Reproduce “HERCULES: Hierarchical Embedding-based Recursive Clustering Using LLMs for Efficient Summarization” on OpenResearch

About this paper

Gabor Petnehazi, Bernadett Aradi

HERCULES introduces a hierarchical clustering framework that integrates Large Language Models to generate human-understandable titles and descriptions for clusters at every level of the hierarchy, supporting text, image, and numeric data. It enhances the interpretability of complex datasets and provides an interactive visualization tool for exploring multi-level data structures.

Key insights

• Systematic integration of LLMs within a recursive clustering framework can effectively translate abstract data groupings into human-understandable titles and descriptions at multiple hierarchical levels.
• Clustering directly on original data embeddings ('direct' representation mode) is more effective for building robust hierarchical structures that enable high-quality, semantically distinct LLM-generated interpretations.
• The quality and speed of the underlying LLM model directly influence the semantic richness and distinctiveness of the generated cluster summaries, impacting overall interpretability.

Summary by alphaXiv.