Although expert-novice differences in various domains have so far been attributed to domain-specific schemata or perceptual-chunks, few past research has addressed the issue of schema acquisition itself. We address this issue in the domain of geometry proof problem-solving. Past literatures on geometry pointed out the significance of the diagrammatic features of problems as the basis of problem-solving memories and as a cue for abstract planning in constructing a proof. Based on this insight, we propose a new perceptual-chunking technique in which the learner chunks such “diagram elements visually grouped together” into a schema, using recognition propagation rules as chunking criteria. The rules represent how human solvers see the diagram elements and geometrical features. We implemented this chunking mechanism on a computer program, PCLEARN, and did some computational experiments to see how the learned chunks contribute to problem-solving improvement. Further, we designed a psychological experiment to examine how human subjects tend to parse the whole diagrams into parts after solving a set of geometry proof problems. Its result shows that the PCLEARN chunking technique can learn what human learner would learn much better than the conventional learners.