Computational model of negotiation skills in virtual artificial agents

Abstract

Negotiation skills represent crucial abilities for engaging in effective social interactions in formal and informal settings. Serious games, intelligent systems and virtual agents can provide solid tools upon which one-to-one training and assessment can be reliably made available. The aim of the present work is to fill the gap between the recent growing interest towards soft skills, and the lack of a robust and modern methodology for supporting their investigation. A computational model for the development of Enact, a 3D virtual intelligent platform for training and testing negotiation skills, will be presented. The serious game allows users to interact with simulated peers in scenarios depicting daily life situations and receive a psychological assessment and adaptive training reflecting their negotiation abilities. To pursue this goal, this work has gone through different research stages, each with a unique methodology, results and discussion described in its specific section. In the first phase, the platform was designed to operationalize the examined negotiation theory, developed and assessed. The negotiation styles considered, consistently with previous findings, have been found not to correlate with personality traits, coping strategies and perceived self-efficacy. The serious game has been widely tested for its usability and underwent two development and release stages aimed at improving its accuracy, usability and likeability. The variables measured by the platform have been found to predict in all cases at least two of the negotiation styles considered. Concerning the user feedback, the game has been judged as useful, more pleasant than the traditional test, and the perceived time spent on the game resulted significantly lower than the real time spent. In the second stage of this research, the game scenarios were used to collect a dataset of documents containing natural language negotiations between users and the virtual agents. The dataset was used to assess the correlations between the personal pronouns' use and the negotiation styles. Results showed that more engaged styles generally used pronouns with a significantly higher frequency than less engaged styles. Styles with a high concern for self showed a higher frequency of singular personal pronouns while styles with a high concern for others used significantly more relational pronouns. The corpus of documents was also used to perform multiclass classification on the negotiation styles using machine learning. Both linear (SVM) and non-linear models (MNB, CNN) performed reliably with a state-of-the-art accuracy.