Pruning Attention Heads of Transformer Models Using A* Search: A Novel Approach to Compress Big NLP Architectures.