Modeling human-machine interaction by means of a sample selection method

Abstract

This paper presents a practical application of Sample Selection techniques to model the process of selecting the next system response of a conversational agent. Our proposal deals with the important problem of imbalanced training data that is usually present in the selected application domain. This process is modeled as a classification task that takes the dialog history as input, and selects the next system response as output. Our proposal improves the classifier’s performance by automatically selecting examples that are difficult to classify during the training phase, considering the criteria of proximity to the border and the typicality of the examples. We present a practical application of this technique for a conversational agent providing railway information. Simulation results support the usefulness of the proposed approach to provide the better selection of the responses of the conversational agent.