This paper develops a structural symbol recognition method with integrated statistical features. It applies spatial organization descriptors to the identified shape features within a fixed visual vocabulary that compose a symbol. It builds an attributed relational graph expressing the spatial relations between those visual vocabulary elements. In order to adapt the chosen vocabulary features to multiple and possible specialized contexts, we study the pertinence of unsupervised clustering to capture significant shape variations within a vocabulary class and thus refine the discriminative power of the method. This unsupervised clustering relies on cross-validation between several different cluster indices. The resulting approach is capable of determining part of the pertinent vocabulary and significantly increases recognition results with respect to the state-of-the-art. It is experimentally validated on complex electrical wiring diagram symbols.