ON WIRING EMOTION TO WORDS: A BAYESIAN MODEL

Language and emotion are deeply entangled. In this dissertation we present a theoretical model that addresses how language and emotions intertwine with one another. To such end, we draw on the several results achieved in emotion theory (either at the psychological and the neurobiological levels) that go under the constructivist umbrella of the Conceptual Act Theory and those related to an emerging theoretical framework for pragmatic inference, the Rational Speech Act framework. We connect these theories and spell such connection in the language of probability, namely in that of Bayesian probabilistic modelling. Our endeavour is addressed to those fields of computer science such as artificial intelligence and machine learning where, in spite of the remarkable progress in the computational processing of language and affect, the study of their intersection is at best at its infancy, in our view. We argue that any further step in such direction only can be afforded by reducing the gap between Affective Science and computational approaches. To pave the way, simulations of the proposed model are presented that account for well known case-studies in pragmatics. In brief, at a high-level abstract representation we consider two interacting agents-in-context, where each agent performs a conceptual act based on interoceptive and exteroceptive sensation, in order to regulate their body budget. The agents communicate, performing communication acts that in turn regulate the agents’ conceptual acts and vice versa, and in this way they create, communicate and share categories, and even add new functions to the world. We implement this framework through two simulations of non-literal language use, namely hyperbole, irony, and a third dealing with politeness, a form of social reasoning. In addition, a fourth simulation concerns the assessment of the stochastic dynamics of the key component of the model, core affect.