In this paper, we present a method for symbol description based on spatio-structural as well as statistical features of visual elementary parts called 'vocabulary'. The extracted vocabulary is first organised into different groups based on their types (e.g., circle, corner). This vocabulary is used as a basis for an Attributed Relational Graph (ARG) where spatial relational descriptors formalise the links between the types, labelled with global shape descriptors. The description is used to globally recognise structure by comparing the signatures. The method is experimentally validated in the context of electrical symbol recognition from wiring diagrams.