DIALOG MANAGEMENT USING ACTIVE LEARNING ALGORITHMS

Abstract

A dialog manager is a component of a dialog system that is responsible for the state and flow of the conversation. In this thesis, we explore the use of active learning for this task. For this thesis, the data we tested on is from a dialog system called “Let’s Go!” [Raux, Antoine, et al. (2005)]. The “Let’s Go!” dialog system tries to determine five slots – bus route, origin location, destination, date and time. It uses a directed design, where it asks for each of these slots sequentially and then, provide the bus schedules for the routes that the user requested as best as it can. The system can also ask for confirmations. We explore the use of active learning for this task by implementing two common greedy criteria: the least confidence criterion and the maximum Gibbs error criterion. We also implement an algorithm that uses information gain.