【摘要】 针对现有图卷积网络在关系抽取任务中存在文本语义，语法表征不准确和在不同树结构上并行化计算较难等问题，本文提出一种基于BERT和注意力引导图卷积网络的关系抽取模型。首先，在模型的输入层使用BERT和Bi-LSTM编码出适应于上下文语境的词向量；其次，对输入的树结构采用最短路径为中心的修剪方式，减少树中的无关信息；最后，在模型中引入多头注意力机制，自动学习不同子空间内对关系提取有用的相关子结构，并在TACRED数据集上进行验证。实验结果表明，相对于基线模型，本文提出的模型显著提高了实体关系抽取的F₁值。更多还原

【Abstract】 To address the problems of text semantics, inaccurate syntactic representation and difficult parallelized computation on different tree structures in existing graphical convolution networks for relationship extraction task, a relation extraction model based on BERT and attention-guided graphical convolution networks is proposed. First, word vectors adapted to the context are encoded in the input layer of the model using BERT and Bi-LSTM. Then the shortest path-centered pruning is applied to the input tree structure to reduce irrelevant information in the tree. Finally, a multi-head attention mechanism is introduced in the model to automatically learn relevant substructure useful for relation extraction in different subspaces. The experimental results on TACRED dataset show that the proposed model in this paper significantly improves the F₁ value of entity relationship extraction compared to the baseline model.更多还原