Can Query Expansion Improve Generalization of Strong Cross-Encoder Rankers?
Topics: AI (Deep Learning), Search Query Processing
The paper titled “Can Query Expansion Improve Generalization of Strong Cross-Encoder Rankers?” explores the impact of query expansion on the performance of second-stage, cross-encoder rankers. The study finds that while traditional query expansion methods benefit weaker models, they often harm stronger rankers. The authors propose a new approach that involves high-quality keyword generation and minimally disruptive query modification, which improves the generalization of strong rankers like RankT5 and MonoT5.
Inventors:
- Minghan Li (University of Waterloo)
- Honglei Zhuang (Google)
- Kai Hui (Google)
- Zhen Qin (Google)
- Jimmy Lin (University of Waterloo)
- Rolf Jagerman (Google)
- Xuanhui Wang (Google)
- Michael Bendersky (Google)