Improvements of Dialogues by Extracting Latent Tasks in Small Talk Dialogue Systems

Abstract

As one of the dialogue systems, there is a small talk dialogue system. This is a dialogue system for small talking with users, and has been treated as a non-task-oriented dialogue system in previous studies. However, it is thought that there are tasks that the speaker wants to accomplish even in small talk. Therefore, in this research, we consider the task that the speaker wants to achieve in the small talk dialogue as a latent task, extract the latent task from the small talk dialogue, and use it to make the dialogue smoother. In order to achieve the above objectives, this study (1) defines latent tasks and investigates to evaluate validity and modify definitions, (2) creates a dataset based on the definitions, (3) we constructed a model for task extraction. In addition, as a result of experiments with the constructed model, about 66% of latent tasks could be correctly extracted.