On the Statistical Significance with Relevance Assessments of Large Language Models

Test collections are an integral part of Information Retrieval (IR) research. They allow researchers to evaluate and compare ranking algorithms in a quick, easy and reproducible way. However, constructing these datasets requires great efforts in manual labelling and logistics, and having only few human relevance judgements can introduce biases in the comparison. Recent research has explored the use of Large Language Models (LLMs) for labelling the relevance of documents for building new retrieval test collections. Their strong text-understanding capabilities and low cost compared to human-made judgements makes them an appealing tool for gathering relevance judgements. Results suggest that LLM-generated labels are promising for IR evaluation in terms of ranking correlation, but nothing is said about the implications in terms of statistical significance. In this work, we look at how LLM-generated judgements preserve the same pairwise significance evaluation as human judgements. Our results show that LLM judgements detect most of the significant differences while maintaining acceptable numbers of false positives. However, we also show that some systems are treated differently under LLM-generated labels, suggesting that evaluation with LLM judgements might not be entirely fair. Our work represents a step forward in the evaluation of statistical testing results provided by LLM judgements. We hope that this will serve as a basis for other researchers to develop reliable models for automatic relevance assessments.