基于剪枝策略的知识增强方面级情感分析  

Aspect based sentiment analysis based on knowledge enhancement of pruning strategy

作　　者：韩虎[1,2] 孔博 何勇禧 徐学锋 HAN Hu;KONG Bo;HE Yongxi;XU Xuefeng(School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China;Gansu Artificial Intelligence and Graphic Image Engineering Research Center,Lanzhou 730070,China)

机构地区：[1]兰州交通大学电子与信息工程学院,甘肃兰州730070 [2]甘肃省人工智能与图形图像工程研究中心,甘肃兰州730070

出　　处：《华中科技大学学报（自然科学版）》2024年第11期140-146,共7页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(62166024)。

摘　　要：为了克服在方面级情感分析模型中因外部知识引入而造成的信息冗余与噪声干扰,提出一种基于剪枝策略的知识增强双通道图卷积网络模型.通道一构造一个语法图卷积模块,面向特定方面重构依赖树,利用剪枝策略去除与特定方面语法距离较远及无关的分支,使用情感知识进行补充,得到降噪后以特定方面为根节点的情感依赖树;通道二构造一个语义图卷积模块,通过门控注意力机制构建注意力权重矩阵挖掘语义信息.最后,利用图卷积网络捕获语法信息和语义信息,并通过交互机制进行信息的共享与融合.在4个公开数据集Twitter,Rest14,Rest15和Rest16上的实验结果表明:该模型的分类准确率分别为74.28%,83.21%,82.47%和90.10%,优于基准模型.To overcome the information redundancy and noise interference caused by the introduction of external knowledge in aspect based sentiment analysis models,a knowledge enhanced dual-channel graph convolutional network model based on pruning strategy was proposed.The channel one constructed a syntactic graph convolutional module,which was tailored to reconstruct the dependency tree for specific aspect,the pruning strategy was used to eliminate branches which were distant from and irrelevant to the syntax of specific aspect,and the tree with affective knowledge was supplemented to obtain a sentiment dependency tree with the specific aspect as its root node after noise reduction.The channel two constructed a semantic graph convolutional module,which built an attention weight matrix through gate attention mechanism to mine semantic information.Finally,the graph convolutional network was used to capture syntactic and semantic information,and the information was shared and fused through an interaction mechanism.Experimental results on four public datasets,Twitter,Rest14,Rest15 and Rest16,show that the classification accuracy of the model is 74.28%,83.21%,82.47%and 90.10%,respectively,which outperforms the benchmark models.

关 键 词：方面级情感分析 图卷积网络 知识增强 剪枝策略 情感知识 注意力机制 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]