Improving Interpretability via Explicit Word Interaction Graph Layer