Fake News Classification
An exploration of fake news classification and the available datasets for the task
Abstract
Accurate classification of news article as legitimate or containing fake information has seldom been a more prominent issue. In this project, a survey of available datasets in news classification was conducted and an evaluation of common language models on the task of text classification was done. In addition to typical classification models like Random Forests and XGBoost, increasingly complex language models and word embeddings such as Word2Vec and ELECTRA were compared and used to vectorize the input and classify articles. The ELECTRA and bi-directional LSTM models achieved the highest classification accuracy at 85%, however simpler models also showed reasonable performances.Among our conclusions, we also draw attention to some common issues in available datasets, some strengths and limitations of simpler word vectorization models like TFID and possible improvements combining simpler and state-of-the-art models.
Repository
Available here