Toxic Comments Multilabel Classfication

This repo contains code for Kaggle Toxic Comment Classification Challenge. Challenge aims to classify Wikipedia Comments based on toxic behaviour. We train models to classify Wikipedia comments in following 6 types of toxicity:

• toxic
• severe_toxic
• obscene
• threat
• insult
• identity_hate

Average ensembling of model with different hyper parameters and embeddings achieved 0.9860 ROC AUC score in competition private leaderboard.

Running Code

Setup

This repo uses keras for creating models. Install all dependency using requirements.txt.

pip install -r requirements.txt

Download stopwords and punkt tokenizer for nltk.

python -m nltk.downloader punkt stopwords

Data

Download train.tsv, test.tsv, sample_submission.csv from kaggle and save in data folder.

Download fasttext or glove(840B or twitter) embeddings and save in data folder.

Train Model

Edit configurations in train_model.py and train using following command:

python train_model.py

Test Model

Predict on custom queries

python query_model.py --model_iteration 100 --queries "I will kill you" "I hate you"