Due to the complexity of the systems, successful modelling of intracellular reaction networks must rely on lumping techniques which systematically reduce the number of variables and parameters. Fortunately, the time scale separation characteristics of biochemical systems provide opportunities for eliminating unnecessary details. Through the proper interpretation of eigenvalues and eigenvectors, this article presents a theoretical basis for systematic model reduction. Results are generalized as a semiheuristic basis for lumping systems without complete kinetic information. It is also illustrated that the simplified system can yield new insight which is otherwise unavailable.