Personality-based dialogue systems with response type tendency control

Abstract

This project proposed a method to build dialogue systems with personalities by controlling the tendencies of the dialogue system to generate certain types of utterances, which is named response type tendencies in this project. This project trained an auto-classifier to label the types of utterances from dialogues in games and movie scripts. A survey was also carried out on the personality of characters who produced the utterances. Characters were grouped based on identified personality traits, and their response type tendencies were extracted from the labeled text. This enabled the calculation of response type tendency for each personality trait. The extracted personality-specific response type tendencies were used to build a GPT-3-based dialogue system (DS-RTTs) with certain personalities. The dialogue systems (DS-RTTs) went through manual evaluation on the personality traits they display in conversation. Our evaluation results show that response type tendency control has a significant impact on the perceived personality of dialogue systems, and that DS-RTTs built using the response type tendencies outperform the baseline method in displaying certain personality traits.