Live Commentary Planning and Generation

Live commentary plays a crucial role in helping audiences interpret high-stakes events such as political debates, central bank press conferences, and corporate earnings calls. Unlike generic summarization, professional commentary requires timely decisions about what to comment on and how to present it, integrating fact-checking, background knowledge, and subjective evaluation. However, little prior work has studied commentary as a structured planning and generation problem. To bridge this gap, we introduce the first multi-domain dataset of Live Commentary Planning and Generation, aligning event transcripts with time synchronized expert analyses and public reactions. Our dataset covers U.S. presidential debates (2016–2024), Federal Open Market Committee press conferences, and corporate earnings calls, enriched with a fine-grained taxonomy of commentary intents (up to 11 categories) and supplemented by Reddit crowd commentary. We define two benchmark tasks: (1) Commentary Planning, predicting the type of commentary given a transcript segment, and (2) Commentary Generation, producing commentary text conditioned on the segment and a target label. Baseline experiments with large language models show that, despite their fluency, models struggle with expert- level commentary, showing the difficulty of integrating contextual reasoning and external knowledge under real-time constraints.