A simple web application demonstrating how a sentiment analysis model can be deployed as an email moderation/filtering system.

Background

The model, a Convolutional Neural Network, was trained by myself on the toxic comments dataset using Google's pre-trained 300-dimensional vector embeddings. GlobalMaxPooling and a generator were used so that sequences of varying length could be processed without padding.

Instructions for testing

If you'd like to test the application yourself, you'll need the "GoogleNews" vectors (GoogleNews-vectors-negative300.bin). You can download them from Kaggle, although I'm not sure if it's the same version that I used for training (obviously, if it's a different version, this might affect model performance).

The following Python packages are required:

• numpy
• pandas
• nltk (for natural language processing)
• keras (to load pre-trained model)
• tensorflow (keras backend)
• gensim (to map words to google's pre-trained vectors)
• wtforms (to build the form for the message)
• flask (to build web applications in Python)
• flask-wtf (form handling extension)
• flask-mail (email extension)
• os + python-dotenv (if you'd like to store your email credentials as environment variables)

The next steps are as follows:

1. clone the repository
2. save the binary file with the google vectors in the model_resources folder
3. replace the email configuration details/credentials in webapp.py with your own
4. create a secret key for CSRF protection, and replace the one in webapp.py with your own

If you'd like to use Mailtrap for email testing (as I did), you can find an instruction on how to get your configuration details and credentials here.

Then run

python webapp.py

and open http://127.0.0.1 in a browser. Select one of the people listed under "current members" (currently, there are only 2 members stored in the database members.db). Complete the contact form on their profile page with your email address, a subject and a message, and click "submit". If your message is classified as "toxic", you'll receive a notification that the message could not be sent. The application will also inform you which types/severity levels of "toxicity" (e.g. "general toxicity" for label toxic) were detected. Otherwise, you should receive a notification that your message was sent, and find the email in your inbox.

If you need to get a lot of "toxicity" out of your system, feel free to try sending multiple nasty messages - however, you shouldn't be able to submit more than 3 such messages. Repeat offenders will have the form fields and submit button disabled.

References

Konstantinos Sechidis, Grigorios Tsoumakas, and Ioannis P. Vlahavas. 2011. On the Stratification of Multi-label Data. In Proceedings of the Machine Learning and Knowledge Discovery in Databases – European Conference, ECML PKDD (part III), pages 145–158, Athens, Greece.

Credits

This tutorial by Will Koehrsen helped me greatly in getting the app up and running.

Photo credits

• Customer icon by Icons8

• website background: photo by Steve Johnson