A Setwise Approach for Effective and Highly Efficient Zero-shot Ranking with Large Language Models
Topics: AI (Deep Learning), Ranking
The document introduces a novel zero-shot document ranking method using large language models (LLMs) called the Setwise prompting approach. This method aims to improve the efficiency and effectiveness of LLM-based zero-shot ranking by comparing multiple documents at each step rather than individual pairs, which reduces computational overhead. The approach is shown to enhance efficiency without compromising effectiveness compared to existing Pointwise, Pairwise, and Listwise methods. The study also provides a comprehensive evaluation framework for LLM-based zero-shot ranking methods.
Inventors: Shengyao Zhuang, Honglei Zhuang, Bevan Koopman, Guido Zuccon (all from Google)