摘要
Aspect-BasedSentimentClassification(ABSC)属于细粒度情感分析任务之一,旨在发现实体方面(Aspect)相关的情感倾向.本论文中提出一种基于胶囊网络的模型:MADC(Model based on Asp-Routing and Doc-Routing Capsule),通过迁移模型将文档级别的特征和语义信息用于方面级情感分析中,针对文档级别和句子级别的的任务,分别使用了基于注意力机制的AspRouting和Doc-Routing动态路由方法,加强了句子级别任务情感分析的可信度.为了让模型识别特定领域词向量的语义信息,文章使用双嵌入词向量加位置信息的表示方法,通过卷积神经网络抽取特征作为胶囊网络的输入,再使用两层动态路由算法使网络共享迁移学习的特征胶囊和主胶囊,最后针对不同的任务使用不同的类胶囊输出向量对方面情感或文档级别情感作出极性预测.文章通过在数据集上与多个框架的对比论证了模型的有效性.
Aspect-Based Sentient Classification(ABSC)is one of the fine-grained sentimental analysis tasks,w hich aims to discover the sentimental polarity related to aspect.In this paper,a model named M ADC(M odel based on ASP-Routing and DOC-Routing Capsule)based on capsule network is proposed.The feature and semantic information of document level are utilized in aspect based sentiment analysis through transform model.For document-level and aspect-level tasks,ASP-Routing and DOC-Routing dynamic routing methods based on attention mechanism are respectively used to enhance the credibility of sentence level task sentiment analysis.First of all,in order to make the model recognize the semantic information of in-domain embedding,this paper uses the representation method of double embedded w ord vectors and position information to extract features as the input of capsule network through convolutional neural network.Then we use tw o-layer dynamic routing algorithm to make the network share the feature capsule and the main capsule of transfer learning.Finally,different class capsules are used for different tasks to predict the polarity of aspect-level and document-level sentimental classification.This paper demonstrates the validity of the model by experiment with several frameworks.
作者
滕磊
严馨
徐广义
周枫
邓忠莹
TENG Lei;YAN Xin;XU Guang-yi;ZHOU Feng;DENG Zhong-ying(Kunming University of Science and Technology,Faculty of Information Engineering and Automation,Kunming 650500,China;Kunming University of Science and Technology,Yunnan Key Laboratory of Artificial Intelligence,Kunming 650500,China;Yunnan Nantian Electronic Information Industry Co.,Ltd,Kunming 650500,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2020年第12期2550-2556,共7页
Journal of Chinese Computer Systems
基金
国家自然科学基金地区科学基金项目(61462055,61562049)资助。
关键词
方面情感分析
胶囊网络
双嵌入
卷积神经网络
动态路由
迁移学习
aspect sentiment analysis
capsule network
double embedding
convolutional neural network
dynamic route
transfer learning
