Coding energy knowledge in constructed responses with explainable NLP models