融入情感和话题信息的中文方面级情感分析  

Chinese aspect-based sentiment analysis integrating sentiment and topic information

作　　者：周法国[1] 孙冬雪 Zhou Faguo;Sun Dongxue(School of Mechanical Electronic&Information Engineering,China University of Mining&Technology-Beijing,Beijing 100083,China)

机构地区：[1]中国矿业大学(北京)机电与信息工程学院,北京100083

出　　处：《计算机应用研究》2022年第12期3614-3619,3625,共7页Application Research of Computers

基　　金：国家自然科学基金资助项目(62072008);中央高校基本科研业务费资助项目(2022YJSJD26)。

摘　　要：近年来的方面级情感分析模型应用图卷积神经网络(GCN)学习语句的语法结构信息,但是在建模时忽略了已知情感词信息和评论所属的已知话题环境,渐渐不能满足中文社交网络情感分析需求。针对以上问题,提出一种基于词典和深度学习软融合的字词双通道模型(2D-SGCN)。该模型首先基于基础情感词典扩展得到微博领域词典,获得领域适用性的情感词;其次使用预训练模型获得字、词初始特征向量,并在字维度融入方面词和话题信息,分别使用Bi-LSTM和融入情感信息的GCN(SGCN)学习全局与局部信息;应用注意力机制得到方面词最终特征并进行多维度融合;最后将话题和方面词结合进行分类纠正。在SemEval-2014的Restaurant数据集上F_(1)为73.67%,在NLPCC2012数据集上F_(1)为91.5%,证明了该模型的有效性。In recent years,the aspect-based sentiment analysis model uses graph convolutional network(GCN)to learn the grammatical structure information of sentences,but ignores the known sentiment word information and the known topic environment belonged to the comments.These gradually can’t meet the needs of sentiment analysis of Chinese social networks.To solve the above problems,this paper proposed a double-channel of char and word(2D-SGCN)model based on soft fusion of dictionary and deep learning.Firstly,it constructed a microblog domain sentiment dictionary based on the expansion of the basic sentiment dictionary,then obtained sentiment words.Secondly,it used the pre-training model to obtain the initial feature vector of the chars and word,and integrated the aspect word and topic information in the char dimension.Then it used the Bi-LSTM and GCN with emotional information(SGCN)models to learn global and local information respectively.And it obtained the final features of aspect words through the attention mechanism and obtained better expression of text through multi-dimensional fusion.Finally,it combined the topic and aspect information for correction and classification.On the Restaurant dataset of SemEval-2014,F_(1) is 73.67%,on the NLPCC2012 dataset,F_(1) is 91.5%,which proves the effectiveness of the 2D-SGCN model.

关 键 词：方面级情感分析 图卷积神经网络 社交网络 情感词典 字词双通道 多维度融合 话题 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]