Open Access

Alexander Heimerl, Tobias Baur, Elisabeth André

• Modelling and recognising affective and mental user states is an urging topic in multiple research fields. This work suggests an approach towards adequate recognition of such states by combining state-of-the-art behaviour recognition classifiers in a transparent and explainable modelling framework that also allows to consider contextual aspects in the inference process. More precisely, in this paper we exemplify the idea of our framework with the recognition of conversational engagement in bi-directional conversations. We introduce a multi-modal annotation scheme for conversational engagement. We further introduce our hybrid approach that combines the accuracy of state-of-the art machine learning techniques, such as deep learning, with the capabilities of Bayesian Networks that are inherently interpretable and feature an important aspect that modern approaches are lacking - causal inference. In an evaluation on a large multi-modal corpus of bi-directional conversations, we show thatModelling and recognising affective and mental user states is an urging topic in multiple research fields. This work suggests an approach towards adequate recognition of such states by combining state-of-the-art behaviour recognition classifiers in a transparent and explainable modelling framework that also allows to consider contextual aspects in the inference process. More precisely, in this paper we exemplify the idea of our framework with the recognition of conversational engagement in bi-directional conversations. We introduce a multi-modal annotation scheme for conversational engagement. We further introduce our hybrid approach that combines the accuracy of state-of-the art machine learning techniques, such as deep learning, with the capabilities of Bayesian Networks that are inherently interpretable and feature an important aspect that modern approaches are lacking - causal inference. In an evaluation on a large multi-modal corpus of bi-directional conversations, we show that this hybrid approach can even outperform state-of-the-art black-box approaches by considering context information and causal relations.…