This paper describes the design and the first steps of implementation of Ofr (Optical Formula Recognition), a system for extracting and understanding mathematical expressions in printed documents. Our approach clearly separate OCR step, geometrical treatments and syntactic analysis. In this paper we focus on the third part: we define a class of context-sensitive graph grammars for mathematical formulas, study their properties and show how to remove their ambiguities (by adding contexts in rules) to define efficient parsing. This method is based on a ``critical pairs'' approach in the sense of Knuth-Bendix algorithm.