Fake news detection is still one major concern in post-pandemic era. As the technological barrier is way lower than before, everyone can publish we-media contents. Detecting fake news is therefore much more important currently. This gives rise to our incentive to construct and train the model to automatically detect the fake news.
The train set has 6420 entries of news. The test set has 2110 entries of news. Both train and test set has a distribution of roughly 50% real and 50% fake to avoid bias and overfit.
First, we will remove the website information and use regular expression to extract the related words in lower case as data preprocessing before embedding.
GloVe: combine the advantages of word2vec and count based. TF-IDF
traditional machine learning: SVM neural network model: Transformer