Edge Features Enhanced Graph Attention Network for Relation Extraction.

Dependency trees of sentences contain much structural information that is useful for capturing long-range relations between words in the text. In order to distill the useless information, the pruning strategy is introduced into the dependency tree for preprocessing. However, most hard-pruning strategies for selecting relevant partial dependency structures are too rough and have poor generalization performance. In this work, we propose an extension of the graph attention network for relation extraction task, which makes use of the whole dependency tree and its edge features. The graph attention layer in our model can implicitly prune the neighbor nodes of each node by assigning different weights according to the content. The edge feature information makes the pruning strategy trainable and non-discrete. Our model can be viewed as a soft-pruning approach strategy that automatically learns the relationship between different nodes in the full dependency tree. The results on various datasets show that our model utilizes the structural information of the dependency tree better and gets the state-of-the-art results.