Leveraging transfer learning for detecting misinformation on social media

Misinformation is a pervasive problem on social media platforms, with the potential to cause harm to individuals and society as a whole. Many computational techniques have been employed to tackle the misinformation in the online ecosystem. Curbing misinformation on online social networks is an important sub-field of this research. In this paper, we explore the use of deep learning to detect misinformation on Reddit, which is one of the popular online forums. The proposed approach combines natural language processing and deep learning techniques to identify posts that contain misinformation and flag them for further review. We evaluate our model on a benchmark dataset of Reddit posts and show that it is able to identify misinformation with a high degree of precision. We use transfer learning by incorporating transformer-generated embeddings in our system. The Transformer based model is fine-tuned on Stanford natural language Inference dataset. We subsequently use a simple deep neural network to classify the posts as true or fake. Our model outperforms the models reported in the literature on the unimodal text data, on the benchmark dataset Fakeddit.