一种用于关系抽取的双层时空图卷积神经网络

A bi-layer spatiotemporal graph convolutional neural network for relation extraction

关系抽取是自然语言处理中一项基础的上游任务.句子的结构信息在某种意义上蕴含了实体及其关系信息,有助于提高关系抽取的准确率,然而使用现有自然语言处理(Natural Language Processing,NLP)语言工具进行句法分析时会引入一定的错误传播问题,且现有的基于图结构的关系抽取模型在一定程度上忽略了句子的时序信息.通过结合双向长短时记忆网络(Bi-directional Long Short-Term Memory,Bi LSTM)捕获句子序列的上下文关系,同时使用传统条件随机场(Conditional Random Field,CRF)的关系标注结果矫正NLP工具的错误传播问题,提出了一种用于关系抽取的双层时空图卷积神经网络(Bilayer Spatiotemporal Graph Convolution Neural Network,Bi SpGCN)模型.该模型在中文糖尿病数据集和中文人物关系数据集上的实验结果表明,相较于传统的多头注意力引导的图卷积神经网络(Attention Guided Graph Convolutional Networks for Relation Extraction,AGGCN)模型,BiSpGCN模型能够充分利用句子的有效信息,具有更好的关系抽取性能.

Relation extraction is a basic upstream task in natural language processing.In a sense,the structural information of sentences contains entity and relationship information,helping to improve the accuracy of relationship extraction.However,when using the existing natural language processing(NLP)language tools for syntactic analysis,some error propagation problems will be introduced.The existing relationship extraction model based on graph structure ignores the temporal information of sentences to a certain extent.By combining Bi-directional Long Short Term Memory(BiLSTM)to capture the context of sentence sequences,and using the relationship annotation results of traditional Conditional Random Field(CRF),the error propagation problem of NLP tool was corrected.A Bilayer Spatiotemporal Graph Convolution Neural Network(BiSpGCN)model for relationship extraction was proposed.The experimental results on the Chinese Diabetes dataset and Chinese Character relational dataset show that the BiSpGCN model can take full advantage of the effective information of sentences and has better relationship extraction performance compared with the traditional Attention Guided Graph Convolutional Networks for Relation Extraction(AGGCN)model.