New paper to appear at EACL 2026 main conference, and it's now up on arxiv: arxiv.org/pdf/2505.17536. The character limit here is insane, so I'll let the screenshots speak for themselves. We put together a new dataset for conversational role attribution & thread disentanglement.
Mark your calendars for our second Cultural Analytics Talk Series lecture w/ @ucbids.bsky.social!
@ischool.uw.edu Prof. @mellymeldubs.bsky.social will discuss how authors and readers are using #AI to generate fiction writing.
📅 Oct. 24, 12:15 - 1:30 pm
📍 210 South Hall, Online
Beautifully put. I share much of this vision of what information science as a discipline has been and should be going forward.
Some movies make you feel seen and forgiven. First thought after I managed to pick myself up from the triumphant Project Hail Mary. Haven't felt this way since Everything Everywhere All at Once. I find social media exhausting, but I was compelled to come back and share this with you.
Never thought this day would come. We're officially past the midpoint of the semester, and the students who are brave enough to stay in my sometimes crazy—but always radical—interdisciplinary experiment are now submitting their final project proposals. We've covered a lot of ground in just 7 weeks!
To ensure wider reach: our Cultural Analytics class will meet in a larger classroom, 210 South Hall, on 9/2. (And no, pun not intended.)
I believe in two things, perhaps now even more firmly than ever: a.) SFT is still the cheapest way to get what you want; b.) subtle forms of memorization, and the inherent, inevitable tension between the model's parametric knowledge and the conditioning context is worthy of further study.
The unique affordance of this work is conversational role attribution, and the pattern there is also quite clear: we see women are significantly more typically cast in listening roles (addressees or side-participants) than as speakers.
I am amused to see that floor-claiming patterns represented in the four long-running TV series here are nearly identical to the 800+ movies we studied back in our 2023 paper: men speak significantly more than women, but women start new conversations at a rate proportional to their speaking time.
Gemini is currently miles ahead of the competition in processing integrated audio-visual signals, and conversation structure understanding is not just a text problem. At the same time, if we use inter-annotator agreement as a proxy for the human ceiling, we still have a long way to go.
