A Comparative Study of Classical and Deep Classifiers for Textual Addressee Detection in Human-Human-Machine Conversations

Abstract

The problem of addressee detection (AD) arises in multiparty conversations involving several dialogue agents. In order to maintain such conversations in a realistic manner, an automatic spoken dialogue system is supposed to distinguish between computer- and human-directed utterances since the latter utterances either need to be processed in a specific way or should be completely ignored by the system. In the present paper, we consider AD to be a text classification problem and model three aspects of users’ speech (syntactical, lexical, and semantical) that are relevant to AD in German. We compare simple classifiers operating with supervised text representations learned from in-domain data and more advanced neural network-based models operating with unsupervised text representations learned from in- and out-of-domain data. The latter models provide a small yet significant AD performance improvement over the classical ones on the Smart Video Corpus. A neural network-based semantical model determines the context of the first four words of an utterance to be the most informative for AD, significantly surpasses syntactical and lexical text classifiers and keeps up with a baseline multimodal metaclassifier that utilises acoustical information in addition to textual data. We also propose an effective approach to building representations for out-of-vocabulary words.