Paper Summary: Inference Time Re-ranker Relevance Feedback for Neural Information Retrieval

Today, I'm discussing the Inference Time Re-ranker Relevance Feedback for Neural Information Retrieval. This video explores a solution provided by Allen AI, UIUC, and UW to improve the traditional bi encoder relevance relation. The approach involves re-ranking at inference time instead of a post bi encoder. I explain the traditional systems and the limitations of using a cross encoder model. Then, I dive into the team's approach of updating the query encoding with language distilled information from the cross encoder. The video also highlights the team's results, which consistently outperform other methods, including BM25 and Colbert. Paper link: https://arxiv.org/abs/2305.11744