SEUPD2425-RACOON at LongEval: A Novel Approach To Information Retrieval With LLM-Based Query Expansion And Temporal Relevance Feedback Techniques

This paper presents the work of the Racoon group from the University of Padua on the LongEval Conference and Labs of the Evaluation Forum (CLEF) 2025 Lab in Task 1 "LongEval-Web Retrieval" [1]. The task was focused on developing Information Retrieval (IR) systems resilient to the temporal evolution of Web documents. To address this challenge, our approach combines traditional IR techniques with several novel components. Specifically, our Search Engine (SE) systems incorporate a French lemmatizer within the analyzer, a Large Language Model (LLM)based Query Expansion (QE) module, a Relevance Feedback (RF) mechanism based on query assessment over the preceding months, and an LLM-based reranking strategy. Overall, our approach, which integrates Relevance Feedback with Deep Learning (DL) reranking, shows significant improvements over the baseline and maintains good short-term temporal robustness on the test set, although its effectiveness declines over the long term.