talks
invited talks, seminars, and presentations
-
In this talk, I present ConGraT, a self‑supervised approach for aligning text and graph representations on text‑attributed graphs. I cover the core contrastive objective that brings together a language model and a graph neural network, show results on node/text classification, link prediction, and language modeling, and discuss how the method surfaces textually grounded communities in social networks.
-
This talk surveys how authenticity, consent, privacy, and documentation break down in today’s foundation‑model training pipelines. I highlight evidence from recent audits, outline practical tooling and standards, and share a roadmap for researchers, developers, and policymakers to make training data more transparent and trustworthy.
-
I present an empirical comparison of how 1,600+ news events spread across Twitter (X) and U.S. talk radio. The talk covers datasets, methods, and key findings—Twitter moves faster, decays more quickly, and is more negative and outraged—plus what this means for studying media ecosystems and agenda setting.
-
A demo‑driven talk introducing AudienceView, an LLM‑powered assistant for interpreting large volumes of audience comments. I walk through clustering, linking themes to concrete examples, visualizing sentiment and distribution, and how the tool fits newsroom workflows—along with limitations and lessons from user interviews.
-
I share what large‑scale analyses of professional dubbing reveal about real production priorities: vocal naturalness and translation quality often matter more than strict lip‑sync. The talk connects these insights to automatic dubbing, evaluation, and model design under real‑world constraints.