Overview
When analyzing topics and themes that emerge in collections of conversations, it is first necessary to find “highlights”—semantically or emotionally significant moments where a participant expresses a core narrative, struggle, or opinion—which can act as core units for sensemaking. Manually searching for and tagging these moments is often one of the most labor-intensive bottlenecks in qualitative analysis. In this work, we study how large language models (LLMs) may be leveraged to automatically identify and extract highlights by processing both text and audio-derived affective information. Based on these insights, we introduce an interactive highlighting tool designed to support qualitative researchers in real-world workflows in a flexible, interpretable way.
