Automated Fake News Mitigation Using Reinforcement Learning and Large Language Models

Fake news on social networks poses significant threats to information integrity and societal stability. From influencing political outcomes to inciting real-world violence, the consequences of fake news underscore the urgent need for effective counter-approaches. While efforts from journalists, fact-checking organizations, social network platforms, and computer scientists have made significant advances in combating fake news, two major challenges persist. First, the dynamic nature of social networks presents challenges to developing effective mitigation strategies, with limited studies on automatic fake news mitigation and understanding of information propagation. Second, the need for large-scale fact-checking services highlights the limitations of manual approaches and the necessity for automated solutions to generate counter-responses to fake news claims. Existing research has primarily focused on veracity assessment, with limited studies on generating counter-responses. To address these challenges, we investigate the possible improvements of automated fake news mitigation methods from two perspectives: network-oriented and content-oriented fake news mitigation. For network-oriented mitigation, several mitigation policies have been proposed by investigating the underlying information propagation and the challenges when applying to real-world mitigation; for content-oriented mitigation, a large language model based fake news mitigation approach has been proposed, aiming for generating evidence-grounded counter-responses to fake news claims. In this thesis, three significant contributions to the field of automated fake news mitigation have been proposed and are summarized as follows. First, we propose a reinforcement learning framework to learn a mitigation policy that selects multiple debunkers dynamically within budget for each stage so that the selected debunkers can maximize the overall cumulative mitigation effect across stages. To address the issue of selecting debunkers from an exponentially large search space, we proposed DQN-FSP, a novel approach that extends deep Q-networks with future state prediction. This method optimizes the selection of debunkers in multi-stage campaigns, maximizing cumulative mitigation effects while minimizing overlap under budget constraints. DQN-FSP significantly outperforms existing baselines in identifying cost-effective debunkers, demonstrating its efficacy in large-scale social network environments. Second, we propose a realistic setting for reinforcement learning-driven fake news mitigation approaches where the reward to decisions is episodic and utilize self-imitation learning to optimize the mitigation policy. We further improve the current self-imitation learning approach by introducing NAGASIL, which uniquely harnesses both positive and negative experiences and incorporates augmented states to better capture the environment dynamics. It overcomes the key challenge of interleaved propagation effects, leading to more robust and effective debunker selection strategies in practical applications. Finally, we propose to leverage the power of large language models to generate tailored, evidence-grounded counter-responses to fake news claims. To enhance the factual accuracy of LLM-generated content, we propose a fine-grained text-based critique feedback framework called MisMitiFact. This framework employs an ensemble of models trained on faithfulness-checking tasks. The training data is auto-generated from factual descriptions of related evidence collected from readily available fact-checking sites. MisMitiFact significantly improves the reliability of automated approaches for generating responses to counter fake news. Extensive experiments on both synthetic and real-world social network datasets validate the superiority of our proposed methods over state-of-the-art baselines in terms of mitigation effectiveness and content faithfulness. By proposing more efficient network-oriented mitigation approaches and LLM-based counter-response generation techniques, this research contributes to the development of more efficient and accurate fake news mitigation strategies. These advancements represent significant steps toward safeguarding the integrity of information ecosystems against the threat of fake news.<p></p>