参数化稠密-稀疏融合混合检索:在BEIR SciFact上基于Qdrant调优排名-分数混合
原标题:Parameterized Dense-Sparse Fusion for Hybrid Retrieval: Tuning a Rank-Score Mix on BEIR SciFact with Qdrant
AI 摘要
该论文研究了一种参数化混合检索排序器,将稠密嵌入列表与稀疏词法列表融合,包含稠密先验、分数-排名混合、RRF平滑、列表几何系数和路由边距等显式参数。作者在SciFact训练集(809查询)上网格搜索参数,冻结后在测试集(300查询)上评估,调优后的排名-分数混合达到0.753 nDCG@10和0.889 recall@10,优于稠密BGE(0.742/0.871)和等权RRF(0.707 nDCG@10)。在九个BEIR zip宏平均上,等权RRF未超过稠密(0.479 vs. 0.519 nDCG@10),但在全部20个索引单元上独立重复训练-冻结流程后,该方法在20/20上超过等权RRF,在16/20上超过稠密。作者强调这些系数是数据集特定的,其他语料库应复用搜索范围而非直接复制SciFact的参数点。
正文节选
Parameterized Dense–Sparse Fusion for Hybrid Retrieval: Tuning a Rank–Score Mix on BEIR SciFact with Qdrant Abstract We study a parameterized hybrid ranker that fuses a dense embedding list and a sparse lexical list. The method has a small, explicit parameter vector: a dense prior , a score-versus-rank mix , an RRF smoothing , optional list-geometry coefficients that move per query, and a router margin that can turn sparse search off. We grid-search those ranges on SciFact train (809 queries) an