Efficient statistical approaches have been recently proposed for natural language understanding in the context of dialogue systems. However, these approaches are trained on data semantically annotated at the segmental level, which increases the production cost of these resources. This kind of semantic 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-step automatic method for semantic annotation. The first step is an implementation of the latent Dirichlet allocation aiming at discovering concepts in a dialogue corpus. Then this knowledge is used as a bootstrap to infer automatically a segmentation of a word sequence into concepts using either integer linear optimisation or stochastic word alignment models (IBM models). The relation between automatically-derived and manually-defined task-dependent concepts is evaluated on a spoken dialogue task with a reference annotation.