基于多任务学习的微博谣言检测方法

Microblog Rumor Detection Method Based on Multi-task Learning

微博谣言的广泛传播给当今社会造成了日益严峻的负面影响。基于深度神经网络的方法存在缺少大量带标签的数据。研究发现,谣言经常伴随负面情感,而非谣言则伴随正面情感,考虑到谣言与非谣言之间表现出的相反情感倾向性,提出一种将谣言检测和情感分析这两个高度相关的任务结合起来学习的多任务学习方法,为了尽可能多地挖掘不同任务之间的关联,全面分析谣言检测任务的特征,设计了一个由BERT和BiGRU联合的多任务学习框架(BERT-BiGRU-MTL,BBiGM)。利用权值共享的方法对两个任务进行联合训练,同时提取出任务之间的共同特征和针对谣言检测任务的特定特征,利用情感分析任务辅助谣言检测。研究结果表明,该方法在准确率、精确率、F1值评测指标上优于采用单任务学习的方法。

The widespread dissemination of Weibo rumors have caused an increasingly severe negative impact on today’s society.The method based on deep neural network has the problem of lack of a large amount of labeled data.The research has found that rumors are often accompanied by negative emotions,while non-rumors are accompanied by positive emo-tions.Taking into account the opposite emotional tendencies between rumors and non-rumors,a method is proposed to highly correlate rumors detection and sentiment analysis.The multi-task learning method that combines the tasks of BERT and BiGRU,in order to mine as many associations between different tasks as possible,and comprehensively analyze the characteristics of the rumor detection task,a multi-task learning framework(BERT-BiGRU-MTL,BBiGM).The weight sharing method is used to jointly train the two tasks,and at the same time,the common features between the tasks and the specific features for the rumor detection task are extracted,and the sentiment analysis task is used to assist the rumor detec-tion.The research results show that this method is better than the single-task learning method in terms of accuracy,preci-sion and F1 value evaluation index.