This article presents an approach to extracting the information expressed in a corpus of texts and to produce a summary. Several variants of extractive methods of text summarization have been implemented and evaluated. Their main originality lies in the exploitation of structures called CDS (which stands for Clause Description Structure) derived from an semantic role labeling component and not directly from the sentences composing the texts. The summary obtained is a subset of the CDSs from the original corpus; this format will allow the detection of textual inconsistencies. In this work, we re-transform the summarized CDS into text to allow comparison of our approach with those of the literature. First results are very encouraging: the proposed methods generally outperform implementations of reference methods.