Distant Supervision Relation Extraction With Graph Convolutional Networks

Distant supervision has advantages of generating training data automatically for relation extraction by aligning triples in Knowledge Base with large-scale corpora,but it suffers from noisy labels which damage the performance of relation extraction.Some recent methods at-tempt to incorporate background knowledge to enhance the performance of relation extraction.However,there still exist two major limitations.Firstly,these methods are tailored for a specific type of information which is not enough to cover most of the cases.Secondly,the introduced background knowledge may contain noise.To address these limitations,we propose a novel Edge-Reasoning Hybrid Graph model(ER-HG),which can incorporate heterogeneous background information on a graph framework,such as entity types and text relation paths.We also propose an Edge-Reasoning Graph Con-volutional Network(ER-GCN)to learn a better graph representation and design a supervised attention mechanism to alleviate the side effect of noise knowledge.The main contributions of this thesis are as follows:1)We propose a novel hybrid graph model,which can incorporate heterogeneous background knowledge in a graph framework.The graph framework is flexible to integrate multiple nodes even with several missing cases.2)we propose an edge-reasoning graph convolutional network(ER-GCN)to encode entire graph and get a better discriminative representation.The ER-GCN can take advantage of the correlation between different nodes and control effective information flow to the neighbor node.We even utilize a supervised attention mechanism to learn weights of each knowledge node to alleviate the side effect of noise knowledge.3)We conduct extensive experiment in three real-world datasets.Experimental results demon-strate that our model outperforms the state-of-the-art methods significantly in various eval-uation metrics.In the real world scenario,entity-related and text-related background knowledge can be obtained through the vertical field.Research on introducing heterogeneous background knowl-edge can effectively improve relation extraction model to better utilize these pieces of knowl-edge and make better relation prediction.