Different rankers on different subcollections
T Jones, F Scholer, A Turpin, S Mizzaro, M Sanderson, A Hanbury, G Kazai, A Rauber, N Fuhr
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) | Springer International Publishing | Published : 2015
Recent work has shown that when documents in a TREC ad hoc collection are partitioned, different rankers will perform optimally on different partitions. This result suggests that choosing different highly effective rankers for each partition and merging the results, should be able to improve overall effectiveness. Analyzing results from a novel oracle merge process, we demonstrate that this is not the case: selecting the best performing ranker on each subcollection is very unlikely to outperform just using a single best ranker across the whole collection.