In a lot of existing legal knowledge based applications, the underlying legal model does not retain isomorphism with the original legal text. Parts of the text have not been modelled, and other parts have been simplified to single if-then-else clauses. This makes these models difficult to validate, maintain and re-use. In this article we propose to make an intermediate model, in which the original structure of the legal text is still represented. A first step is the detection and classification of norms in sentences in legal texts. We present a classification of norms or provisions based on the analyses of a large body of Dutch law. The classification covers all text constructs found thus far and we claim the use of typical sentence structures enables automation of a large part of this task.