Intentionality and speech generation methods that create characterization

Abstract

In order to give chatbots characterized personality, stylistic features and speech intention response patterns were extracted from dialogues in works that have been successful in the character business were used. Based on the extracted results, an estimation system for speech intention was constructed, and speech was generated in a large-scale language model. Specifically, a system was constructed to estimate 13 types of intentions by performing fine tuning on the learned model. The classification accuracy was 74.67% when output that was confirmed by the analyst and deemed acceptable was included as a correct answer. The accuracy of generating appropriate responses to utterances was 75.90%. In the future, it is believed that it will be possible to generate more characterful speech by manipulating the character's response intention according to the intention estimation.