LRQ-Fact: LLM-Generated Relevant Questions for Multimodal Fact-Checking

Human fact-checkers have specialized domain knowledge that allows them to formulate precise questions to verify information accuracy. However, this expert-driven approach is labor-intensive and is not scalable, especially when dealing with complex multimodal misinformation. In this paper, we propose a fully-automated framework, LRQ-Fact, for multimodal fact-checking. Firstly, the framework leverages Vision-Language Models (VLMs) and Large Language Models (LLMs) to generate comprehensive questions and answers for probing multimodal content. Next, a rule-based decision-maker module evaluates both the original content and the generated questions and answers to assess the overall veracity. Extensive experiments on two benchmarks show that LRQ-Fact improves detection accuracy for multimodal misinformation. Moreover, we evaluate its generalizability across different model backbones, offering valuable insights for further refinement.

A Semi-supervised Fake News Detection using Sentiment Encoding and LSTM with Self-Attention

Micro-blogs and cyber-space social networks are the main communication mediums to receive and share news nowadays. As a side effect, however, the networks can disseminate fake news that harms individuals and the society. Several methods have been developed to detect fake news, but the majority require large sets of manually labeled data to attain the application-level accuracy. Due to the strict privacy policies, the required data are often inaccessible or limited to some specific topics. On the other side, quite diverse and abundant unlabeled data on social media suggests that with a few labeled data, the problem of detecting fake news could be tackled via semi-supervised learning. Here, we propose a semi-supervised self-learning method in which a sentiment analysis is acquired by some state-of-the-art pretrained models. Our learning model is trained in a semi-supervised fashion and incorporates LSTM with self-attention layers. We benchmark our model on a dataset with 20,000 news content along with their feedback, which shows better performance in precision, recall, and measures compared to competitive methods in fake news detection.