natepill/toxic_text_classifier_te

Toxic Text Classification Report

Link to Report: https://docs.google.com/document/d/1fKOzErfzN-hjsj542BrOv3S-YGs-LaEKKYYcSyLIOTc/edit?usp=sharing

Link to Google Colab Notebook (view-only): https://colab.research.google.com/drive/1GukAgm5Hgtz-qiZhYtrJQRSOluJ2croE

About the Project

Since the objective is simply to identify the percentage of how toxic a piece of text is, the the only feature column that is needed is "toxic". The pieces of texts that are classified as True for any of the other forms of toxicity are also classified as True under Toxic. This means that if a piece of text is classified as "severe_toxic" or "threat", it is also classified as True for "Toxic".

Model Architectures w/ 10-Fold Cross Validated Results

Simple LSTM

Bi-LSTM with Attention Mechanism

Bi-LSTM with CNN

Next Steps

• 1st Place Kaggle Competition Solution Overview: https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge/discussion/52557
• Training on Upsampling
• Denser networks
• Grid Search Parameter tuning across all models