Who did what to whom is a major focus in natural language understanding, which is right the aim of semantic role labeling (SRL). Although SRL is naturally essential to text comprehension tasks, it is surprisingly ignored in previous work. This paper thus makes the first attempt to let SRL enhance text comprehension and inference through specifying verbal arguments and their corresponding semantic roles. In terms of deep learning models, our embeddings are enhanced by semantic role labels for more fine-grained semantics. We show that the salient labels can be conveniently added to existing models and significantly improve deep learning models in challenging text comprehension tasks. Extensive experiments on benchmark machine reading comprehension and inference datasets verify that the proposed semantic learning helps our system reach new state-of-the-art.

I Know What You Want: Semantic Learning for Text Comprehension