User Interaction-Driven Refinement of Dense Retrieval Models

After recognizing their need for information, users formulate search queries that reflect their intent.These queries are then submitted to a search engine with the expectation that the results will be relevant and address their information needs.Users rely on the returned results aligning with their intent and context.This highlights the increasing need for advanced retrieval models that can adapt not only to the user's query but also to their broader informational requirements.While significant progress has been made in improving dense retrieval systems to prioritize the most relevant results, few efforts have focused on integrating user history.In this research, we aim to present an approach that incorporates user interaction history, such as clicks and query reformulations, into dense retrievers to achieve more accurate rankings.To measure progress toward this goal, we will utilize the Trip Click benchmark, created from approximately 4M click log entries in the context of a health web search engine.Our objective is to develop a model that enhances dense retrieval results by leveraging user action history, thereby increasing the likelihood that users will find relevant documents more quickly and end their search sessions with greater satisfaction.