This is a auto moderator that helps us to track rude comments on Stack Overflow in Russian
The app is live with the besic model ever: https://chat.stackexchange.com/rooms/73258/. Admin app is depoyed on http://benice.rudevs.ru/
The app is structured in the following way.
- We create a database and upload initial data.
- The initial data contains ~6K of comments (1.5K rude comments and the rest is normal ones).
- We build a logistic regression model and store it in the database.
- Each 10 minutes the app pulls new comments from Stack Exchange servers via public API and stores it locally in the database.
- Each 10 minutes the app analyses comments in the database with the logistic regression model and marks comments as rude if necessary.
- If a comment is marked as rude it will appear in an RSS feed
http://benice.rudevs.ru/comments/feed/. - The feed is added to a chat room. Anyone can monitor this chat and see if something wrong happens on our site.
I've built an additional “admin” module where we can
- See how the features are distributed among all analysed comments.
- See ROC for all models.
- Verify comments that were marked as rude. When there are more than 50 verified comments the model gets retrained and starts reanalysing all comments in the database.



