Fine Tuning Query Representation using Iterative Negative Feedback

Authors: Mudit Chaudhary, Dhawal Gupta
Venue: Course Research Project for COMPSCI 646 Information Retrieval at UMass Amherst (Prof. Hamed Zamani)
Abstract: Dense retrieval methods such as DPR are becoming increasingly popular due to their superior performance over traditional retrieval methods. Furthermore, dense retrieval methods are able to model the semantic meaning of the query and the documents. One of the ways to improve the retrieval performance of these methods is to use a re-ranker. Often, the re-rankers (e.g., cross-encoders) are computationally expensive and impractical for fast re-ranking on a large list of documents. Alternative ways to improve the performance is to use query representation refinement methods to re-rank the documents. Majority of the research focuses on using positive relevance feedback for improving the dense query representation. Work on using negative relevance feedback for dense query refinement is severely limited. In this paper, we propose a simple and computationally inexpensive model for query refinement using negative relevance feedback to re-rank documents. We perform experiments on the MS-Marco passage re-ranking task and aim to refine dense query representations from DPR.
Report Link: Link