UNHIDE: UNsupervised Human-Interpretable Dialogue Exploration

ACL ARR 2025 February Submission8099 Authors

Abstract: Organizations increasingly rely on human or virtual agents for customer service, generating vast amounts of dialogue data. Interpreting this data is essential for improving communication and customer satisfaction but remains challenging due to its volume, complexity, and nuances. We present UnHIDE, an unsupervised framework that supports human interpretation of dialogue data by discovering dialogue flows and computing interpretable metrics. These reveal insights into common paths, complexity, sentiment progression, and response speed. The potential of UnHIDE is showcased in dialogues generated with the previous variables, which can be observed in the produced analysis, demonstrating UnHIDE's potential for uncovering meaningful dialogue structures and improving conversational strategies.

Paper Type: Short

Research Area: Dialogue and Interactive Systems

Research Area Keywords: evaluation and metrics; task-oriented;applications;automatic evaluation;analysis;dialogue;flow;

Contribution Types: Data resources, Data analysis

Languages Studied: English

Submission Number: 8099