Algorithm CPLCL (Classification by Preferential Clustered Link) for the classification graphs of syntactic variations, introduced by Ibekwe & SanJuan in ECAI'98, has the mathematical properties of single link clustering, since it is partly founded on the concept of component related to a subgraph, while avoiding the chain effect as much as possible. In spite of these properties, this algorithm of hierarchical classification does not make it possible to avoid the formation of some classes of too great dimension. However, the visualization of subgraphs corresponding to these classes with the AiSee interface highlights obvious structural graph properties, which enable their decomposition with recent formal tools. This motivates us in developing underlying graph-theoretical aspects with algorithm CPCL presented in TermWatch and in defining a decomposition algorithm for these classes.