Plan.txt

At present Multilingual Hate Speech Detection using Transformers focuses on binary classification. We present NoahBERT (No Hate BERT) trained on MLMA multi-class/multi-domain English dataset and m-NoahBERT (multilingual NoahBERT) trained on translated MLMA using GoogleTranslate API.

Offensive vs Fearful vs Abusive vs Hateful

1. Preprocess the Dataset and get the required columns: decide what features and classes we want

2. Determining values of evaluation metrics (F1, P, R) on BERT-base (this means training te last layer of BERT to get the values)

3. Developing/designing NoahBERT (adding another FF layer to BERT and training it), and get the evaluation metrics 

4. Multilinguality: train and test on mBERT-base (Google), and develop/design m-NoahBERT

5. Test for Robustness (what happens when data/labels are noisy, adultering the Test Data but keeping the Train and Val data clean)