A comparison of addressee detection methods for multiparty conversations

Several algorithms have recently been proposed for recognizing addressees in a group conversational setting. These algorithms can rely on a variety of factors including previous conversational roles, gaze, and type of dialogue act. Both statistical supervised machine learning algorithms as well as rule based methods have been developed. In this paper, we compare several algorithms developed for several different genres of multiparty dialogue, and propose a new synthesis algorithm that matches the performance of machine learning algorithms while maintaining the transparency of semantically meaningful rule-based algorithms.