This paper proposes a fuzzy neural network, named the Knowledge-Based Neural Network with Trapezoid Fuzzy Set inputs (KBNN/TFS), which processes fuzzy input in a trapezoid membership function. To facilitate the processing of fuzzy information, an LR-fuzzy interval is employed. The initial structure of KBNN/ TFS is constructed using existing partail fuzzy rules and then revised by neural learning. These partial domain theories may be incorrect or incomplete. A consistency checking algorithm is proposed for verifying the initial knowledge and the revised fuzzy rules in order to detect redundant rules, conflicting rules and subsumed rules. In addition to fuzzy rule verification, fuzzy rule refinement and generation are proposed. We show the workings of the proposed model in an example called the Knowledge Base Evaluator (KBE). The results show that the proposed algorithm can detect inconsistencies in KBNN/TFS. By removing the inconsistencies and applying a rule insertion mechanism, the learning of the neural network is greatly improved. Furthermore, a consistent fuzzy rule base is obtained.