I3A: An Intelligent Interactive Information Agent Model for Information Retrieval
ChengXiang Zhai · 2022
Information Retrieval (IR) can be broadly interpreted as an interactive process for connecting users with the right information at the right time to finish a task, where interaction can be multimodal (e.g., using text, speech, and gestures) and connection can be made in multiple ways (e.g., querying, browsing, and recommendation). Although many formal IR models have been developed, the existing models are generally restricted to modeling the problem of ranking information items in response to a user's query without much consideration of user interaction. As a result, how to develop a general formal model that can cover all the variations of interactive IR (IIR) remains an open challenge. In this talk, I will discuss how we can address this challenge and present a general formal model for IR, called Intelligent Interactive Information Agent (I3A) model, which provides a unified theoretical foundation for both optimizing and evaluating sophisticated IIR algorithms and application systems. In I3A, an IIR system is modeled generally as an intelligent interactive information agent which plays an interactive cooperative "game" with its user(s), where both parties would take turns to "make moves" and interact with each other with a common objective of helping a user finish a task with minimum overall user effort. The optimization of IIR can be formally modelled as the agent optimizing a sequence of interaction decisions in response to each user action in a Bayesian decision framework. I will discuss how to refine the various components of the decision framework to make I3A operational and how multiple existing models, such as the Interface Card Model, the Probability Ranking Principle for IIR, formal models of users, and online learning to rank, can all be covered in the general I3A model. The I3A model also naturally suggests a new general methodology of evaluating IIR systems using search simulation.