基于时间关系的Bi-LSTM+GCN因果关系抽取

针对传统时间关系只应用在机器学习方向关系抽取的问题, 提出一种基于序列标注实体识别的关系抽取方法. 先构建双向长短期记忆网络(Bi-LSTM)模型进行特征提取, 再输入时间关系作为特征矩阵进行图卷积. 实验结果表明: 时间关系能提高因果关系抽取效果, 并且包含时间关系的Bi-LSTM+GCN模型能有效抽取因果事件; 带有时间关系的Bi-LSTM+GCN模型获得因果关系的抽取结果优于传统方法因果关系的抽取结果.

Bi-LSTM+GCN Causality Extraction Based on Time Relationship

Aiming at the problem that traditional time relationships were only applied in the direction of machine learning, we proposed a relationship extraction method based on sequence labeling entity recognition. We first constructed Bi-LSTM model for feature extraction, and then input time relationship as a characteristic matrix for graph convolution. The experimental results show that the time relationship can improve the effect of causality extraction, and the Bi-LSTM+GCN model containing time relationship can effectively extract causal events, and the results of causality extraction of the Bi-LSTM+GCN model with time relationship are better than those of traditional methods.