purify

The extension uses a language model to predict whether or not they post on social media contains Fake News. It uses five categories to label the posts: Discrediting, Ideologized Propaganda, False Media, False Supporting, and Government Propaganda. In case if the post is classified as a positive, tag highlighting the corresponding category appears under it.

To prediction the labels of a post, the application uses a fine-tuned ELECTRA model for the Georgian language. The model itself is open source and is located at https://huggingface.co/jnz/electra-ka-fake-news-tagging

To fine-tune the model 15 000 positive and 92 000 negative posts had been collected from fact-checking organizations operating in Georgia, between them: https://disinfoobserver.ge https://mythdetector.ge

To maximize the model predictions, the extension can collect user feedback and retrain the model on collected entries.

The code for extension and API are both open source and at the following repositories:

Chrome Extension: https://github.com/purify-app/google-chrome-extention