Affective Natural Language Generation

The automatic generation of natural language messages (or NLG) has been employed in many computer systems and for various purposes, which go from providing information on a particular subject, to instructing on how to perform some -complexaction, to arguing about a particular claim (for an overview, see [24]). In the majority of systems that have been designed in the last decade, these messages are adapted, in their content and presentation style, to the context in which they have to be applied: that is, to the Hearer’s characteristics (represented in a “mental model” of the Hearer), to the application domain and, generally in a more implicit way, to the Speaker’s characteristics. Adaptation is based on “strong” assumptions about the Hearer’s mental state and the way this state is influenced by communication of each individual component of the message, and by understanding of the relationships among these components. The aspects of the mental state that are represented, in the large majority of cases, are the Hearers’ beliefs and knowledge of domain topics and their goals about the domain state; in some cases, this may extend to representing other aspects, such as the ability to perform domain actions, the interest towards topics, the preference towards domain states. When a Speaker’s model is represented as well, this includes second-order beliefs and goals of the same type. Generally seen as informative tools, these systems give little space to representation of less rational aspects, such as emotions, of both Speaker and Hearer. However, natural language communication is influenced by a number of factors, of which the more rational ones constitute only a subset. Especially when communication occurs in “delicate” scenarios, such as a medical setting, the Hearer cannot be expected to coolly react to what is being said. Many studies in the health behaviour research have shown that patients’ attention and understanding is highly affected by their emotional involvement [1,26]. It therefore appears worthwhile to investigate whether, when and how emotions, personalities and other extra-rational factors should be taken into account when designing a NLG tool. By paraphrasing Rose Picard [22], we define Affective Natural Language Generation as “NLG that relates to, arises from or deliberately influences emotions or other non-strictly rational aspects of the Hearer”: these aspects (denoted with the generic term of “attitudes”) include personality traits, emotions and highly-placed values. We argue for the need for affective NLG by considering