Hierarchical Attention Graph Convolution Networks for Relationship Extraction

Attention mechanisms have gained significant prominence in the realm of natural language processing. When it comes to relation extraction in the form of a graph, the complex sentence structure and uncertain entity positions pose considerable challenges. To address these limitations, we propose a hierarchical attention graph convolutional neural network (HAGCN) for relation extraction. The HAGCN model reflect the extraction process at multiple levels spanning the graph, sentence, and word levels; In addition, the proposed model incorporates three levels of attention mechanisms which precisely adjust feature vector distribution in the semantic space to prioritize essential vocabulary information and facilitate the extraction of dependency distances between nodes and entities, enhancing the comprehension of intricate sentence structures. The experimental results of TACRED and SemEval indicate that the proposed model performs better than other baseline models.

I. Introduction

Relationship extraction (RE) task is a crucial task in knowledge graph construction that involves identifying and determining specific relationships between entities mentioned in text. In the field of natural language processing (NLP) [1], RE serves as an upstream task for various high-level tasks such as information extraction [2], knowledge graph construction [3], and question answering systems [4]. As a result, it has garnered significant research attention in recent years.