Abstract
For a response generation model that reflects the characteristics (e.g., a person's interests and preferences) to work practically, it is required that the model has acquired a person embedding space that can be interpolated to enable response generation by a speaker who corresponds to an intermediate speaker between different persons and that the person embedding is easy to control. In this study, we trained the model using a large amount of dialogue data with user identifiers, which are suitable for acquiring an interpolable person embedding space, and dialogue data with persona sentences (sentences describing the characteristics of a person), which are highly controllable for person representation, by mixing the two types of dialogue data. We propose a dialogue model that can generate responses via this person embedding. To demonstrate the effectiveness of the proposed method, we compared it with a conventional response generation model that does not explicitly model persona embedding and evaluated the interpolability and controllability of the persona embedding obtained by the proposed method.
