A Weighted Diffusion Graph Convolutional Network for Relation Extraction
Jiusheng Chen, Zhenlin Li, Hang Yu, Xiaoyu Zhang · Journal of Electrical and Computer Engineering · 2024
Currently, graph convolutional network (GCN) is widely used in relation extraction (RE) tasks. Within RE tasks in the form of directed graphs, the placement of entities in the sentence context generates a large number of remote entity nodes in the directed graph. However, the GCN has an acceptance domain limitation and can only aggregate consecutive neighbor nodes, making it ineffective for remote node information extraction. To tackle the above issues, a new model GCN based on weighted diffusion is proposed, including two improvements. First, the pretrained BERT model is introduced to extract the semantic features from the original sentences, and then the diffusion convolution is modified by the weighted matrix using the Gaussian kernel function to calculate the distance of nodes on the graph, enabling the diffusion convolution to selectively aggregate information of remote nodes. The effect of remote node information aggregation can be improved by shortening the distance of related nodes and increasing the distance of unrelated nodes. Second, an identity mapping is added to ensure that the weight matrix can match each layer adaptively. The efficacy of our proposed model has been demonstrated through experimental results on two publicly available datasets, TACRED and SemEval, which have shown that it outperforms other baseline models.