A computer-implemented method, computer system, and computer program product for improving a natural language generation knowledge base is provided. The method may include detecting user input comprising a natural language expression, generating an erroneous natural language response based on the user input and the knowledge base, determining user feedback corresponding to the erroneous natural language response, wherein the user feedback comprises an indication of an end-user reaction to the erroneous natural language response, determining an improvable performance metric with respect to the knowledge base, and updating the knowledge base based on the improvable performance metric, wherein the knowledge base comprises an explicit model of language corresponding to the erroneous natural language response, and wherein updating the knowledge base comprises updating the explicit model of language based on the user input, the erroneous natural language response, and the user feedback.