LLM-Rank: An Unsupervised Keyword Extraction Method Using Local Large Language Models

Large Language Models (LLMs) have shown strong potential in keyword extraction by capturing deep contextual information. However, most existing methods rely on proprietary APIs, raising concerns around cost, privacy, and reproducibility, and often suffer from issues like hallucinated or irrelevant keywords. To address these limitations, we introduce LLM-Rank, a fully unsupervised framework that combines a locally deployed LLM (e.g., Phi-3) with a lightweight post-processing ranking module. Our contributions include: (1) proposing LLM-Rank, a novel framework for local keyword extraction; (2) analyzing the effect of LLM size on extraction quality; (3) evaluating multiple post-processing strategies; and (4) benchmarking performance on four public datasets. Experimental results show that LLM-Rank outperforms state-of-the-art baselines on the NUS dataset, demonstrating the feasibility of accurate, privacy-preserving keyword extraction using local LLMs. Our code has been publicly released on GitHub to support future research.