基于记忆迭代网络的多模态情感分析  

Multimodal sentiment analysis based on memory iterative network

作　　者：赖河蒗[1] 王丽丽 李玲俐[1] 许学添[1] 陈丽仪[1] LAI Helang;WANG Lili;LI Lingli;XU Xuetian;CHEN Liyi(Department of Information Administration,Guangdong Justice Police Vocational College,Guangzhou 510520,China;School of Basic Education,Guangdong Teachers College of Foreign Language and Arts,Guangzhou 510640,China)

机构地区：[1]广东司法警官职业学院信息管理系,广州510520 [2]广东省外语艺术职业学院基础教育学院,广州510640

出　　处：《智能计算机与应用》2024年第11期200-205,共6页Intelligent Computer and Applications

基　　金：广东省科技创新战略专项资金(大学生科技创新培育)项目(pdjh2024b593);广东省教育科学规划课题(高等教育专项)(2024GXJK989);广东省普通高校特色创新类项目(2023KTSCX250,2024WTSCX300,2023KTSCX295);广东省第二批高职院校高水平专业群建设项目(GSPZYQ2021123);广东司法警官职业学院第五届院级课题(2023YB08)。

摘　　要：在开展多模态情感分析的过程中,需要考虑2方面的关键信息:一是每种模态的内部信息,二是各模态之间的交互信息。针对如何有效地捕获以上核心信息的问题,提出一种记忆迭代网络模型。将每种模态的序列数据分别输入到其各自的循环神经网络以及注意力机制模块,处理后将所有模态序列级联并送到记忆迭代模块进行迭代处理,迭代结束后将数据经过完全连接层,降低维度后进行分类。在公开的数据集CMU-MOSI、ICT-MMMO以及YouTube上进行的实验表明,记忆迭代网络模型的准确率和F1值均有所提升。In the process of multimodal sentiment analysis,two key aspects need to be considered:one is the internal information of each modality,and the other is the interaction information between each modality.Aiming at the problem of how to effectively capture the above core information,a memory iterative network model is proposed.Each modal sequence data is input into its own recurrent neural network and attention mechanism module respectively.After processing,all modal sequences are concatenated and sent to the memory iteration module for iterative processing.After the iteration,the data is passed through the fully connected layer and classified after reducing the dimension.Experiments on the public datasets CMU-MOSI,ICT-MMMO,and YouTube show that both the accuracy and F1 value of the memory iterative network model are improved.

关 键 词：记忆网络 迭代 注意力机制 多模态 情感分析 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]