Incorporating Factual Knowledge For Rumor-Debunking Text Generation Technology

The wide propagation of rumors is extremely harmful to society,as it violates the individual rights of citizens or social organizations,causes social panic,undermines the social trust system,and even endangers national security.Many researchers have investigated relevant methods of rumor-debunking,but the current work still requires a lot of manual labor,existing problems of burdensome work content and low efficiency of rumor-debunking.The rumor-debunking text generation method can greatly reduce the labor and time cost of the rumor-debunking work,and it plays a critical role to automatically restrain the propagation of rumors in time and effectively reduce the social damage of rumors.However,the effective method of rumor-debunking is still underexplored.The difficulties in rumor-debunking text generation are as following:(1)The rumor-debunking text generation is highly dependent on external knowledge,and the information of the rumor itself cannot support the rumor-debunking text generation.(2)Rumor-debunking text is a logical rumor-debunking explanation,remaining difficulties in the long text generation.(3)Currently,there is no corpus design criterion for the task of rumor-debunking text generation.Aiming the above difficulties,the main research contents and innovations are as follows:1.Aiming at the problem that rumor-debunking text generation is highly dependent on external knowledge,this thesis introduces external knowledge in the form of knowledge triples and proposes a knowledge triples-based knowledge text generation method,which pays more attention to the internal elements and relative position information of knowledge triples.Experimental results show that the proposed method is superior to the classical Seq2seq method.2.Aiming at the problem of difficulty in long text generation,this thesis proposes a two-step method via knowledge constraint.It not only alleviates the problem of difficulty in long text generation,but also makes the generated rumor-debunking text more logical.Experimental results show that the two-step method achieves a BLEU score of 35.01,which is higher than the one-step GPT-2 method by 19.96 points.3.For the first time,this thesis proposes a corpus design criterion for the task of rumor-debunking text generation.The criterion takes full account of external knowledge dependence of rumor-debunking text generation,and defines that the dataset for this task consists of "rumor text","knowledge text" and "rumor-debunking conclusion".