Impact of agent’s answers variability on its believability and human-likeness and consequent chatbot improvements

Although globally less efficient than advanced dialogue systems, the chatbot approach allows people to easily design conversational agents. We suggest that one of their main drawbacks, their lack of believability, could be bypassed through the addition of variability in their answers, particularly when the variations depend on previous interactions or on particular parameters defining the agent. We validate the legitimacy of that hypothesis in two steps: first through simple additions to our chatbot-like framework (DIVA), we show it is technically feasible to simulate degrees of variability in answers. Then through an experiment done on 21 subjects interacting with two among six DIVA agents with different degrees of variability in a classical meeting scenario, we show that agents with an advanced variability in their answers are indeed perceived as more believable, human-like, and globally, more satisfying.