融合位置特征注意力与关系增强机制的远程监督关系抽取

Distantly Supervised Relation Extraction with Position Feature Attention and Relation Enhancement

实体关系抽取是构建知识图谱的主要任务之一,旨在确定句子中实体之间的关系类别.远程监督关系抽取方法通过将远程知识库与文本数据对齐来自动标记数据,已成为处理关系抽取任务的主要方式.为解决远程关系抽取不能充分利用单词之间的位置关系信息,并且没有考虑重叠关系之间语义相关性的问题,本文提出一种融合位置特征注意力和关系增强机制的远程监督关系抽取模型.该模型使用基于高斯算法的位置特征注意力机制重新分配句子中单词的权重,并且采用分段卷积神经网络和词级注意力来捕获句子特征.然后,利用基于自注意力的关系增强机制来捕获重叠关系之间的语义关联.在NYT10公共数据集上的实验结果表明,本文模型的性能优于所比较的基线关系抽取模型.

Entity relation extraction is a tasks of knowledge graph construction,the purpose of this task is to determine the classes of relation between entities in sentences.Distantly supervised relation extraction methods,which automatically label data by aligning distant knowledge bases with text data,have become the main way to handle relation extraction tasks.To address the problems that distant relation extraction fails to fully use the position relation information between words and does not consider the semantic relevance between overlapping relations,we propose a distantly supervised relation extraction model with position feature attention and relation enhancement mechanism.The model uses the Gaussian algorithm-based position feature attention mechanism to reassign the weight of words in a sentence.Then,it employs PCNN and word-level attention to capture sentence features.Then,it uses the self-attention-based relation enhancement mechanism to capture the semantic correlation between overlapping relations.We conduct experiments on NYT10 public data set,and the results show that the model in our paper outperforms the compared baseline relation extraction model.