LRTA: A Transparent Neural-Symbolic Reasoning Framework with Modular Supervision for Visual Question Answering