基于领域嵌入的循环膨胀卷积神经网络评价对象抽取方法  

Aspect extraction based on domain embedding iterated dilated convolutions neural network

作　　者：陈积常[1] 周武 Chen Jichang;Zhou Wu(Infomation Engineering College,Nanning University,Nanning 530200,China;School of Computer Science,South China Normal University,Guangzhou 510631,China)

机构地区：[1]南宁学院信息工程学院,南宁530200 [2]华南师范大学计算机学院,广州510631

出　　处：《国外电子测量技术》2022年第1期20-27,共8页Foreign Electronic Measurement Technology

基　　金：南宁学院教授工程项目(2021JSGC01)资助。

摘　　要：在评价对象抽取任务当中,基于循环神经网络的方法具有前向依赖性且未能利用图形处理器(graphics processing unit,GPU)并行计算的能力,而基于传统卷积神经网络的方法则存在语义覆盖能力有限的问题。针对所述问题,为充分发挥模型的计算能力以及语义覆盖能力,首次将循环膨胀卷积神经网络应用于该任务,并融入领域嵌入特征,提出了一种基于领域嵌入的循环膨胀卷积神经网络模型(domain embedding iterated dilated convolutions neural network,DE-IDCNN)。在评价对象抽取任务数据集L14、R14、R15、R16的实验中,取得的F1值分别为81.85、80.90、72.23、73.26,与基线模型相比取得了两个数据集实验效果的领先。经进一步实验验证,DE-IDCNN模型实现了更高的计算效率以及表现出更好的语义覆盖能力。In the task of aspect extraction,the method based on recurrent neural network has forward dependence and fails to utilize the parallel computing ability of GPU,while the method based on traditional convolutional neural network has the problem of limited semantic coverage.In order to solve these problems,make the model fully optimize the computational efficiency and utilize the semantic coverage ability,this paper applies the iterated dilation convolution neural network with domain embedding feature for the first time,and proposes a model based on domain embedding iterated dilation convolution neural network(DE-IDCNN).In the experiment for task related data sets L14,R14,R15 and R16,the F1 values were 81.85,80.90,72.23 and 73.26 respectively.Compared with the baseline model,the experimental results of the two data sets are in the lead,and further experiments show that the DE-IDCNN achieves higher computational efficiency and better semantic coverage.

关 键 词：评价对象抽取 领域嵌入 膨胀卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]