Smoothing Conversation Using Dialogue Agents Accompanied by Advisory Agent Inside

There are many attempts to support conversation. Their progress requires a large number of datasets through experiments with human subjects. However, it costs a huge amount of time and effort. This paper proposes a method to collect datasets from agents that simulate human conversations, especially chatting. Dialogue between agents has a problem of quick convergence because agents cannot deepen the conversation. The proposed method incorporates an advisory agent inside the dialogue agent. The advisory agent gives advice to the dialogue agent on the next topic to talk about based on the dialogue content. The advisory agent also has a discernment function to select the timing of the advice. To examine the usefulness of the proposed method, the paper compares conversation between conventional methods, between the proposed methods, and between humans. This paper defined inter-sentence similarity as the degree of similarity of the current sentence to the sentence immediately preceding it. Higher values indicate convergence into conversation depends while lower values indicate successful development of the conversation topics. The proposed method can maintain a low value over time, whereas the conventional one puts a high value as time goes on. Actual logs also show less converged conversation between proposed agents. This result means that the proposed method is superior to the conventional one in terms of developing and deepening the topics.