In this article, we propose a methodology to automatically construct fuzzy rules in a Fuzzy Rule-Based System that detects emotions from an actor's performance during a show. The idea is to collect video data from a lot of performances of the same show from which it should be possible to visualize all the emotions and intents or more precisely ''intent graphs''. To do this, we analyze the collected data defining low-level descriptors that are converted into high-level characterizations thanks to aggregations. The following step is to partition correctly the universes in classes, depending on the retrieved data. Finally the fuzzy rules built from the classes are used to label the detected emotions.