This paper focuses on the task that consists in automatically structuring free texts according to semantic principles, hence requiring a discourse analysis. We show that the task can be rephrased as a machine learning task in which the algorithm is supposed to take an optimal decision from the range of complex interacting constraints. The approach is implemented and evaluated taking the example of Health Practices Guidelines, medium-size documents intended to describe common practices that should be followed by physicians. Our approach outperforms previous approaches limited to sentence boundaries or requiring a lot of manual work.