Training efficient statistical approaches for natural language understanding generally requires data with segmental semantic annotations. Unfortunately, building such resources is costly. In this paper, we propose an approach that produces annotations in an unsu-pervised way. The first step is an implementation of latent Dirichlet allocation that produces a set of topics with probabilities for each topic to be associated with a word in a sentence. This knowledge is then used as a bootstrap to infer a segmentation of a word sentence into topics using either integer linear optimisation or stochastic word alignment models (IBM models) to produce the final semantic annotation. The relation between automatically-derived topics and task-dependent concepts is evaluated on a spoken dialogue task with an available reference annotation.