Robust Semi-Supervised Fake News Recognition by Effective Augmentations and Ensemble of Diverse Deep Learners

Nowadays, most people obtain information from social media networks, where news accompanied by photos and videos attracts readers more than traditional ones. However, these advantages are often misused some publishers to disseminate fake rapidly, thereby adversely affecting individuals societies. Thus, the early detection of posts is crucial. Developing an automatic content-based detector ideal way overcome this issue. Given that generation rate in drastic labeling a huge amount data required fully supervised models expensive time consuming, not beneficial real applications. To address limitation, study presents semi-supervised method utilizing ensemble diverse deep learners, effective augmentations, distribution-aware pseudo-labeling technique. Here, proposed hybrid loss function enforces learners have accurate classification performance while attending different parts content. Moreover, augmentations enhance robustness prevent overfitting effectively. Diverse utilized annotate unlabeled accurately update their parameters confident predicted curriculum way, enhancing quality pseudo labels model. we utilize encoded sentences pre-trained transformer models, such as XLNET, parameter sharing build light on common feature extractor module. Consequently, number less existing methods, experiments conducted three public datasets reveal consistently outperforms state-of-the-art with proportions labeled across all evaluated datasets.