Abstract: The ability to automatically recognize a sketch accurately is important to computer-based diagramming. Many recognition techniques have been proposed but few researchers have reported the use of formal methods to select the most appropriate ink features for recognition algorithms. We have used a statistical approach to identify the most important distinguishing features of ink for dividing text and shapes. We implemented these into an existing recognition engine and conducted a comparative evaluation. Our feature set more successfully classified a range of common diagram elements than two existing dividers.