A novel technique that applies the neural-network learning strategy of back-propagation to recognize semantically inconsistent rules is presented. When the rule strengths of most rules are semantically consistent, semantically inconsistent rules can be recognized if their strengths are weakened or change signs after training with correct samples. In each training cycle, the discrepancies in the belief values of goal hypotheses are propagated backward and the strengths of rules responsible for such discrepancies are modified appropriately. A function called consistent-shift is defined for measuring the shift of a rule strength in the direction consistent with the strength assigned before training and is a critical component of this technique. The viability of this technique has been demonstrated in two examples and one practical domain.