Assessing the Benefits of Model Ensembles in Neural Re-ranking for Passage Retrieval
Published in European Conference on Information Retrieval (ECIR), 2021
Luís Borges, Bruno Martins, and Jamie Callan
This study uses Fast Geometric Ensembling to produce multiple neural re-rankers and compares ways to combine their outputs, including score averaging, rank fusion, and supervised learning-to-rank. Experiments on MS MARCO show gains in ranking quality from ensembling, with particularly strong results from supervised combination methods.