Improving Interpretability via Explicit Word Interaction Graph Layer

Recent NLP literature has seen growing interest in improving model interpretability. Along this direction, we propose a trainable neural network layer that learns global interaction graph between words and then selects more informative using the learned word interactions. Our layer, call WIGRAPH, can plug into any network-based text classifiers right after its embedding layer. Across multiple SOTA models various datasets, demonstrate adding WIGRAPH substantially improves models' interpretability enhances prediction performance at same time.