SAGE: Segmenting and Grouping Data Effectively using Large Language Models

Abstract

Grouping is a technique used to organize data into manageable pieces, reducing cognitive load and enabling users to focus on discovering higher-level insights and generating new questions. However, creating groups remains a challenge, often requiring users to have prior domain knowledge or an understanding of the underlying structure of the data. We introduce SAGE, a novel technique that leverages the knowledge base and pattern recognition abilities of large language models (LLMs) to segment and group data with domainawareness. We instantiate our technique through two structures: bins and highlights; bins are contiguous, non-overlapping ranges that segment a single field into groups; highlights are multi-field intersections of ranges that surface broader groups in the data. We integrate these structures into Olli, an open-source tool that converts data visualizations into accessible, keyboard-navigable textual formats to facilitate a study with 15 blind and low-vision (BLV) participants, recognizing them as experts in assessing agency. Through this study, we evaluate how SAGE impacts a user’s interpretation of data and visualizations, and find our technique provides a rich contextual framework for users to independently scaffold their initial sensemaking process.