Embedded Files
Fri, July 3, 2026 – Time: 09:10 – 10:00 – Room: Regatta
Title: Discourse Structure Guided NLP Models for Fine-grained Media Bias Analysis
Abstract: Thinking about more and more powerful pretrained models and an increasing context window size supported by recent LLMs, do we still need explicit discourse structures to guide semantic reasoning? In this talk, I will present our research on fine-grained sentence-level media bias analysis that shows that incorporating shallow discourse structures or event relation graphs enables NLP models to better understand broader context and recognize subtle sentence-level ideological bias. News media play a vast role in shaping public opinion not just by supplying information, but by selecting, packaging, and shaping that information to persuade as well. Sentence-level media bias analysis is challenging and aims to identify sentences within an article that can illuminate and explain the overall bias of the entire article. This talk will first show that understanding the discourse role of a sentence in telling a news story, as well as its discourse relation with nearby sentences, can help reveal the ideological leanings of the author even when the sentence itself appears merely neutral or factual. This talk will further show that analyzing events with respect to other events in the same document or across documents is critical for identifying bias sentences.
Bio: Dr. Ruihong Huang is an associate professor in the Department of Computer Science & Engineering at Texas A&M University (TAMU), College Station. She is also an adjunct associate professor in McWilliams School of Biomedical Informatics at UTHealth Houston. Huang received her PhD in computer science at the University of Utah and completed a postdoc at Stanford University. Her research is focused on event-centric NLP, discourse analysis, dialogue and pragmatics, LLM evaluation, LLM safety and moral reasoning. She is a recipient of the US National Science Foundation CAREER award (2020).
Fri, July 3, 2026 – Time: 13:30 – 14:20 – Room: Regatta
Title: Studying and Simulating Discourse Failures in Interactive AI Systems
Abstract: In this talk, we will go over two projects that study human-AI communication in conversational interfaces (e.g., ChatGPT). With Lost in Conversation, we simulate multi-turn underspecified conversations that involve a user gradually revealing their full intent over time to the AI system, finding that though this is common in real-world interactions, this leads to large degradation in model performance. We then discuss the need for more realistic user simulation that can model user behaviors specifically in the case of User-AI conversational exchange, and discuss its role in evaluating AI systems.
Bio: Philippe is a Research Scientist at Microsoft Research, based in New York and a member of the AI Interaction and Learning (MSR AIIL) group. His research lies at the intersection of two fields: NLP and HCI, with a general interest in understanding how NLP/AI technology can enable effective consumption and production of information for users (which has involved working on topics such as summarization, simplification, question generation, etc.). Recently, he has taken an interest in studying human-AI interaction in conversational settings.
Page updated
Google Sites
Report abuse