面向对话文本的实体关系抽取方法研究

Research on Entity Relationship Extraction for Dialogue Text

对话关系抽取旨在预测对话文本中的实体对之间的关系。相比于规范文本,对话文本中的信息密度较低,存在关系的实体对通常出现在不同对话语句中。针对以上问题,本文提出了一个融入图神经网络的对话关系抽取模型,该模型使用BERT模型学习到对话的语义特征,根据关键信息构建关系图,并采用图注意力机制来对文本中重要的节点进行学习。模型将学习过后的特征放入到多视图图卷积模块,然后利用潜在多视图捕获到更多有用的特征,最后通过结合文本全局信息和细化图后的信息对两个实体对关系进行分类。实验结果表明,与其他模型相比,在公用数据集DialogRE上所提出的模型性能有明显的提升。

Dialogue relationship extraction aims to predict the relationships between entity pairs in dialogue text.Compared to standard text,the information density in dialogue text is lower,and entity pairs with relationships typically appear in different dialogue sentences.In response to the above issues,this article proposes a dialogue relationship extraction model incorporating graph neural networks.The model uses the BERT model to learn the semantic features of the dialogue,then constructs a relationship graph based on key information,and uses graph attention mechanism to learn important nodes in the text.The model places the learned features into the multi view graph convolution module,and then captures more useful features using potential multi views.Finally,the relationship between the two entities is classified by combining global text information and refined graph information.The experimental results show that the model proposed on the common dataset DialogRE has significant performance improvements compared to other models.