Generalizable Tip-of-the-Tongue Retrieval with LLM Re-ranking
Published in ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), 2024
Luís Borges, Rohan Jha, Jamie Callan, and Bruno Martins
This work studies how retrieval systems generalize across domains when users describe items they cannot name. It introduces a multi-domain dataset and evaluates first-stage retrieval followed by zero-shot GPT-4 re-ranking using item titles. The experiments show benefits from training across domains and from applying an LLM to retrieved candidates.