A Study On Style-based Multi-task Rumor Detection Method

Nowadays,access to information through the We Chat and other online social media has gradually replaced the traditional media to view news and has become a habit of more and more people,who can get all kinds of information anytime and anywhere.However,rumors can also spread rapidly and widely through the developed online social media,thus endangering the stability of the country and society.Studies have found that,in addition to the textual content of rumors,their style also differs from that of real information,for example,rumors are generally shorter than real information and use more words with strong emotions.At the same time,a series of social background information generated during the rumor transmission process can also be used for judgment.Therefore,inspired by multi-task rumor learning,this dissertation focuses on multi-task rumor detection,rumor style and social background information,which are related as follows:(1)Design and validation of style feature representation and social context information feature representation.In this dissertation,we design and define a style feature representation for rumors,and then use a Bi-LSTM to verify its validity,and conduct experiments on different datasets to show that the style features designed in this dissertation are valid.In this dissertation,we also design a social background information feature representation based on related research,and also use a Bi-LSTM network to verify its validity,and its experimental results on different data sets also prove its validity.(2)A multi-task rumor detection approach based on style and hierarchical attention mechanism.A multi-task rumor detection model SC-HA-MTL based on style and hierarchical attention mechanism is proposed by combining the text content and style-based rumor detection tasks in a multi-task learning approach for deeper information interaction and using a hierarchical attention mechanism to mitigate the effect of noise.The experimental results demonstrate the effectiveness and generality of the style and hierarchical attention-based multitask rumor detection model on both English and Chinese datasets by comparing it with existing state-of-the-art detection models on the Pheme and Weibo datasets.(3)A multi-information dual-sharing rumor detection method.Based on the SC-HA-MTL and the validation of the social context information feature representation,two parallel information sharing modules are added to the social context information-based rumor detection task and two information sharing methods are designed to implement the task centered on the textual content-based rumor detection task,finally forming a multi-information dual-sharing based rumor detection model MIDS.The results show that the MIDS can achieve better detection results than the advanced multitasking rumor detection algorithm and the communication structure-based detection algorithm.