| The openness and convenience of social media reduce the cost of writing and sharing false information,which poses a great threat to national security,social stability and the ecological of information systems.Moreover,information on social media is characterized by a large amount,high deceptive nature and multiple modalities,manual detection has limitations in timeliness,coverage and effectiveness.Therefore,it is important to study automated false information detection technology.Nowadays,the model based on deep learning has achieved superior results.However,on the one hand,the black box of deep learning makes the model lack transparency and evidence supporting.On the other hand,detecting false information requires faithfulness,which raises the requirement for the credibility of the algorithm.How to deal with the contradiction between the two and how to improve the effectiveness of false information by using the tedious and complicated clues are the main challenges faced by the work.To tackle the above challenges,the work aims to make full use of all kinds of clues related to the post and mine the implicit correlation between the clues,to improve the performance and trusted of the false information detection algorithm.And our work follows the requirements of the trusted AI proposed by International Business Machines Corporation(IBM)to improve the interpretability and robustness of the algorithm,and carries out the following research.Firstly,in order to capture and model the semantic relationship of information posts and user comments,we propose an explainable false information detection approach based on the signed-attention mechanism.The method applies the Semantic Hilbert Space to simulate quantum-like phenomena in human language,and designs the signed-attention mechanism to simultaneously capture the relationship between the post and comments in terms of stance and importance,so as to improve the transparency and post-hoc explainability of the model.Secondly,aiming at the implicit relationship between information text content and visual content,the paper proposes a multi-modal false information detection technology based on adversarial training.The visual-text pre-training model and co-visual-text multi-head attention mechanism are used to fully explore the relationship between different levels of information text and visual cues.Besides,three kinds of malicious perturbations are designed for multi-modal data,and adversarial training is used to improve the model’s robustness.For each model,the paper conducts experiments on two real-world datasets,and the experimental results show the effectiveness of the proposed methods.Finally,based on the above methods,we design and implement a trustworthy false information classification system.The system includes the complete user interaction in the process from data processing,model training to result display.And users can adjust the parameters to build a model suitable for their tasks according to specific needs. |
