Utilizing Pseudo-Relevance Feedback in Fusion-based Retrieval

The usage of positive relevance feedback in fusion-based retrieval was previously shown to be very useful. Yet, in many retrieval use-cases, no actual relevance feedback may be available. With the absence of relevance data, pseudo-relevance feedback models have been suggested as an alternative. Encouraged by the previous success of using positive relevance feedback in fusion-based retrieval, in this work, we study the usage of pseudo-relevance feedback in this setting as well. We build on top of an existing approach that was originally designed for utilizing positive relevance feedback and adapt it to pseudo-relevance feedback. To this end, we propose a novel approach for estimating document (pseudo) relevance labels. Our labeling approach is better tailored to the fusion-based retrieval setting and provides favorable retrieval quality results.