|
|
--- |
|
|
language: en |
|
|
tags: |
|
|
- text-classification |
|
|
- abusive-language |
|
|
- hate-speech |
|
|
- toxicity |
|
|
- cyberviolence |
|
|
- abusive-language-detection |
|
|
- BERT |
|
|
license: mit |
|
|
--- |
|
|
|
|
|
# AbuseBERT |
|
|
|
|
|
## Model Description |
|
|
|
|
|
**AbuseBERT** is a **BERT-based classification model** fine-tuned for **abusive language detection**, optimized for **cross-dataset generalization**. |
|
|
|
|
|
> Abusive language detection models often suffer from poor generalization due to **sampling and lexical biases** in individual datasets. Our approach addresses this by integrating **ten publicly available abusive language datasets**, harmonizing labels and preprocessing textual samples to create a **broader and more representative training distribution**. |
|
|
|
|
|
**Key Findings:** |
|
|
- Individual dataset models: average F1 = **0.60** |
|
|
- Integrated model: F1 = **0.84** |
|
|
- Dataset contribution to performance improvements correlates with **lexical diversity (0.71 correlation)** |
|
|
- Integration exposes models to diverse abuse patterns, enhancing **real-world generalization** |
|
|
|
|
|
--- |
|
|
|
|
|
## Conclusion / Takeaways |
|
|
|
|
|
- No single dataset captures the full spectrum of abusive language; each dataset reflects a **limited slice** of the problem space. |
|
|
- Systematically integrating ten heterogeneous datasets significantly improves classification performance on a **held-out benchmark**. |
|
|
- Lexically dissimilar datasets contribute more to **enhancing generalization**. |
|
|
- The integrated model demonstrates superior **cross-dataset performance** compared to models trained on individual datasets. |
|
|
|
|
|
--- |
|
|
|
|
|
## Paper Reference |
|
|
|
|
|
Samaneh Hosseini Moghaddam, Kelly Lyons, Frank Rudzicz, Cheryl Regehr, Vivek Goel, Kaitlyn Regehr, |
|
|
“**Enhancing machine learning in abusive language detection with dataset aggregation**,” in *Proc. 35th IEEE Int. Conf. Collaborative Advances in Software Computing (CASC)*, 2025. |
|
|
|
|
|
--- |
|
|
|
|
|
## Intended Use |
|
|
|
|
|
**Recommended:** |
|
|
- Detecting abusive language in text from social media or online platforms |
|
|
- Research on bias mitigation and cross-dataset generalization |
|
|
- Supporting safe and inclusive online environments |
|
|
|
|
|
**Not Recommended:** |
|
|
- Fully automated moderation without human oversight |
|
|
- High-stakes legal or policy decisions |
|
|
|
|
|
--- |
|
|
|
|
|
## Usage Example |
|
|
|
|
|
```python |
|
|
from transformers import AutoTokenizer, AutoModelForSequenceClassification |
|
|
import torch |
|
|
|
|
|
# Load model and tokenizer |
|
|
tokenizer = AutoTokenizer.from_pretrained("Samanehmoghaddam/AbuseBERT") |
|
|
model = AutoModelForSequenceClassification.from_pretrained("Samanehmoghaddam/AbuseBERT") |
|
|
|
|
|
# Sample input |
|
|
text = "Your example text here." |
|
|
inputs = tokenizer(text, return_tensors="pt") |
|
|
outputs = model(**inputs) |
|
|
|
|
|
# Predicted label |
|
|
predicted_label = torch.argmax(outputs.logits, dim=1).item() |
|
|
print(f"Predicted label: {predicted_label}") |
|
|
|
