Research On Rumor Processing Based On Multi-task Learning

With the rapid development of the Internet and the extensive use of social platforms,social platforms such as Twitter,Facebook and Weibo,etc.have gradually become primary source of information for multitudinous users.However,due to the characteristics of social platforms,i.e.,easy to share and difficult to supervise in real time,a large number of unverified news(i.e.rumors)are widespread among users,which may,to some certain,have negative impact on social platforms,individuals' daily lives and even the entire society.Therefore,more and more attention has been paid to the rumor processing based on the social platforms.It has become a hot research topic to detect rumors and determine the veracity of the rumors early on the social platform to minimize the effects of the rumor spreading.Rumor processing can be divided into four sub-tasks: rumor detection,rumor tracking,stance classification,and rumor verification.Each sub-task can be studied separately,but with the development of multi-task learning,the excellent performance of jointly training related tasks in many fields has been proved.Especially,the multi-task learning based rumor processing has also witnessed great development.This thesis focuses on the task of rumors verification,which aims at verify and divide rumors into three categories according to their veracity: True rumors(T),False rumors(F)and Unverified(U).The main research contributions of this thesis are given as follows:(1)A new rumor verification model based on multi-task learning named RV-ML,is proposed to improve the accuracy of rumor verification.Specifically,RV-ML uses the task of stance classification as auxiliary to rumor verification.In RV-ML framework,the shared long short-term memory(LSTM)layer for both tasks is firstly designed to learn the shared representation of the two tasks.Then,for rumor verification,a convolution neural network(CNN)with varying filter window sizes is utilized to capture multiple local features from the high-level representation obtained by the shared layer.Then the fully connection layer completes the tasks of stance classification and rumor verification respectively.Compared with a single-task rumor verification model and two existing multi-task rumor verification models(MTL and MT-ES),experiments on two real rumor datasets PHEME and Rumour Eval,show that the RV-ML is superior to these benchmark schemes in terms of accuracy,macro F1 and F1 scores for each category.(2)Considering that the attention mechanism is widely used in Natural Language Processing,and especially its variant self-attention mechanism can effectively learn the internal dependence of the text,our work intentionally adds the module of self-attention mechanism after the shared LSTM layer in RV-ML,which adaptively adjusts the weight of shared representation extracted from LSTM,and better learn the semantic feature relationship.Experiments on PHEME and Rumour Eval datasets illustrate that the modified model makes the better performance in rumor verification.(3)Based on the observation although the goal of CNN in rumor verification specific task is to extract local features of different distances in the shared representation,the local features are limited by the window size of filters,and in contrast,the self-attention mechanism has no such distance limit.Therefore multi-head attention mechanism,which integrated by multiple self-attentions,is used to replace the CNN module in the task of rumor verification.Experimental results on PHEME and Rumour Eval illustrate that our improvement to RV-ML achieves better performance in accuracy,Macro F1 and F1 scores for each category of rumor veracity.