Improvement of the Construction of an Active Listening Dialogue Model using Reinforcement Learning

Abstract

In recent years, the percentage of the elderly population to the total population in Japan has reached an all-time high, and the number of dementia patients has been increasing at the same time as the aging of the Japanese population. “Validation” is a communication method to attend to the emotions of dementia patients. However, it is difficult to dedicate enough time with individual patients due to the recent shortage of human resources for caregivers. In a previous our study, the authors constructed an active listening dialogue model using reinforcement learning and evaluated the performance of the model. In this paper, we attempt to improve the speech templates and the number of templates when creating the reward model, which were among the issues in the previous study. The experimental results showed that the average reward was higher than that of the previous model, and result of the mechanical evaluation was better with an increase in the percentage of active listening. However, there were issues with the number of learning steps and the evaluation method in the human evaluation.