Predicting Relevant Conversation Turns for Improved Retrieval in Multi-Turn Conversational Search

This technical report presents the work of Università della Svizzera italiana in TREC CAsT 2019. TREC CAsT was set up to advance research on conversational search systems. The goal of the track is to create a reusable benchmark for open-domain information-centric conversational dialogues and to establish a concrete and standard collection of data with information needs to make systems directly comparable. Given the complexity of natural language and the evolution of user’s information need in a conversation with multiple turns, finding relevant context is not always straightforward. We developed a neural model for identifying relevant turn(s) corresponding to the given turn. Our model reformulates the information need of the user to take into account the conversational context to enhance the ad-hoc passage retrieval performance. Two of our runs also employ neural re-ranking of the passages post-retrieval. One of our runs was able to achieve above-median performance.