Toward Practical Usage of the Attention Mechanism as a Tool for Interpretability