基于迁移学习的微博谣言检测方法  

Detection method of micro blog rumor based on transfer learning

作　　者：沈瑞琳 潘伟民[1] 张海军[1] SHEN Rui-lin;PAN Wei-min;ZHANG Hai-jun(School of Computer Science and Technology,Xinjiang Normal University,Urumqi 830054,China)

机构地区：[1]新疆师范大学计算机科学技术学院,新疆乌鲁木齐830054

出　　处：《计算机工程与设计》2021年第12期3534-3539,共6页Computer Engineering and Design

基　　金：2020年新疆维吾尔自治区研究生教育改革创新计划基金项目(XJ2020G235);新疆师范大学重点实验室基金项目(XJNUSYS2019B13);国家自然科学基金委NSFC-新疆联合基金重点支持基金项目(U1703261)。

摘　　要：为解决基于深度神经网络的微博谣言检测工作中带标签数据稀缺的问题,提出一种基于迁移学习的微博谣言检测方法。利用双层双向的门控循环单元和卷积神经网络组成的联合模型作为特征提取器,利用丰富的评论数据对联合神经网络进行预训练,将训练好的特征提取层迁移到微博谣言检测任务中,通过区分微调和斜三角学习率两种微调策略对特征提取层进行调整,使其适应于目标任务。实验结果表明,采用迁移学习方法的联合神经网络能有效提高微博谣言检测的准确率。To solve the problem of the scarcity of tagged data in the detection of micro blog rumors based on deep neural network,a method of micro blog rumors detection based on transfer learning was proposed.The joint model composed of two-layer bi-directional gating loop unit and convolutional neural network was used as feature extractor,rich review data were used to pre train the joint neural network,the trained feature extraction layer was migrated to the micro blog rumor detection task,and the feature extraction layer was adjusted by distinguishing the fine-tuning and the oblique triangle learning rate to make it suitable for the target task.Experimental results show that the combined neural network based on transfer learning method can improve the accuracy of rumor detection effectively.

关 键 词：谣言检测 迁移学习 微博 深度神经网络 预训练 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]