Toxic Comment Classification Flask App

This project was originally developed for the NLP Project 3 workshop@Vancouver School of AI - Github & Meetup page - to demonstrate how to deploy a toxic comment text classification model using Heroku and Flask. However, to keep things simple, we built this in a hackthon way and did not put much effort in enforcing good practices.

For fun and still for fun, I am going to take it a step further by adding test-driven development (TDD), DevOps (CI/CD), container (Docker), and deployment (I am sure this was in the original workshop, but I just mention it to make it sounds fancy :).

Project Overview

Build a classification model that can distinguish between toxic and non-toxic comments and use the model in a real-life application.

The meetups serve as guidance. The goal is for all attendees to build a good machine learning model that can be used in a real-life application. We encourage all attendees to apply creativity to this project. There are no limits.

Installation Requirements

All code is written in Python. Please use this guide to get Python and Jupyter Notebook up and running.

Project Setup

Training

For those who want to walk though the whole process from training to deployment, you need to download the data to train the model

To download the data, run:

python src/download.py

This will download the training data and pre-trained embedding file in :

• ./assets/data/train.csv
• ./assets/embedding/fasttext-crawl-300d-2m/crawl-300d-2M.vec

To train the model, run:

python src/train_classifier.py

This will train a pooled GRU with FastText embedding. The text preprocessor and the model will be seriallized and stored in:

• ./assets/model/preprocessor.pkl
• ./assets/model/model.h5 Note: This took ~1 hour to train on Intel Core i7-HQ CPU

Prediction

To get a feeling of doing predictions, run:

python src/predict.py

# output:
# Corgi is stupid          - Toxicity: [0.99293655]
# good boy                 - Toxicity: [0.02075008]
# School of AI is awesome  - Toxicity: [0.01223523]
# F**K                     - Toxicity: [0.90747666]

Deployment

TODO

Resources

The project uses data from Kaggle's Toxic Comment Classification Challenge. The data can be found here.

If you are struggling with implementing some of the concepts discussed at the meetup, check out the slides notebook as guidance. There are also many kernels specific to the toxic comment challenge that you can refer to get some inspiration or help.

Alternatively, ask for assistance on Slack. That's what this community is all about :)

Meetup Contributors

Akshi Chaudhary

Johannes Harmse

Xinbin Huang

Peter Lin

Johannes G