As the US stock market tumbled most this year due to Covid-19 fears. Because there is news and social media to quickly spread the information. Social media can act as a fastest medium to spread information. But sometimes the news is just fake. So in order to curb such flow of information. We bring in a system that will combat this from the very root.
What it does
Our application using Pytorch to train a model over certain text data. Which in future could be a set of Facebook posts. From training the model what we try to achieve is the ability to determine the posts that tend to distribute false emotions. So the usecase as user posts on Facebook then the post content will go through our model and it will help label toxic or non-toxic. So if the information is toxic that means it is trying to arouse something unwanted among the masses. Then the most could be labeled for the readers and then it could be up to them to decide whether they want to react to it.
How we built it
Using pytorch we trained upon the dataset we had from opensource. We used BERT model which gave us an accuracy of 94.31. Then we used a FastText model to train on the dataset and that gave us an accuracy of 90.4. Then we ran it on a Python Flask application so that we can conveniently access the model from our browser.
Challenges we ran into
Find the dataset for the project. For it to match the exact usecase
Accomplishments that we're proud of
We could build a prototype and deploy to heroku
What we learned
We learnt about Pytorch. How the text on the social media flows. We also learnt the effectiveness of quality of news on social media.
What's next for False Posts
Next we would like to use ensemble to combine these models. Then we would try and integrate with Facebook posts API and then see how we deliver promising results from the content of Facebook posts.