Statistical response method and learning data acquisition using gamified crowdsourcing for a non-task-oriented dialogue agent

Abstract

This paper presents a proposal of a construction method for non-task-oriented dialogue agents (chatbots) that are based on the statistical response method. The method prepares candidate utterances in advance. From the data, it learns which utterances are suitable for context. Therefore, a dialogue agent constructed using our method automatically selects a suitable utterance depending on a context from candidate utterances. This paper also proposes a low-cost quality-assured method of learning data acquisition for the proposed response method. The method uses crowdsourcing and brings game mechanics to data acquisition. Results of an experiment using learning data obtained using the proposed data acquisition method demonstrate that the appropriate utterance is selected with high accuracy.