Pseudo Relevance Feedback with Deep Language Models and Dense Retrievers: Successes and Pitfalls

Pseudo Relevance Feedback (PRF) is known to improve the effectiveness of bag-of-words retrievers. At same time, deep language models have been shown outperform traditional rerankers. However, it unclear how integrate PRF directly with emergent models. This article addresses this gap by investigating methods for integrating signals rerankers and dense retrievers based on We consider text-based, vector-based hybrid approaches investigate different ways combining scoring relevance signals. An extensive empirical evaluation was conducted across four datasets two task settings (retrieval ranking). Text-based results show that use had a mixed effect datasets. found best achieved when (i) concatenating each passage query, searching new set queries, then aggregating scores; (ii) using Borda aggregate scores from runs. Vector-based enhanced over several metrics. higher query retains either majority or weight within mechanism, shallower signal (i.e., smaller number top-ranked passages) employed, rather than deeper signal. Our method computationally efficient; thus, represents general others can