Source: Bruch, Gai & Ingber 2022: An Analysis of Fusion Functions for Hybrid Retrieval
Paper finding that a tuned convex combination of lexical and semantic scores beats RRF, and that RRF is parameter-sensitive.
Contents
- What: research paper (ACM TOIS), arXiv:2210.11934
- Author: Sebastian Bruch, Siyu Gai, Amir Ingber (Pinecone)
- Date: 2022-10-21 (v2 2023-05-04)
- Original: arXiv:2210.11934
Summary
Compares ways of fusing lexical (BM25) and dense retrieval scores, focusing on convex combination of normalized scores versus Reciprocal Rank Fusion.
Key claims
- A convex combination (CC) of lexical and semantic scores outperforms RRF both in-domain and out-of-domain.
- RRF is sensitive to its parameters, contrary to earlier assumptions.
- The choice of score normalization has little effect on learning CC.
- CC has one parameter and can be tuned with a small set of labelled examples from the target domain.