IRnator: A Framework for Discovering Users Needs from Sets of Suggestions


To tackle complex IR tasks, where users cannot precisely define their needs, interaction is paramount. Both query-reformulation approaches and chatbots are limited for this type of task, since the former only learn to mimic users, while the latter are bounded by the domain they have been trained on. To take a first step towards truly exploratory and interactive IR, we introduce a framework, where users navigate document collections by expressing their preference among sets of queries proposed by the system at each step – thus refining the knowledge about the user’s information need. Our training approach, based on self-supervised and reinforcement learning techniques, aims at minimizing the amount of interactions required to reach relevant queries, and thus documents, for users. We experimentally show that the introduced framework enables efficient learning from interactions with simple user bots, that are demonstrated to generalize well in real-world settings.