Generalising Across Domains in Video Misinformation Detection

This paper proposes and demonstrates the application of pre training and transfer learning in a deep neural network model to identify misinformation in YouTube videos, based on their auto-generated captions. Pre-training, in this context, is the training of smaller sub-models on domain-specific data, before those smaller models, together with all trained weights, are integrated together into one larger model. Transfer learning takes place when the larger model is able to utilise the knowledge gained by sub-models to generate new conclusions. The given model also utilises pre-trained Bidirectional Encoder Representations from Transformers (BERT) layers, further utilising the idea of pre-training and leveraging the concept of transfer learning to maximise efficacy. An appreciable improvement in overall quality is demonstrated when the proposed method is used, as compared to a naive model built to tackle this same problem. The goal of this research is to better understand the extent to which the given techniques and model architectures contribute to effective classification of misinformation across a variety of different domains.