hate-speech-severity-bert

Hate Speech Severity Predictor — BERT

Model Description

This is a fine-tuned BERT model (bert-base-uncased) for hate speech severity prediction, developed as part of an MSc research project at the University of Moratuwa, Sri Lanka.

The model predicts hate speech severity across three levels:

• Level 0 — Non-hate Speech
• Level 1 — Mild / Offensive
• Level 2 — Severe Hate Speech

It also produces a continuous severity score S in [0,1]: S = 0.0 x P(Level 0) + 0.5 x P(Level 1) + 1.0 x P(Level 2)

Model Details

• Developed by: J.A.U.S. Jayakody (239817M), University of Moratuwa
• Supervised by: Dr. Supunmali Ahangama
• Base model: bert-base-uncased
• Language: English
• License: MIT

Dataset

Fine-tuned on HateXplain (Mathew et al., 2021):

• 20,148 posts from Twitter and Gab
• Stratified 70-15-15 train-validation-test split

Training Details

• Epochs: 3 (best checkpoint: Epoch 2)
• Batch size: 16
• Learning rate: 2e-5
• Max sequence length: 128 tokens
• Class weighting: Balanced
• Hardware: Tesla T4 GPU

Evaluation Results

Severity Prediction Metrics:

• Spearman Correlation: 0.714
• Pearson Correlation: 0.720
• Mean Absolute Error: 0.212
• RMSE: 0.292

How to Use

from transformers import BertForSequenceClassification, BertTokenizer
import torch
import torch.nn.functional as F
import numpy as np

tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
model = BertForSequenceClassification.from_pretrained('UdaniSJ/hate-speech-severity-bert')
model.eval()

def predict_severity(text):
    inputs = tokenizer(text, return_tensors='pt',
                       truncation=True, max_length=128)
    with torch.no_grad():
        outputs = model(**inputs)
        probs = F.softmax(outputs.logits, dim=1).numpy()[0]
    score = 0.0*probs[0] + 0.5*probs[1] + 1.0*probs[2]
    level = int(np.argmax(probs))
    names = {0:'Non-hate', 1:'Mild', 2:'Severe'}
    return {'level': names[level], 'score': round(float(score),3)}

print(predict_severity("I love all people regardless of background"))

Live Demo

https://huggingface.co/spaces/UdaniSJ/hate-speech-severity-predictor

Limitations

• Trained on English social media content only
• May exhibit lexical over-reliance on identity terms
• Context-aware adjustment partially mitigates reclaimed language misclassification

References

• Mathew et al. (2021). HateXplain. AAAI 2021.
• Devlin et al. (2019). BERT. NAACL 2019.
• Ribeiro et al. (2016). LIME. KDD 2016.
• Lundberg and Lee (2017). SHAP. NeurIPS 2017.