Modelling Emotion Dynamics in Chatbots with Neural Hawkes Processes

Abstract

Conversation partners tend to stick to a particular emotional state unless some external motivation excited them to change that state. Usually, the excitation comes from the other conversation partner. This preliminary study investigates how an Artificial Intelligence model can provide excitation for the other partner during a dyadic text-based conversation. As a first step, we propose a Neural Emotion Hawkes Process architecture (NEHP) for predicting future emotion dynamics of the other conversation partner. Moreover, we hypothesize that NEHP can facilitate learning of distinguishable consequences of different excitation strategies, and thus it allows for goal-directed excitation behavior by integrating with chatbot agents. We evaluate our preliminary model on two public datasets, each with different emotion taxonomies. Our preliminary results show promising emotion prediction accuracy over future conversation turns. Furthermore, our model captures meaningful excitation without being trained on explicit excitation ground-truths as practiced in earlier studies.