Core Concepts
It is possible to improve the generalization of strong neural rankers, such as RankT5 and MonoT5, by leveraging prompt engineering and aggregating the ranking results of each expanded query via fusion.
Abstract
The paper examines the impact of query expansion on the generalization of strong cross-encoder rankers, which has been an under-explored area. The authors first apply popular query expansion methods, such as RM3 and generative techniques, to state-of-the-art cross-encoder rankers and observe a deterioration in their zero-shot performance.
The authors identify two key steps for successful query expansion with cross-encoders: high-quality keyword generation and minimal-disruptive query modification. They propose a framework that leverages an instruction-following language model to generate high-quality keywords through a reasoning chain, and then combines the ranking results of each expanded query dynamically using self-consistency and reciprocal rank weighting.
Experiments on the BEIR and TREC Deep Learning 2019/2020 benchmarks show that this approach can improve the nDCG@10 scores of both MonoT5 and RankT5, pointing to a promising direction for applying query expansion to strong cross-encoder rankers.
Stats
The nDCG@10 score of RankT5 on TREC DL 2019 improved from 0.737 to 0.751 using the proposed method.
The nDCG@10 score of MonoT5 on TREC DL 2019 improved from 0.695 to 0.724 using the proposed method.
The nDCG@10 score of RankT5 on the BEIR Wiki+News dataset improved from 0.557 to 0.570 using the proposed method.
The nDCG@10 score of MonoT5 on the BEIR Wiki+News dataset improved from 0.519 to 0.532 using the proposed method.
Quotes
"Can Query Expansion Improve Generalization of Strong Cross-Encoder Rankers?"
"It is possible to improve the generalization of a strong neural ranker, by prompt engineering and aggregating the ranking results of each expanded query via fusion."