Abstract
Artificial intelligence pursues reasoning, planning, learning, and utilizing knowledge in a complex domain, such as teaching a machine to tell jokes, which catches many scholars’ eyes for these years. Subjected to the humor as being a double way interaction, both the comprehension and the performance are limited by individuals, though, we attempt to generate a joke in a certain way. We shift a previous vision on how to generate a joke, into a perspective of what to express based on the elaborate layout. Considering that a joke consists of a sequence of events, we set up a new method by using reinforcement learning. Following by the theory of ontological semantic theory of humor (OSTH) and Incongruity Theory, we provide a feasible reward schema to learn a good policy.