Hierarchical Attention Based Spatial-Temporal Graph-to-Sequence Learning for Grounded Video Description.