Core Concepts
LISTT5, a listwise reranking model based on Fusion-in-Decoder, outperforms existing models in zero-shot retrieval tasks.
Abstract
LISTT5 is a novel reranking approach that handles multiple candidate passages efficiently. It improves efficiency and performance in zero-shot retrieval tasks compared to state-of-the-art models. The model overcomes limitations of previous listwise rerankers and demonstrates robustness to positional bias. By leveraging the Fusion-in-Decoder architecture, LISTT5 provides significant gains in NDCG@10 scores and computational efficiency.
Stats
LISTT5 (1) outperforms RankT5 baseline with +1.3 gain in NDCG@10 score.
LISTT5 has an efficiency comparable to pointwise ranking models and surpasses the efficiency of previous listwise ranking models.
Quotes
"LISTT5 provides significant contributions in computational efficiency, robustness to positional bias, and zero-shot performance."
"Efficiency improvements are demonstrated through FLOPs analysis, showcasing the advantages of LISTT5 over existing models."