Effect Of Persona Updates.pdf

The effect of adding personas on response generation in Persona dialogue systems is analyzed. Existing approaches to persona dialogue systems attempt to incorporate some textual descriptions as explicit profiles of the system, but the persona of such systems is fixed at the beginning and cannot be automatically updated from a pre-defined profile. A new problem setting is considered where a persona dialogue system automatically updates its own persona according to the dialogue history. A persona dialogue system with a persona addition mechanism is proposed and the effects of possible problems are investigated. The system's response history is used to extract new personas, which are then added to the existing persona. The impact of the new persona on response generation is evaluated using metrics such as distinctness and consistency score. The results show that the system's response contains a similar distribution of persona attribute categories as the user's input, and that about 17% of the responses contain the same attribute category as the user's input. The study also investigates the effect of filtering on the addition of new personas, including the removal of duplicates and contradictions.