CALLIOPE is a top‑down framework for discovering topic hierarchies in massive unstructured corpora such as LLM conversations, while providing differential‑privacy guarantees throughout the pipeline. Conventional and recent LLM‑augmented approaches first embed texts into vectors and then cluster them; CALLIOPE instead leverages the LLM’s reasoning abilities directly for topic assignment, hierarchy construction, and privacy noise injection, eliminating the need for vector embeddings.
Experiments show that under differential‑privacy constraints, especially in the high‑privacy regime with a small privacy budget, CALLIOPE achieves noticeably higher utility than embedding‑based baselines. We assess output quality with standard clustering metrics (e.g., silhouette score, NMI) and several novel utility metrics introduced in this work.
Review