DenseCNN-ATT:实体关系抽取的密集连接卷积神经网络  

DenseCNN-ATT:Dense Connected Convolutional Neural Network for Entity Relationship Extraction

作　　者：李雅欣 王佳英 单菁[1,2] 邵明阳 LI Yaxin;WANG Jiaying;SHAN Jing;SHAO Mingyang(School of Information and Control Engineering,Shenyang Jianzhu University,Shenyang 110168;School of Software,Shenyang University of Technology,Shenyang 110023)

机构地区：[1]沈阳建筑大学信息与控制工程学院,沈阳110168 [2]沈阳工业大学软件学院,沈阳110023

出　　处：《计算机与数字工程》2021年第12期2483-2489,共7页Computer & Digital Engineering

基　　金：国家自然科学基金-青年基金项目“海量文本数据下实体智能近似匹配的关键技术研究”(编号:61702346);国家自然科学基金-青年基金项目“基于跨平台资源整合的团队形成深度策略关键技术研究”(编号:61702345);辽宁省教育厅基金项目“面向生物大数据的局部近似匹配技术研究”(编号:lnqn202015)资助。

摘　　要：在远程监督(Distant Supervision,DS)实体关系抽取任务中,采用远程监督的方式虽然可以产生大量的标注数据,但是这种方法产生的数据集充满大量的噪声数据,从而会降低关系抽取的性能。为此,我们针对现有深度学习使用浅层和单一深层神经网络模型提取特征的局限,设计了一个融合注意力机制的密集连接卷积神经网络模型——DenseCNN-ATT,该模型采用五层卷积深度的CNN,构成密集连接卷积模块作为句子编码器,通过增加特征通道数量来提高特征传递,减少了特征梯度的消失现象;此外,为进一步减少噪声影响,论文将网络的最大池化结果融合注意力机制,通过强调句子权重,来提升关系抽取性能。该模型在NYT数据集上的平均准确率达到了83.2%,相比于目前效果较好的浅层网络PCNN+ATT和深层网络ResCNN-9提升了9%~11%。实验证明,该模型能够充分利用有效的实例关系,在综合性能上明显优于目前效果较好的主流模型。In the task of distant supervision(DS)and entity relationship extraction,a large number of labeled data can be generated by DS.However,data sets generated by this kind of method contain a lot of noisy data,which can reduce the performance of relation extraction.In order to overcome the limitation of shallow layer and single deep layer neural network model used in the deep learning for feature extraction,a dense connected convolutional neural network model combined with attention mechanism model:DenseCNN-ATT is designed.This model adopts convolution neural network(CNN)with 5 convolution layers in depth,which constructs dense connected convolution module as sentence encoder.Feature transmission is enhanced by adding feature channels,so the gradient disappearance is alleviated.In addition,to further reduce the impact of noise,the max-pooling results are combined with attention mechanism and the sentence weight is emphasized to improve the performance of relation extraction.The re⁃sults show that the average accuracy reaches 83.2%at NYT-Freebase dataset,which is 9%~11%higher than PCNN+ATT and ResCNN-9.This model can make full use of effective instance relationship,and its comprehensive performance is better than main⁃stream models.

关 键 词：密集连接 关系抽取 注意力机制 卷积神经网络 远程监督 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]