Dialogue management in conversational agents through psychology of persuasion and machine learning

To be really effective, conversational agents must integrate well with the characteristics of the humans with whom they interact. This exploratory study focuses on a method for integrating well-assessed methods from the field of social psychology in the design of task-oriented conversational agents in which the dialogue management module is developed through machine learning. In particular, the aim is to achieve agents whose policies could take into account the psychological features of the human interactants to deliver personalized and more effective messages. The paper presents the psychological study performed and outlines the overall theoretical architecture of the software framework proposed. On the psychosocial side, we first assessed the effectiveness of differently framed messages aimed to reducing red meat consumption taking the Theory of Planned Behavior (TPB) as the psychosocial model of reference. Turning to the machine learning field, the resulting Structural Equation Model (SEM) was first translated into a probabilistic predictor using Dynamic Bayesian Network (DBN). In turn, such DBN became the fundamental element of a Partially Observable Markov Decision Processes (POMDP) in a reinforcement learning setting. The possibility to elicit complete interaction policies was then studied by applying Neural Monte Carlo Tree Search (Neural MCTS) methods. The results thus obtained introduce the possibility to develop new multidisciplinary and integrated techniques for the development of automated dialogue managing systems.