采用多尺度注意力机制的远程监督关系抽取

Multi-level Attention Mechanism Based Distant Supervision for Relation Extraction

针对目前大多数关系抽取模型中局部特征及全局特征利用不充分的缺点,该文提出一种采用多尺度注意力机制的远程监督关系抽取模型。在词语层面,通过在池化层构建权重矩阵来衡量词语与关系的相关程度,从而捕捉句子中重要的语义特征;在句子层面,采用注意力机制将预测关系与句子进行相关性比较,获得句子级别的重要信息。模型在NYT数据集上平均准确率达到78%,表明该模型能够有效地利用多尺度特征,并且提高远程关系抽取任务的准确率。

To best exploit the local and global feature,we proposed a distant supervised relation extraction model based on multi-level attention mechanism.We employ an attention matrix in pooling layer to capture the word-level sematic feature which indicates the relevant relationship between input words and relations.Moreover,we adopted sentence-level attention mechanism to compare the relationship between sentences and predicted relations.Experimental results show that the mean accuracy of the proposed model achieves 78% in the NYT data set,indicating an effective use of multi-level feature and better performance of distant relation extraction task.