This paper describes a new approach for multime- dia documents indexing and addresses the problem of automati- cally detecting a large number of visual concepts. Though using a multi-label approaches are used in some works, concepts detectors are often trained independently. We propose a model that takes into account the detection of not only a target concept but also other ones and regroups in terms of semantics similar samples. The expected benefit from such a combination is to consider the relationships between concepts in order to reclassify the results of an initial indexing system. Experiments on the TRECVID 2012 data are presented and discussed. Our method has significantly improved a quite good baseline system performance up to +6% on mean average precision.