Datasets:
Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,275 @@
|
|
1 |
-
---
|
2 |
-
license: mit
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: mit
|
3 |
+
task_categories:
|
4 |
+
- text-classification
|
5 |
+
language:
|
6 |
+
- id
|
7 |
+
tags:
|
8 |
+
- hate-speech-detection
|
9 |
+
- abusive-language
|
10 |
+
- text-classification
|
11 |
+
- indonesian
|
12 |
+
- social-media
|
13 |
+
- nlp
|
14 |
+
- content-moderation
|
15 |
+
- multi-label-classification
|
16 |
+
size_categories:
|
17 |
+
- 10K<n<100K
|
18 |
+
---
|
19 |
+
|
20 |
+
# Indonesian Hate Speech Detection Dataset
|
21 |
+
|
22 |
+
## Dataset Summary
|
23 |
+
|
24 |
+
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.
|
25 |
+
|
26 |
+
## Dataset Details
|
27 |
+
|
28 |
+
- **Total Samples**: 13,169 Indonesian tweets
|
29 |
+
- **Language**: Indonesian (Bahasa Indonesia)
|
30 |
+
- **Annotation Type**: Multi-label binary classification
|
31 |
+
- **Labels**: 12 different hate speech and abusive language categories
|
32 |
+
- **Format**: CSV file
|
33 |
+
- **Text Length**: 4-561 characters (average: 114 characters)
|
34 |
+
|
35 |
+
## Label Categories
|
36 |
+
|
37 |
+
### Primary Classifications
|
38 |
+
| Label | Description | Positive Cases | Percentage |
|
39 |
+
|-------|-------------|----------------|------------|
|
40 |
+
| `HS` | **Hate Speech** - General hate speech detection | 5,561 | 42.2% |
|
41 |
+
| `Abusive` | **Abusive Language** - Offensive or abusive content | 5,043 | 38.3% |
|
42 |
+
|
43 |
+
### Target-Based Classifications
|
44 |
+
| Label | Description | Positive Cases | Percentage |
|
45 |
+
|-------|-------------|----------------|------------|
|
46 |
+
| `HS_Individual` | Hate speech targeting specific individuals | 3,575 | 27.1% |
|
47 |
+
| `HS_Group` | Hate speech targeting groups/communities | 1,986 | 15.1% |
|
48 |
+
| `HS_Religion` | Religious hate speech | 793 | 6.0% |
|
49 |
+
| `HS_Race` | Racial/ethnic hate speech | 566 | 4.3% |
|
50 |
+
| `HS_Physical` | Physical appearance-based hate speech | 323 | 2.5% |
|
51 |
+
| `HS_Gender` | Gender-based hate speech | 306 | 2.3% |
|
52 |
+
| `HS_Other` | Other types of hate speech | 3,740 | 28.4% |
|
53 |
+
|
54 |
+
### Intensity Classifications
|
55 |
+
| Label | Description | Positive Cases | Percentage |
|
56 |
+
|-------|-------------|----------------|------------|
|
57 |
+
| `HS_Weak` | Weak/mild hate speech | 3,383 | 25.7% |
|
58 |
+
| `HS_Moderate` | Moderate hate speech | 1,705 | 12.9% |
|
59 |
+
| `HS_Strong` | Strong/severe hate speech | 473 | 3.6% |
|
60 |
+
|
61 |
+
## Key Statistics
|
62 |
+
|
63 |
+
**Text Characteristics:**
|
64 |
+
- **Average tweet length**: 114 characters
|
65 |
+
- **Shortest tweet**: 4 characters
|
66 |
+
- **Longest tweet**: 561 characters
|
67 |
+
- **Language**: Indonesian (Bahasa Indonesia)
|
68 |
+
|
69 |
+
**Label Distribution:**
|
70 |
+
- **Balanced primary labels**: ~42% hate speech, ~38% abusive
|
71 |
+
- **Imbalanced target categories**: Physical (2.5%) to Individual (27.1%)
|
72 |
+
- **Severity pyramid**: Weak (25.7%) > Moderate (12.9%) > Strong (3.6%)
|
73 |
+
|
74 |
+
## Use Cases
|
75 |
+
|
76 |
+
This dataset is ideal for:
|
77 |
+
|
78 |
+
- **Multi-label Text Classification**: Train models to detect multiple types of hate speech
|
79 |
+
- **Indonesian NLP**: Develop language-specific content moderation systems
|
80 |
+
- **Social Media Monitoring**: Build automated detection for Indonesian platforms
|
81 |
+
- **Severity Assessment**: Create models that classify hate speech intensity
|
82 |
+
- **Target Analysis**: Understand different targets of hate speech
|
83 |
+
- **Content Moderation**: Deploy real-time filtering systems
|
84 |
+
- **Research**: Study hate speech patterns in Indonesian social media
|
85 |
+
|
86 |
