Towards Automated Dialog Personalization using MBTI Personality Indicators

As conversational interfaces mature in both capacity and usage, the need to personalize towards specific user characteristics becomes apparent, in order to improve users' acceptance, satisfaction and trust in the conversations. We utilize the concept of Myers-Briggs personality type indicators in order to adapt chatbot behavior. In a user study, we investigate the impact and realization of the so-called "law of attraction" by providing users with a chatbot that mirrors their own personality. This entails predicting the personality from the user behavior, in this work chat messages, by utilizing a pre-trained language model rather than composing many resources like lexicons. We conduct a user study with aligned and misaligned personality and analyze the effect on usability. Results show that alignment significantly improves major usability factors such as satisfaction, perceived naturalness, recommendation likelihood, appropriateness and trustworthiness of our interaction. Further, comparing different language models, contrastive learning approaches outperform previous methods. Predicting the thinking vs. feeling and introversion vs. extroversion indicator dichotomies, we achieve 76.14% F-1 and 69.11% F-1, respectively, with setting a new state-of-the-art performance in the literature for the former. Finally, our work adds transparency to the design of linguistic personality cues, hitherto rarely reported in the literature.