Hybrid Retrieval and Multi-stage Text Ranking Solution at TREC 2022 Deep Learning Track

TL;DR

This paper describes a hybrid retrieval and multi-stage ranking system for large-scale text retrieval, combining traditional and neural methods, achieving top ranks in TREC 2022 evaluations.

Contribution

The paper introduces a novel hybrid retrieval approach and a multi-stage ranking framework with a lightweight sub-ranking module, improving text ranking performance.

Findings

Achieved 1st place in passage ranking at TREC 2022.

Achieved 4th place in document ranking at TREC 2022.

Demonstrated effectiveness of combined sparse and dense retrieval methods.

Abstract

Large-scale text retrieval technology has been widely used in various practical business scenarios. This paper presents our systems for the TREC 2022 Deep Learning Track. We explain the hybrid text retrieval and multi-stage text ranking method adopted in our solution. The retrieval stage combined the two structures of traditional sparse retrieval and neural dense retrieval. In the ranking stage, in addition to the full interaction-based ranking model built on large pre-trained language model, we also proposes a lightweight sub-ranking module to further enhance the final text ranking performance. Evaluation results demonstrate the effectiveness of our proposed approach. Our models achieve the 1st and 4th rank on the test set of passage ranking and document ranking respectively.