Deep learning for rumour detection and claim veracity assessment on social media

Abstract

Online social networks, such as Twitter, Facebook, and Weibo, have become crucial platforms for news consumption, but they are also prone to the rapid spread of misinformation, leading to public deception. Therefore, the automatic detection and verification of rumours play a vital role in safeguarding society's trust. This thesis investigates deep learning approaches for rumour detection and claim veracity assessment on social media, encompassing multimodal source-based rumour detection, user credibility-enhanced rumour detection, propagation graph-based rumour verification, and the incorporation of external evidence for veracity assessment. First, we investigate multimodal rumour detection by focusing on classifying user-generated image-text pairs on social media. To handle the diverse multimedia content, we introduce a novel model called the Crossmodal Bipolar Attention Network (CBAN), which incorporates both positive and negative attention mechanisms. Experimental results demonstrate the superior performance of CBAN, compared to existing methods for multimodal rumour detection. Additionally, the proposed CBAN has shown promising performance in other multimodal image-text classification tasks, including sentiment analysis, sarcasm detection, and hate-speech detection. Moreover, the thesis presents an early detection approach that utilizes textual claims and source author credibility to identify rumours. By leveraging pretrained language models and transforming author-aware rumour detection into a text classification problem, our proposed method enhances detection accuracy. Additionally, we introduce a Layer-Wise Parameter-Efficient Tuning (LWPET) strategy to optimize pretrained language model parameters, reducing computation and memory requirements during fine-tuning. In the pursuit of an efficient stream classification framework for early fine-grained rumour classification based on community response, we introduce the Causal Diffused Graph-Transformer Network (CDGTN). CDGTN incorporates Source-Guided Incremental Attention Pooling (SGIAP) and a Stacked Early Classification Loss (SecLoss) to improve early classification effectiveness. Furthermore, we propose a continued inference algorithm based on prefix-sum to enhance efficiency. Experimental results on multiple datasets confirm the effectiveness and efficiency of CDGTN. To address the challenge of assessing the veracity of claims on social media, particularly those lacking contextual information, we propose the Dual-Stream Cross-Attention Network (DSCAN). DSCAN combines social response and external evidence using a dual attention mechanism. Experimental results demonstrate the significant performance improvement of DSCAN, which is evaluated on extended datasets containing relevant evidence retrieved from web search engines. Lastly, this thesis explores the integration of recent conversational-based instruction-following language models with external evidence retrieval for fact-checking purposes. This improves the accessibility of the fact-checking system to more general use. By leveraging search engines to retrieve evidence and enhancing the knowledge of a pretrained language model, our approach, called FactLLaMA, achieves state-of-the-art performance in fact-checking tasks by bridging the gap between model knowledge and up-to-date information. In summary, the research presented in this thesis contributes significantly to the field of rumour claim detection and claim veracity assessment on social media. The proposed deep learning techniques and models demonstrate their effectiveness in addressing key challenges, outperforming existing methods on various benchmark datasets. These contributions have important implications for combating misinformation and promoting the dissemination of accurate information on online platforms.