Code for Kaggle competition https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge

This script achieves 0.057 on LB.

First, install required libraries:

pip install nltk keras tqdm scikit-learn

Download embeddings. I used fastText crawl-300d-2M.vec. It can be found here: https://github.com/facebookresearch/fastText/blob/master/docs/english-vectors.md

Download competition's data. The links are here: https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge/data

Don't forget to extract files from archives

Next, run

python fit_predict.py train.csv test.csv crawl-300d-2M.vec

You will need some time to train a model. It takes ~3-4 hours on GTX 1080 Ti. In the finish, there will be file toxic_results/submit which you will be able to submit on Kaggle.