摘要： 基于卷积神经网络的远程监督关系抽取方法提取的特征单一，且标准交叉熵损失函数未能较好处理数据集中正负样本比例不均衡的情况。为此，提出一种基于深度残差神经网络的远程监督关系抽取模型，通过改进交叉熵聚焦损失函数，提取句子中的深层语义特征，同时降低损失函数中负样本的权重，避免在NYT-Freebase标准数据集中引入NA关系类别的噪音。实验结果表明，该模型能增强深度残差神经网络对含噪音数据的表示学习能力，有效提高远程监督关系抽取任务的分类准确率。

关键词: 交叉熵损失函数, 残差学习, 远程监督模型, 关系抽取, 卷积神经网络

Abstract: Distant supervision relation extraction based on Convolutional Neural Network(CNN) can extract only single feature,and the standard cross-entropy loss function is not sufficient in balancing the ratio of positive samples and negative samples in datasets.To address the problem,this paper proposes a relation extraction model using distant supervision based on deep residual neural network,which improves the cross-entropy focal loss function to extract deep semantic features of a sentence.Also,the weight of simple negative samples in the loss function is reduced to introduce noise of the NA relation into standard NYT-Freebase dataset.Experimental results show that the model can enhance the ability of deep residual neural network to represent and learn sample data with noise,increasing the classification accuracy rate in relation extraction tasks using distant supervision.

Key words: cross-entropy loss function, residual learning, distant supervision model, relation extraction, Convolutional Neural Network(CNN)

中图分类号: 

• TP391