首页 | 本学科首页   官方微博 | 高级检索  
     

多任务交互式学习网络的方面情感分析
引用本文:宋婷,潘理虎,陈战伟. 多任务交互式学习网络的方面情感分析[J]. 计算机工程与应用, 2022, 58(19): 202-208. DOI: 10.3778/j.issn.1002-8331.2203-0485
作者姓名:宋婷  潘理虎  陈战伟
作者单位:1.太原科技大学 计算机科学与技术学院,太原 0300242.中国移动通信集团山西有限公司,太原 030001
摘    要:方面情感分析传统方法采用方面词抽取-情感预测的独立学习模式,未充分利用两模块的联合信息及训练过程中有价值的信息。提出基于消息传递机制的多任务交互式学习网络,模型采用细粒度属性级分类任务和篇章级分类任务联合训练,设计消息传递显式地对任务交互进行建模,通过共享隐藏变量迭代传递信息,有助于特征学习和推理。方面情感分析模块提出词级信息交互机制以及观点词抽取——情感预测信息传递通道,实现双注意力机制;利用池化操作嵌入多层GRU网络实现篇章级任务预测。设计迭代算法在方面级和篇章级任务间交替训练,通过三个数据集上的实验对比,结果表明模型在每个子任务的F1分数、模型整体性能、篇章级任务网络性能上均得到有效提高。

关 键 词:方面情感  多任务交互  消息传递机制  词级交互  双注意力  

Aspect Sentiment Analysis Based on Multi-Task Interactive Learning Network
SONG Ting,PAN Lihu,CHEN Zhanwei. Aspect Sentiment Analysis Based on Multi-Task Interactive Learning Network[J]. Computer Engineering and Applications, 2022, 58(19): 202-208. DOI: 10.3778/j.issn.1002-8331.2203-0485
Authors:SONG Ting  PAN Lihu  CHEN Zhanwei
Affiliation:1.School of Computer Science and Technology, Taiyuan University of Science and Technology, Taiyuan 030024, China2.China Mobile Communications Group Shanxi Co., Ltd., Taiyuan 030001, China
Abstract:The traditional method of aspect-based sentiment analysis adopts the independent model of aspect-word extraction and emotion prediction, which does not fully utilize the joint information and the valuable information in the training process. Multi-tasking interactive learning network based on message passing mechanism is proposed, which using fine-grained property level task and chapter level task joint training. A message passing mechanism explicitly to modeling of interactions between tasks is designed, by sharing hidden variable iteration pass information. Word-level information interaction module and aspect opinion-word extraction/aspect sentiment prediction information transmission channel are designed to realize the dual attention mechanism. Multi layer GRU network combined with pooling operation is used to realize chapter level task prediction. Iterative algorithm is designed on three data sets to train alternately between aspect and chapter tasks. The effectiveness and feasibility of the model are verified by comparing the F1 score of each task, the overall performance of the model, and the network performance of the discourse level task.
Keywords:aspect emotion,multi-task interaction  message passing mechanism  word-level interaction,dual attention,
点击此处可从《计算机工程与应用》浏览原始摘要信息
点击此处可从《计算机工程与应用》下载免费的PDF全文
设为首页 | 免责声明 | 关于勤云 | 加入收藏

Copyright©北京勤云科技发展有限公司  京ICP备09084417号