Most recent efficient statistical approaches for natural language understanding require a segmental annotation of training data. Such an annotation implies both to determine the concepts in a sentence and to link them to their corresponding word segments. In this paper we propose a two-steps alternative to the fully manual annotation of data: an initial unsupervised concept discovery, based on latent Dirichlet allocation, is followed by an automatic segmentation using integer linear optimisation. The relation between discovered topics and task-dependent concepts is evaluated on a spoken dialogue task for which a reference annotation is available. Topics and concepts are shown close enough to achieve a potential reduction of one half of the manual annotation cost.