I don't see it. I think the current paradigm - translate your word problem into appropriate LP code, read and interpret the results - is ideal, because the models are so good at the translation and interpretation steps and the code is still auditable.