Better Social Media

Better Social Media is a project that attempts to build a Chrome extension that can filter out content across various categories on a user’s Twitter feed. The categories -- toxicity, profanity, deressiveness, and sexual explicitness -- represent sentiment or rhetorical properties that we consider the most salient and addressable problems within text analysis. The goal of this project is to address what we consider to be a large part of what makes current social media platforms toxic: echo chambers of negativity, anger, and despair.

Chrome Extension

To allow our models to interact with our actual Twitter feeds, we created a Chrome extension as our interface to do so. Users can use sliders to adjust the strictness of the filters, and the tweets with predicted scores that exceed the thresholds in any category will get whited out.

Models

Depressiveness

We collected 10,000 tweets using the tweepy package and annotated each tweet by ourselves. Then, we transformed the tweets into feature vectors using a TF-IDF Vectorizer. Finally, we trained a XgBoost classifier to detect the tweets' depressiveness. The notebook documenting our modeling process can be found here.

Toxicity, Profanity, and Sexually Explicitness

For the other three models, we used 80,000 Reddit comments pre-labeled using Perspective API as our training data. Then, using similar preprocessing and vectorizing methods, we trained three separate classifiers to detect these three properties in tweets. The notebook documenting our modeling process can be found here.

Deployed API

After we determined our final models, we packaged each of the TF-IDF vectorizers as well as the models into Pickle files to prepare them for reuse. Using Flask, we built a RESTful API that can be called in our Chrome extension. The API was then deployed on Heroku. There are two endpoints available for POST requests.

Depressiveness

Predictions for a tweet's depressiveness can be retrieved using a POST request:
POST https://better-social-media.herokuapp.com/api/depressive

Sample Input

{
    "tweet": "todays energy feels so depressive and ugly ugh I can’t wait the day to be over"
}

Sample Output

Returns prediction scaled from 0 to 1.

{
    "prediction": "0.98030853"
}

Toxicity, Profanity, and Sexually Explicitness

Predictions for a tweet's toxicity, profanity, and sexually explicitness can be retrieved using a POST request:
POST https://better-social-media.herokuapp.com/api/depressive

Sample Input

{
    "tweet": "3 stages in life 1.birth 2.the fuck is this? 3.death"
}

Sample Output

Returns prediction scaled from 1 to 10.

{
    "profanity": "10.0",
    "sexually": "7.0",
    "toxicity": "9.0"
}

Requirements

In order to run the api and models locally, install the required python packages using pip install -r requirements.txt after cloning the repository.

Project Authors

• Andy Chen
• Saurabh Budholiya
• Cameron Milne