by Dan Klein, and from 2016-2017, he was a research scientist at Semantic Machines.