README.md

---
license: mit
task_categories:
- text-classification
language:
- id
tags:
- hate-speech-detection
- abusive-language
- text-classification
- indonesian
- social-media
- nlp
- content-moderation
- multi-label-classification
size_categories:
- 10K<n<100K
---

# Indonesian Hate Speech Detection Dataset

## Dataset Summary

This dataset contains **13,169 Indonesian tweets** annotated for hate speech detection and abusive language classification. The dataset provides comprehensive multi-label annotations covering different types of hate speech, target categories, and intensity levels, making it valuable for building robust content moderation systems for Indonesian social media.

## Dataset Details

- **Total Samples**: 13,169 Indonesian tweets
- **Language**: Indonesian (Bahasa Indonesia)
- **Annotation Type**: Multi-label binary classification
- **Labels**: 12 different hate speech and abusive language categories
- **Format**: CSV file
- **Text Length**: 4-561 characters (average: 114 characters)

## Label Categories

### Primary Classifications
| Label | Description | Positive Cases | Percentage |
|-------|-------------|----------------|------------|
| `HS` | **Hate Speech** - General hate speech detection | 5,561 | 42.2% |
| `Abusive` | **Abusive Language** - Offensive or abusive content | 5,043 | 38.3% |

### Target-Based Classifications
| Label | Description | Positive Cases | Percentage |
|-------|-------------|----------------|------------|
| `HS_Individual` | Hate speech targeting specific individuals | 3,575 | 27.1% |
| `HS_Group` | Hate speech targeting groups/communities | 1,986 | 15.1% |
| `HS_Religion` | Religious hate speech | 793 | 6.0% |
| `HS_Race` | Racial/ethnic hate speech | 566 | 4.3% |
| `HS_Physical` | Physical appearance-based hate speech | 323 | 2.5% |
| `HS_Gender` | Gender-based hate speech | 306 | 2.3% |
| `HS_Other` | Other types of hate speech | 3,740 | 28.4% |

### Intensity Classifications
| Label | Description | Positive Cases | Percentage |
|-------|-------------|----------------|------------|
| `HS_Weak` | Weak/mild hate speech | 3,383 | 25.7% |
| `HS_Moderate` | Moderate hate speech | 1,705 | 12.9% |
| `HS_Strong` | Strong/severe hate speech | 473 | 3.6% |

## Key Statistics

**Text Characteristics:**
- **Average tweet length**: 114 characters
- **Shortest tweet**: 4 characters
- **Longest tweet**: 561 characters
- **Language**: Indonesian (Bahasa Indonesia)

**Label Distribution:**
- **Balanced primary labels**: ~42% hate speech, ~38% abusive
- **Imbalanced target categories**: Physical (2.5%) to Individual (27.1%)
- **Severity pyramid**: Weak (25.7%) > Moderate (12.9%) > Strong (3.6%)

## Use Cases

This dataset is ideal for:

- **Multi-label Text Classification**: Train models to detect multiple types of hate speech
- **Indonesian NLP**: Develop language-specific content moderation systems
- **Social Media Monitoring**: Build automated detection for Indonesian platforms
- **Severity Assessment**: Create models that classify hate speech intensity
- **Target Analysis**: Understand different targets of hate speech
- **Content Moderation**: Deploy real-time filtering systems
- **Research**: Study hate speech patterns in Indonesian social media

## Quick Start

```python
import pandas as pd
from sklearn.model_selection import train_test_split
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.multioutput import MultiOutputClassifier
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import classification_report

# Load dataset
df = pd.read_csv('data.csv')

# Prepare features and targets
X = df['Tweet']
y = df[['HS', 'Abusive', 'HS_Individual', 'HS_Group', 'HS_Religion', 
        'HS_Race', 'HS_Physical', 'HS_Gender', 'HS_Other']]

# Split data
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, random_state=42
)

# Vectorize text
vectorizer = TfidfVectorizer(max_features=10000, ngram_range=(1, 2))
X_train_vec = vectorizer.fit_transform(X_train)
X_test_vec = vectorizer.transform(X_test)

# Train multi-label classifier
classifier = MultiOutputClassifier(LogisticRegression(random_state=42))
classifier.fit(X_train_vec, y_train)

# Evaluate
y_pred = classifier.predict(X_test_vec)
print("Multi-label Classification Report:")
for i, label in enumerate(y.columns):
    print(f"\n{label}:")
    print(classification_report(y_test.iloc[:, i], y_pred[:, i]))
```

## Advanced Usage Examples

### Intensity-Based Classification
```python
# Focus on hate speech intensity levels
intensity_labels = ['HS_Weak', 'HS_Moderate', 'HS_Strong']
hate_speech_data = df[df['HS'] == 1]  # Only hate speech samples

# Multi-class intensity classification
y_intensity = hate_speech_data[intensity_labels]
```

