Gated graph convolutional network with enhanced representation and joint attention for distant supervised heterogeneous relation extraction

Xiang Ying, Zechen Meng, Mankun Zhao, Mei Yu, Shirui Pan, Xuewei Li

Abstract: Distant supervised relation extraction which is to extract heterogeneous relations from text data without manual annotation has been widely used in decision-making tasks such as question answering or recommendation system. However, existing distant supervised methods inevitably accompany with the wrong labelling problem. They typically use attention mechanism to select valid instances while ignore the core of relation extraction, i.e., entity pairs and relations. To address this problem, in this paper we incorporate enhanced representations into a gated graph convolutional network to enrich the background information and further improve the attention mechanism to focus on the most relevant relation. Specifically, in the proposed framework, 1) we introduce a triplet enhanced word representation method to focus on not only position information but also entity pair and implicit relation information in a sentence; 2) we use a Gated Rectified Linear Units (GRLU) module to integrate triplet information into an instance so as to achieve the purpose of enhancing sentence-level features; and 3) we employ sentence-relation joint attention over multiple instances and multiple relations, which is expected to dynamically reduce the weights of those noisy instances and enhance the bag representation. Extensive experiments on two popular datasets show that our model achieves significant improvement over all baseline methods.