一种用于基于方面情感分析的深度分层网络模型  

A Deep Hierarchical Neural Network Model for Aspect-based Sentiment Analysis

作　　者：刘全[1,2,3] 梁斌 徐进[1] 周倩[1] LIU Quan;LIANG Bin;XU Jin;ZHOU Qian(School of Computer Science and Technology,Soochow University,Suzhou,Jiangsu 215006;Collaborative Innovation Center of Novel Software Technology and Industrialization,Nanjing 210000;Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education,Jilin University,Changchun 130012)

机构地区：[1]苏州大学计算机科学与技术学院,江苏苏州215006 [2]软件新技术与产业化协同创新中心,南京210000 [3]吉林大学符号计算与知识工程教育部重点实验室,长春130012

出　　处：《计算机学报》2018年第12期2637-2652,共16页Chinese Journal of Computers

基　　金：国家自然科学基金(61272005;61303108;61373094;61472262;61502323;61502329;61772355);江苏省自然科学基金(BK2012616);江苏省高校自然科学研究项目(13KJB520020;16KJB520041);江苏省高等学校自然科学研究重大项目(18KJA520011);吉林大学符号计算与知识工程教育部重点实验室基金项目(93K172014K04);苏州市应用基础研究计划工业部分(SYG201422;SYG201308)资助~~

摘　　要：近年来,基于方面情感分析已成为自然语言处理领域的研究热点之一.结合注意力机制的深度网络模型在基于方面情感分析任务中取得了令人瞩目的成功,针对以独立句子作为网络模型输入的方法无法获取句子间相互关系,以及仅使用词语层注意力机制难以充分获取同一评论中句子间的相互联系等问题,提出一种结合区域卷积神经网络和分层长短期记忆网络(Regional Convolutional Neural Network-Hierarchical Long Short-Term Memory,RCNN-HLSTM)的深度分层网络模型用在基于方面情感分析任务中.该模型通过区域CNN既可以保留不同句子在评论中的时序关系也可以大大降低仅使用LSTM网络的时间代价.此外,该模型利用一个分层LSTM网络来获取待分类句子内部词语之间的相互联系,以及待分类句子和评论中其他句子之间的情感特征信息.通过词语层和句子层注意力机制能有效获取特定方面在句子中的局部特征和整个评论中的长距离依赖关系,弥补了仅使用词语层注意力机制的不足.最后在多种语言的不同领域数据集上进行实验,取得了比传统的深度网络模型、结合注意力机制的深度网络模型以及考虑句子间关系的双向分层LSTM网络模型更好的分类效果.Aspect-based sentiment analysis has become one of the research hotspots in the field of natural language processing （NLP） in recent years. Different from ordinary sentiment analysis, aspect-based sentiment classification is a fine - grained task of sentiment analysis in the field of NLP, which need to infer different sentiment polarity of different aspects in the same sentence, because there will be more than one aspect in the same sentence usually. Previous studies, in general, usually consider the independent sentence as the input of neural networks and only focus on the given aspect in each sentence in the process of training. These approaches, however, will ignore the long-distance dependency of the given aspect in the entire long text and cannot take full advantage of the context relations of different sentences across the same review, which is bad for those ambiguous sentences and short sentences of a review in the training process. To address these problems, this paper proposes a hierarchical model of combining regional convolutional neural network and hierarchical long short-term memory （HRCNN-LSTM） for the task of aspect-based sentiment classification on long text customer review. This approach can extract more feature information of independent sentence and the relations of different sentences in the whole review via combining regional CNN and hierarchical LSTM, and is able to infer the sentiment polarity of different aspects discriminatively without any external information such as semantic dependency parsing. We divide a long text review into several regions based on different targets of aspect in the sentences, and then we utilize a regional CNN to receive different independent regions of the review to extract the information across the entire review. This regional CNN is able to capture the long-distance dependency of the concerned aspect across the whole review and keep the order of different regions, as well as, save the training time of using LSTM network only. Meanwhile, dividing regions

关 键 词：深度学习 基于方面情感分析 分层模型 循环神经网络 卷积神经网络 注意力机 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]