This paper presents a N-R method of acquiring multi-level reasoning production rules automatically based on NN(neural network), so as to resolve the bottle neckproblem in building a knowledge - based system-knowledge acquisition. This methodadopts a characteristic NN structure model and corresponding learning algorithm, so thatN -R method can fix the number of hidden nodes dynamically and form new conceptswhich are not defined in the sample set in the course of learning. Having been trained, NNis transformed into an RN (reasoning network) with the proposed transform algorithm,from which productionrules can be obtained easily.