Research On Rumor Detection Based On Nested LSTM With Multi-Source Loss

In 2020,the New Coronavirus(COVID-19)epidemic has spread globally.Reports and dis-cussions on issues such as the source and spread of the COVID-19 are mixed with many statements that are not true.Rumors and conspiracy theories are rampant on social networks and some traditional media.The spread of rumors will not only affect the government's emergency response and cause economic losses,but also intensify the negative emotions of the receivers and even have a serious impact on the masses and society.Therefore,the detec-tion and recognition of Internet rumors have important practical significance.The research goal of the thesis is to realize the auto rumor detection of social networks such as Weibo and Twitter.Scholars have proposed a series of deep learning methods combined with feature engineer-ing to transform rumor detection into classification problems and have achieved remarkable results in recent years.However,most researchers separate tweets from users' comments and thus cannot effectively associate the post of the poster with all the comment content and the characteristics that change over time during the propagation process.Therefore,the paper aims to solve the problem of incomplete deep semantic features in mining large-scale tweets about rumors of existing deep learning models.The thesis proposes the Sentence and Words Position Embedding Based on Nested Long Short-Term Memory algorithm(SPNLSTM).The main research contents and contributions of the paper are as follows:1.For the problem that some existing rumor detection models ignore the important fea-tures of user comments over time,the SPNLSTM algorithm uses multisampling tech-nology to decompose large-scale text into small-scale samples.The algorithm makes full use of the time series correlation characteristics of the divided data and converts the newly generated time series data into time series source data.Meanwhile,the SPNLSTM algorithm uses Bi-directional Long Short-Term Memory as the basic unit to extract the semantic features of the temporal sequence to provide valuable informa-tion for the rumor detection work.2.Based on the phenomenon of "Alignment Disaster" that current models tend to ap-pear in variable-length texts,the SPNLSTM uses sentences as a unit and combines with automatic labeling technology to effectively mine both the deep context of each sentence of same users' tweets and the semantic relevance between the word segmen-tation within the same sentence.Further more,the model collects multi-source loss into the loss function to update the hyperparameters of the network.3.In the paper,the rumor detection algorithm is modeled on two sets of Chinese data sets and two sets of English data sets.The thesis adopts the evaluation metrics of accuracy,precision,and recall.The experiment shows that the SPNLSTM algorithm effectively reduces the hyperparameter scale of the network and achieves better results.The SPNLSTM algorithm has a strong generalization ability.