Existing systems or architectures hardly provide any way to localize sub-parts of multimedia objects (e.g. sub regions of images, persons, events…), which represents hidden semantics of resources. To simplify and automate discovering hidden connections between such resources, we describe and evaluate in this paper, an algorithm for creating semantic relationships between multimedia news resources, giving a contextual schema (represented in RDF) as a result. This latter, which could eventually be used under any retrieval system, is integrated in our main multimodal retrieval system.We have also proposed and introduced a special measure of accuracy since evaluation relies on users’ intentions. An experimental evaluation of our algorithm is presented, showing encouraging results.