Enhancing Rumour Detection: A Hybrid Deep Learning Approach with ELMo Embeddings & CNN

Microblogs have gained significant traction as a news media source in recent times. On the contrary, unsupported information that is widely shared on social media platforms has the potential to drive significant disruption by misguiding individuals. This is exemplified by synthetic text, commonly referred to as ’readfakes,’ which further spreads online disinformation. It is crucial to develop models that possess the ability to detect and prevent the spread of rumours. Linguistic characteristics are easily obtainable and serve as crucial attributes in the initial phases of dissemination. Concurrently, the choice of features is substantial with respect to the interpretability and performance of the classifier. This research presents a hybrid model that addresses the urgent requirement for an automated rumour detection system. The model incorporates deep learning, more specifically a Convolutional Neural Network (CNN), and ELMo embeddings that have undergone refinement using a Multinomial Naive Bayes classifier. Training and evaluating this model with the PHEME rumour dataset. The text is transformed into ELMo embeddings at the outset and then proceeds to traverse the CNN architecture in order to extract features. The optimized vector is then used to train the Multinomial Naive Bayes classifier, which is positioned in the CNN’s output layer, to categorize rumors.