Improving Query Representations for Dense Retrieval with Pseudo Relevance Feedback: A Reproducibility Study

Pseudo-Relevance Feedback (PRF) utilises the relevance signals from top-k passages first round of retrieval to perform a second aiming improve search effectiveness. A recent research direction has been study and development PRF methods for deep language model based rankers, in particular context dense retrievers. Dense retrievers provide trade off between effectiveness, which is often reduced compared more complex neural query latency, also making pipeline efficient. The introduction motivated as an attempt further their In this paper, we reproduce method with retrievers, called ANCE-PRF. This concatenates text that feedback form new input, then encoded into representation using newly trained encoder on original retriever used retrieval. While can potentially be applied any existing prior work studied it only ANCE retriever. We reproducibility ANCE-PRF terms both its training (encoding signal) inference (ranking) steps. extend empirical analysis provided investigate effect hyper-parameters govern process robustness across these different settings. Finally, contribute generalisability when other than are encoding signal.