LLM-as-a-Judge: automated evaluation of search query parsing using large language models.

Abstract

Introduction: The adoption of Large Language Models (LLMs) in search systems necessitates new evaluation methodologies beyond traditional rule-based or manual approaches.

Methods: We propose a general framework for evaluating structured outputs using LLMs, focusing on search query parsing within an online classified platform. Our approach leverages LLMs' contextual reasoning capabilities through three evaluation methodologies: Pointwise, Pairwise, and Pass/Fail assessments. Additionally, we introduce a Contextual Evaluation Prompt Routing strategy to improve reliability and reduce hallucinations.

Results: Experiments conducted on both small- and large-scale datasets demonstrate that LLM-based evaluation achieves approximately 90% agreement with human judgments.

Discussion: These results validate LLM-driven evaluation as a scalable, interpretable, and effective alternative to traditional evaluation methods, providing robust query parsing for real-world search systems.