Structural relationships for large-scale learning of answer re-ranking