+
## Quick Start
|
87 |
+
|
88 |
+
```python
|
89 |
+
import pandas as pd
|
90 |
+
from sklearn.model_selection import train_test_split
|
91 |
+
from sklearn.feature_extraction.text import TfidfVectorizer
|
92 |
+
from sklearn.multioutput import MultiOutputClassifier
|
93 |
+
from sklearn.linear_model import LogisticRegression
|
94 |
+
from sklearn.metrics import classification_report
|
95 |
+
|
96 |
+
# Load dataset
|
97 |
+
df = pd.read_csv('data.csv')
|
98 |
+
|
99 |
+
# Prepare features and targets
|
100 |
+
X = df['Tweet']
|
101 |
+
y = df[['HS', 'Abusive', 'HS_Individual', 'HS_Group', 'HS_Religion',
|
102 |
+
'HS_Race', 'HS_Physical', 'HS_Gender', 'HS_Other']]
|
103 |
+
|
104 |
+
# Split data
|
105 |
+
X_train, X_test, y_train, y_test = train_test_split(
|
106 |
+
X, y, test_size=0.2, random_state=42
|
107 |
+
)
|
108 |
+
|
109 |
+
# Vectorize text
|
110 |
+
vectorizer = TfidfVectorizer(max_features=10000, ngram_range=(1, 2))
|
111 |
+
X_train_vec = vectorizer.fit_transform(X_train)
|
112 |
+
X_test_vec = vectorizer.transform(X_test)
|
113 |
+
|
114 |
+
# Train multi-label classifier
|
115 |
+
classifier = MultiOutputClassifier(LogisticRegression(random_state=42))
|
116 |
+
classifier.fit(X_train_vec, y_train)
|
117 |
+
|
118 |
+
# Evaluate
|
119 |
+
y_pred = classifier.predict(X_test_vec)
|
120 |
+
print("Multi-label Classification Report:")
|
121 |
+
for i, label in enumerate(y.columns):
|
122 |
+
print(f"\n{label}:")
|
123 |
+
print(classification_report(y_test.iloc[:, i], y_pred[:, i]))
|
124 |
+
```
|
125 |
+
|
126 |
+
## Advanced Usage Examples
|
127 |
+
|
128 |
+
### Intensity-Based Classification
|
129 |
+
```python
|
130 |
+
# Focus on hate speech intensity levels
|
131 |
+
intensity_labels = ['HS_Weak', 'HS_Moderate', 'HS_Strong']
|
132 |
+
hate_speech_data = df[df['HS'] == 1] # Only hate speech samples
|
133 |
+
|
134 |
+
# Multi-class intensity classification
|
135 |
+
y_intensity = hate_speech_data[intensity_labels]
|
136 |
+
```
|
137 |
+
|
138 |
+
### Target-Specific Models
|
139 |
+
```python
|
140 |
+
# Build specialized models for different targets
|
141 |
+
target_labels = ['HS_Individual', 'HS_Group', 'HS_Religion', 'HS_Race',
|
142 |
+
'HS_Physical', 'HS_Gender', 'HS_Other']
|
143 |
+
|
144 |
+
# Train target-specific classifiers
|
145 |
+
for target in target_labels:
|
146 |
+
# Create binary classifier for each target type
|
147 |
+
pass
|
148 |
+
```
|
149 |
+
|
150 |
+
### Indonesian Text Preprocessing
|
151 |
+
```python
|
152 |
+
import re
|
153 |
+
|
154 |
+
def preprocess_indonesian_text(text):
|
155 |
+
# Convert to lowercase
|
156 |
+
text = text.lower()
|
157 |
+
|
158 |
+
# Remove URLs
|
159 |
+
text = re.sub(r'http\S+|www\S+|https\S+', '', text, flags=re.MULTILINE)
|
160 |
+
|
161 |
+
# Remove user mentions and RT
|
162 |
+
text = re.sub(r'@\w+|rt\s+', '', text)
|
163 |
+
|
164 |
+
# Remove extra whitespace
|
165 |
+
text = re.sub(r'\s+', ' ', text).strip()
|
166 |
+
|
167 |
+
return text
|
168 |
+
|
169 |
+
# Apply preprocessing
|
170 |
+
df['Tweet_processed'] = df['Tweet'].apply(preprocess_indonesian_text)
|
171 |
+
```
|
172 |
+
|
173 |
+
## Model Architecture Suggestions
|
174 |
+
|
175 |
+
### Traditional ML
|
176 |
+
- **TF-IDF + Logistic Regression**: Baseline multi-label classifier
|
177 |
+
- **TF-IDF + SVM**: Better performance on imbalanced classes
|
178 |
+
- **Ensemble Methods**: Random Forest or Gradient Boosting
|
179 |
+
|
180 |
+
### Deep Learning
|
181 |
+
- **BERT-based Models**: Use Indonesian BERT (IndoBERT) for better performance
|
182 |
+
- **Multilingual Models**: mBERT or XLM-R for cross-lingual transfer
|
183 |
+
- **Custom Architecture**: BiLSTM + Attention for sequence modeling
|
184 |
+
|
185 |
+
### Multi-task Learning
|
186 |
+
```python
|
187 |
+
# Hierarchical classification approach
|
