Can Query Expansion Improve Generalization of Strong Cross-Encoder Rankers?

Query expansion has been widely used to improve the search results offirst-stage retrievers, yet its influence on second-stage, cross-encoderrankers remains under-explored. A recent work of Weller et al. [44] shows thatcurrent expansion techniques benefit weaker models such as DPR and BM25 butharm stronger rankers such as MonoT5. In this paper, we re-examine thisconclusion and raise the following question: Can query expansion improvegeneralization of strong cross-encoder rankers? To answer this question, wefirst apply popular query expansion methods to state-of-the-art cross-encoderrankers and verify the deteriorated zero-shot performance. We identify twovital steps for cross-encoders in the experiment: high-quality keywordgeneration and minimal-disruptive query modification. We show that it ispossible to improve the generalization of a strong neural ranker, by promptengineering and aggregating the ranking results of each expanded query viafusion. Specifically, we first call an instruction-following language model togenerate keywords through a reasoning chain. Leveraging self-consistency andreciprocal rank weighting, we further combine the ranking results of eachexpanded query dynamically. Experiments on BEIR and TREC Deep Learning2019/2020 show that the nDCG@10 scores of both MonoT5 and RankT5 followingthese steps are improved, which points out a direction for applying queryexpansion to strong cross-encoder rankers.