Empirical Analysis of Score Fusion Strategies under Pool-Restricted Dense Encoding for Ad Hoc Document Retrieval

Abstract

This paper examines whether restricting dense encoding to the BM25 candidate pool top 100 per query is a viable substitute for full corpus embedding and what score fusion can extract from that restricted budget We implement BM25100 as the first stage on the MS MARCO document development collection 5193 topics train a passage level BERT bi encoder with 24 dimensional compression and evaluate four pool level fusion strategies semantic only reranking linear blending rank stratified blending and reciprocal rank fusion Linear blending at α 085 BM25 heavy yields nDCG10 of 03197 versus 03155 for BM25 alone p 0012 paired bootstrap semantic only reranking falls to 00904 71 relative the remaining variants trail BM25 by 8 11 Under pool restriction reducing the lexical weight consistently degrades effectiveness the passage trained encoder does not compensate for the lost BM25 signal at document granularity Cross encoder reranking and learned feature stacking are under ongoing investigation