188 |
+
# 1. First classify: Normal vs Abusive vs Hate Speech
|
189 |
+
# 2. If Hate Speech: Classify target and intensity
|
190 |
+
# 3. Multi-task loss combining all objectives
|
191 |
+
```
|
192 |
+
|
193 |
+
## Evaluation Metrics
|
194 |
+
|
195 |
+
Given the multi-label and imbalanced nature:
|
196 |
+
|
197 |
+
### Primary Metrics
|
198 |
+
- **F1-Score**: Macro and micro averages
|
199 |
+
- **AUC-ROC**: For each label separately
|
200 |
+
- **Hamming Loss**: Multi-label specific metric
|
201 |
+
- **Precision/Recall**: Per-label analysis
|
202 |
+
|
203 |
+
### Specialized Metrics
|
204 |
+
```python
|
205 |
+
from sklearn.metrics import multilabel_confusion_matrix, jaccard_score
|
206 |
+
|
207 |
+
# Multi-label specific metrics
|
208 |
+
jaccard = jaccard_score(y_true, y_pred, average='macro')
|
209 |
+
hamming = hamming_loss(y_true, y_pred)
|
210 |
+
```
|
211 |
+
|
212 |
+
## Data Quality & Considerations
|
213 |
+
|
214 |
+
### Strengths
|
215 |
+
- ✅ **Comprehensive Labeling**: Multiple dimensions of hate speech
|
216 |
+
- ✅ **Large Scale**: 13K+ samples for robust training
|
217 |
+
- ✅ **Real-world Data**: Actual Indonesian tweets
|
218 |
+
- ✅ **Intensity Levels**: Enables nuanced classification
|
219 |
+
- ✅ **Multiple Targets**: Covers various hate speech types
|
220 |
+
|
221 |
+
### Limitations
|
222 |
+
- ⚠️ **Class Imbalance**: Some categories <5% positive samples
|
223 |
+
- ⚠️ **Language Specific**: Limited to Indonesian context
|
224 |
+
- ⚠️ **Temporal Bias**: Tweet collection timeframe not specified
|
225 |
+
- ⚠️ **Cultural Context**: May not generalize across Indonesian regions
|
226 |
+
|
227 |
+
## Ethical Considerations
|
228 |
+
|
229 |
+
**Content Warning**: This dataset contains hate speech and abusive language examples.
|
230 |
+
|
231 |
+
### Responsible Use
|
232 |
+
- **Research Purpose**: Intended for academic and safety research
|
233 |
+
- **Content Moderation**: Building protective systems
|
234 |
+
- **Bias Awareness**: Monitor for demographic biases in predictions
|
235 |
+
- **Privacy**: Tweets should be handled according to platform policies
|
236 |
+
|
237 |
+
### Not Suitable For
|
238 |
+
- Training generative models that could amplify hate speech
|
239 |
+
- Creating offensive content detection without human oversight
|
240 |
+
- Commercial use without proper ethical review
|
241 |
+
|
242 |
+
## Related Work & Benchmarks
|
243 |
+
|
244 |
+
### Indonesian NLP Resources
|
245 |
+
- **IndoBERT**: Pre-trained Indonesian BERT model
|
246 |
+
- **Indonesian Sentiment**: Related sentiment analysis datasets
|
247 |
+
- **Multilingual Models**: Cross-lingual hate speech detection
|
248 |
+
|
249 |
+
### Benchmark Performance
|
250 |
+
Consider comparing against:
|
251 |
+
- Traditional ML baselines (TF-IDF + SVM)
|
252 |
+
- Pre-trained language models (mBERT, IndoBERT)
|
253 |
+
- Multi-task learning approaches
|
254 |
+
|
255 |
+
## Citation
|
256 |
+
|
257 |
+
```bibtex
|
258 |
+
@dataset{indonesian_hate_speech_2025,
|
259 |
+
title={Indonesian Hate Speech Detection Dataset},
|
260 |
+
year={2025},
|
261 |
+
publisher={Dataset From Kaggle},
|
262 |
+
url={https://huggingface.co/datasets/nahiar/indonesian-hate-speech},
|
263 |
+
note={Multi-label hate speech and abusive language detection for Indonesian social media}
|
264 |
+
}
|
265 |
+
```
|
266 |
+
|
267 |
+
## Acknowledgments
|
268 |
+
|
269 |
+
This dataset contributes to safer Indonesian social media environments and supports research in:
|
270 |
+
- Multilingual content moderation
|
271 |
+
- Southeast Asian NLP
|
272 |
+
- Cross-cultural hate speech patterns
|
273 |
+
- Social media safety systems
|
274 |
+
|
275 |
+
**Note**: Handle this sensitive content responsibly and in accordance with ethical AI principles.
|