Fusion-in-T5: Unifying Document Ranking Signals for Improved Information Retrieval

TL;DR

Fusion-in-T5 (FiT5) is a unified re-ranking model that combines multiple ranking signals into one system, significantly enhancing passage retrieval performance on benchmark datasets by effectively utilizing attention mechanisms.

Contribution

The paper introduces FiT5, a novel unified re-ranking model that integrates diverse ranking signals using global attention, outperforming traditional cascade ranking pipelines.

Findings

FiT5 significantly improves passage ranking performance.

Attention fusion enables joint utilization of multiple ranking signals.

FiT5 detects subtle ranking nuances more effectively.

Abstract

Common document ranking pipelines in search systems are cascade systems that involve multiple ranking layers to integrate different information step-by-step. In this paper, we propose a novel re-ranker Fusion-in-T5 (FiT5), which integrates text matching information, ranking features, and global document information into one single unified model via templated-based input and global attention. Experiments on passage ranking benchmarks MS MARCO and TREC DL show that FiT5, as one single model, significantly improves ranking performance over complex cascade pipelines. Analysis finds that through attention fusion, FiT5 jointly utilizes various forms of ranking information via gradually attending to related documents and ranking features, and improves the detection of subtle nuances. Our code is open-sourced at https://github.com/OpenMatch/FiT5.