### Target-Specific Models
```python
# Build specialized models for different targets
target_labels = ['HS_Individual', 'HS_Group', 'HS_Religion', 'HS_Race', 
                'HS_Physical', 'HS_Gender', 'HS_Other']

# Train target-specific classifiers
for target in target_labels:
    # Create binary classifier for each target type
    pass
```

### Indonesian Text Preprocessing
```python
import re

def preprocess_indonesian_text(text):
    # Convert to lowercase
    text = text.lower()
    
    # Remove URLs
    text = re.sub(r'http\S+|www\S+|https\S+', '', text, flags=re.MULTILINE)
    
    # Remove user mentions and RT
    text = re.sub(r'@\w+|rt\s+', '', text)
    
    # Remove extra whitespace
    text = re.sub(r'\s+', ' ', text).strip()
    
    return text

# Apply preprocessing
df['Tweet_processed'] = df['Tweet'].apply(preprocess_indonesian_text)
```

## Model Architecture Suggestions

### Traditional ML
- **TF-IDF + Logistic Regression**: Baseline multi-label classifier
- **TF-IDF + SVM**: Better performance on imbalanced classes
- **Ensemble Methods**: Random Forest or Gradient Boosting

### Deep Learning
- **BERT-based Models**: Use Indonesian BERT (IndoBERT) for better performance
- **Multilingual Models**: mBERT or XLM-R for cross-lingual transfer
- **Custom Architecture**: BiLSTM + Attention for sequence modeling

### Multi-task Learning
```python
# Hierarchical classification approach
# 1. First classify: Normal vs Abusive vs Hate Speech
# 2. If Hate Speech: Classify target and intensity
# 3. Multi-task loss combining all objectives
```

## Evaluation Metrics

Given the multi-label and imbalanced nature:

### Primary Metrics
- **F1-Score**: Macro and micro averages
- **AUC-ROC**: For each label separately
- **Hamming Loss**: Multi-label specific metric
- **Precision/Recall**: Per-label analysis

### Specialized Metrics
```python
from sklearn.metrics import multilabel_confusion_matrix, jaccard_score

# Multi-label specific metrics
jaccard = jaccard_score(y_true, y_pred, average='macro')
hamming = hamming_loss(y_true, y_pred)
```

## Data Quality & Considerations

### Strengths
- ✅ **Comprehensive Labeling**: Multiple dimensions of hate speech
- ✅ **Large Scale**: 13K+ samples for robust training
- ✅ **Real-world Data**: Actual Indonesian tweets
- ✅ **Intensity Levels**: Enables nuanced classification
- ✅ **Multiple Targets**: Covers various hate speech types

### Limitations
- ⚠️ **Class Imbalance**: Some categories <5% positive samples
- ⚠️ **Language Specific**: Limited to Indonesian context
- ⚠️ **Temporal Bias**: Tweet collection timeframe not specified
- ⚠️ **Cultural Context**: May not generalize across Indonesian regions

## Ethical Considerations

**Content Warning**: This dataset contains hate speech and abusive language examples.

### Responsible Use
- **Research Purpose**: Intended for academic and safety research
- **Content Moderation**: Building protective systems
- **Bias Awareness**: Monitor for demographic biases in predictions
- **Privacy**: Tweets should be handled according to platform policies

### Not Suitable For
- Training generative models that could amplify hate speech
- Creating offensive content detection without human oversight
- Commercial use without proper ethical review

## Related Work & Benchmarks

### Indonesian NLP Resources
- **IndoBERT**: Pre-trained Indonesian BERT model
- **Indonesian Sentiment**: Related sentiment analysis datasets
- **Multilingual Models**: Cross-lingual hate speech detection

### Benchmark Performance
Consider comparing against:
- Traditional ML baselines (TF-IDF + SVM)
- Pre-trained language models (mBERT, IndoBERT)
- Multi-task learning approaches

## Citation

```bibtex
@dataset{indonesian_hate_speech_2025,
  title={Indonesian Hate Speech Detection Dataset},
  year={2025},
  publisher={Dataset From Kaggle},
  url={https://huggingface.co/datasets/nahiar/indonesian-hate-speech},
  note={Multi-label hate speech and abusive language detection for Indonesian social media}
}
```

## Acknowledgments

This dataset contributes to safer Indonesian social media environments and supports research in:
- Multilingual content moderation
- Southeast Asian NLP
- Cross-cultural hate speech patterns
- Social media safety systems

**Note**: Handle this sensitive content responsibly and in accordance with ethical AI principles.