Characterizing Harmful Inputs in Neural Retrieval: From Query Tokens to Relevant Feedback Documents

Abstract: The research includes two projects. In the first project, we explore whether removing specific tokens from a user query can improve retrieval effectiveness in neural retrieval systems. We begin by evaluating how token removal affects performance, and then, based on our findings, we identify which tokens to keep or remove. In the second project, we focus on relevance-feedback-based retrieval, in which users provide a document they deem relevant to the system, which then leverages it to enhance retrieval performance. Previous research has shown that some relevant documents can degrade performance rather than improve it; these documents are referred to as "poison pills." While poison pills have been studied in classical retrieval settings, our research extends this analysis to neural retrieval systems.