Misinformation detection using multitask learning with mutual learning for novelty detection and emotion recognition

摘要

Fake news or misinformation is the information or stories intentionally created to deceive or mislead the readers. Nowadays, social media platforms have become the ripe grounds for misinformation, spreading them in a few minutes, which led to chaos, panic, and potential health hazards among people. The rapid dissemination and a prolific rise in the spread of fake news and misinformation create the most time-critical challenges for the Natural Language Processing (NLP) community. Relevant literature reveals that the presence of an element of surprise in the story is a strong driving force for the rapid dissemination of misinformation, which attracts immediate attention and invokes strong emotional stimulus in the reader. False stories or fake information are written to arouse interest and activate the emotions of people to spread it. Thus, false stories have a higher level of novelty and emotional content than true stories. Hence, Novelty of the news item and recognizing the Emotional state of the reader after reading the item seems two key tasks to tightly couple with misinformation Detection. Previous literature did not explore misinformation detection with mutual learning for novelty detection and emotion recognition to the best of our knowledge. Our current work argues that joint learning of novelty and emotion from the target text makes a strong case for misinformation detection. In this paper, we propose a deep multitask learning framework that jointly performs novelty detection, emotion recognition, and misinformation detection. Our deep multitask model achieves state-of-the-art (SOTA) performance for fake news detection on four benchmark datasets, viz. ByteDance, FNC, Covid-Stance and FNID with 7.73%, 3.69%, 7.95% and 13.38% accuracy gain, respectively. The evaluation shows that our multitask learning framework improves the performance over the single-task framework for four datasets with 7.8%, 28.62%, 11.46%, and 15.66% overall accuracy gain. We claim that textual novelty and emotion are the two key aspects to consider while developing an automatic fake news detection mechanism.