Response Probability Distribution Acquisition for Autonomous Dialogue Generation

In this work we present a conversation system that is aimed at encouraging conversation from older adults, where the responses are obtained automatically from existing dialogue corpora. This system is unique as it is designed to respond with questions specifically with the aim of encouraging the older adults to speak more and to prevent the conversation from falling into a dead halt. Human evaluation for naturalness was conducted on the responses of this system and compared to a rule based and a neural based conversation system. Our system was capable to outperform the neural based system while falling slightly short of the rule based system.