The design of a dataset needed to train a chatbot is most often the result of manual and tedious step. To guarantee the efficiency and objectivity of the annotation, we propose an active learning method based on constraints annotation. It's an iterative approach, relying on a clustering algorithm to segment data and using annotator knowledge to lead clustering from unlabeled question to relevant intents structure. In this paper, we study the optimal modeling parameters to get an exploitable dataset with a minimum of annotations, and show that this approach allows to make a coherent structure for the training of a chatbot.