Does LLM Relevance Labelling Work for Arabic?
TL;DR
It is shown that agreement on binary labels and system ordering is generally high, with no cases of significant opposite conclusions; however, there are cases of false or missed improvements and noticeable limitations when labelling manual, highly performing systems.
Abstract
Large Language Models (LLMs) are increasingly used in Information Retrieval, both within retrieval pipelines and for constructing evaluation resources. Existing studies on using LLMs for IR evaluation, however, focus almost exclusively on English, leaving their applicability to other languages, where evaluation resources are often limited and highly needed, unexplored. We examine the use of LLMs to generate relevance labels for an Arabic test collection (ArTest). Using about 10K relevance labels, generated by three LLMs, and used to order eight automated systems and nine simulated manual systems, we show that agreement on binary labels and system ordering is generally high, with no cases of significant opposite conclusions; however, there are cases of false or missed improvements and noticeable limitations when labelling manual, highly performing systems. These findings align with results reported for English and indicate that LLMs could, with some caveats, offer a viable approach to supporting IR evaluation in languages with limited evaluation resources.