From Labels to Context: Context-Sensitive User Modeling of Affect with Large Language Models

Abstract

Emotion-aware user modeling is increasingly important in adaptive systems, yet many approaches still represent affect as a fixed label inferred from isolated inputs. This limits their ability to capture how emotional meaning develops across context, time, and interaction history. We address this limitation by modeling affect as a user-level, context-dependent phenomenon. We introduce the context sphere, a structured user representation that aggregates interaction history and preserves conversational and temporal context. Building on this representation, we use a guided two-stage Large Language Model (LLM) pipeline to generate affective interpretations over extended interaction context. We evaluate the approach using automated consistency checks and an exploratory qualitative study with five semi-structured interviews. Results provide initial evidence of plausible, interpretable context-sensitive affective interpretation.