False textual information detection, a deep learning approach

Many approaches exist for analysing fact checking for fake news identification, which is the focus of this thesis. Current approaches still perform badly on a large scale due to a lack of authority, or insufficient evidence, or in certain cases reliance on a single piece of evidence.
To address the lack of evidence and the inability of models to generalise across domains, we propose a style-aware model for detecting false information and improving existing performance. We discovered that our model was effective at detecting false information when we evaluated its generalisation ability using news articles and Twitter corpora.
We then propose to improve fact checking performance by incorporating warrants. We developed a highly efficient prediction model based on the results and demonstrated that incorporating is beneficial for fact checking. Due to a lack of external warrant data, we develop a novel model for generating warrants that aid in determining the credibility of a claim. The results indicate that when a pre-trained language model is combined with a multi-agent model, high-quality, diverse warrants are generated that contribute to task performance improvement.
To resolve a biased opinion and making rational judgments, we propose a model that can generate multiple perspectives on the claim. Experiments confirm that our Perspectives Generation model allows for the generation of diverse perspectives with a higher degree of quality and diversity than any other baseline model.
Additionally, we propose to improve the model's detection capability by generating an explainable alternative factual claim assisting the reader in identifying subtle issues that result in factual errors. The examination demonstrates that it does indeed increase the veracity of the claim.
Finally, current research has focused on stance detection and fact checking separately, we propose a unified model that integrates both tasks. Classification results demonstrate that our proposed model outperforms state-of-the-art methods.