What's Purify?





Purify allows to spot and highlight the posts on social media that falls into one of the predefined categories:

Discrediting
,
False Media
,
False Supporting
or
Propaganda
.

The detailed information about the Facebook fact-checking policy can be viewed at the link.


Goal

The goal of the project is to protect the online community from manipulated content and create an additional obstacle against disinformation campaigns.

Development

The application was developed during the hackathon - Hacking the Fake News, organized by Digital Communication Network, Myth Detector, ForSet, and Media Development Foundation.

How Does it Work?

Model

Model is a collection of interconnected decision gates that can classify the text between two categories. To learn how to distinguish the sentences, the model is trained using the examples of pre-labeled sentences, that were categorized by researchers working at fact-checking organizations.

The model is based on ELECTRA architecture,
the final version of the fine-tuned model can be accessed here , with some examples.

Data

To fine-tune the model 15 000 positive and 92 000 negative posts had been collected from fact-checking organizations operating across Georgia: factcheck.ge disinfoobserver.ge mythdetector.ge mediachecker.ge

Browser Extension

The browser extension reads the Facebook feed and checks whether posts, fall into predefined categories. If the model's prediction is positive, the tag appears above the post, highlighting the predicted category.

Your Feedback

Your feedback would be highly appreciated, as it would further increase the accuracy of the model.

In case you agree with the prediction of a model, you can press the 'feedback > agree' button, on the right side, of the prediction, and press 'disagree' in case you think the predictions are not accurate.

The model is constantly been retrained, using the data collected through your feedback.

Code

The code for the application is open for contributions on Github.