A Discord bot which can classify comment sentiments on the fly.
This model can classify comment sentiments into 3 classes: Positive, Neutral and Negative.
Yelp dataset. Originally this dataset contains around 8 million samples of reviews collected by Yelp over the years. These samples have stars rating from 1 to 5. For my usecase, I have grouped the star ratings to 3 broad categories.
- 1 and 2 stars - Negative
- 3 stars - Neutral
- 4 and 5 stars - Positive
But since I had limited resources, I used 1.2 million reviews. Another reason why I used these many reviews was also to maintain equal number of samples for all three classes. I took 512k samples for each class.
This is their distribution:
The model used here is a combination of 1d convolutions and bidirectional LSTMs with also an added Embedding layer of 50 dimensions. The model had a test accuracy of ≈ 82%. You can use the predictor.py file to test the code in your computer.
The bot has been made using Discord's discord.py api. Its super easy to learn and make bots. I was able to deploy my bot online thanks to Heroku's free dyno. The scripts I have used are:
- bot.py : This script runs the bot and is the main program for the bot
- requirements.txt: This script contains all the necessary libraries required to run the bot on Heroku.
- Procfile: This file tells Heroku server what file does it needs to run first.
Join this server to test the bot!