arabic-toxic

Arabic Multi-Label Toxic Content Classifier

A fast and lightweight Arabic NLP package for toxic content detection, content moderation, social media analysis, and Arabic text safety classification.

The package supports multi-label classification and can detect several categories of toxic and harmful content in Arabic text.

---

Installation

pip install arabic-toxic

---

Quick Start

Single Text Classification

from arabic_toxic import classify

result = classify("يلعن امك يا ابن الحمار")

print(result)

Output

{
    "text": "يلعن امك يا ابن الحمار",
    "labels": ["Cussing", "Hatred"],
    "top_label": "Cussing",
    "is_toxic": True,
    "scores": {
        "Cussing": 99.999,
        "Hatred": 94.404,
        "Sexual": 2.938,
        "Racial": 0.762,
        "Appearance": 0.033,
        "Violence": 0.027,
        "NOT": 0.001
    }
}

---

Toxic / Non-Toxic Detection

The classify() function returns an is_toxic field that directly tells whether the input text is toxic.

from arabic_toxic import classify

result = classify("أنت إنسان محترم")

print(result["is_toxic"])

Output

False

Full Example

from arabic_toxic import classify

print(classify("أنت إنسان محترم"))
print(classify("يلعن امك يا ابن الحمار"))

Output

{
    "text": "أنت إنسان محترم",
    "labels": ["NOT"],
    "top_label": "NOT",
    "is_toxic": False,
    "scores": {
        "Hatred": 23.802,
        "Appearance": 23.436,
        "Racial": 13.917,
        "NOT": 11.666,
        "Cussing": 10.767,
        "Sexual": 3.805,
        "Violence": 1.656
    }
}

{
    "text": "يلعن امك يا ابن الحمار",
    "labels": ["Cussing", "Hatred"],
    "top_label": "Cussing",
    "is_toxic": True,
    "scores": {
        "Cussing": 99.999,
        "Hatred": 94.404,
        "Sexual": 2.938,
        "Racial": 0.762,
        "Appearance": 0.033,
        "Violence": 0.027,
        "NOT": 0.001
    }
}

---

Classifying a List of Texts

The current version can classify a list of texts by applying classify() to each item.

from arabic_toxic import classify

texts = [
    "أنت إنسان محترم",
    "شكراً لك على المساعدة",
    "يلعن امك يا ابن الحمار",
    "الشكل والتفكير حمار",
    "ليبرالية شوية صعاليك"
]

results = [classify(text) for text in texts]

for result in results:
    print(result)

Output

{
    "text": "أنت إنسان محترم",
    "labels": ["NOT"],
    "top_label": "NOT",
    "is_toxic": False,
    "scores": {
        "Hatred": 23.802,
        "Appearance": 23.436,
        "Racial": 13.917,
        "NOT": 11.666,
        "Cussing": 10.767,
        "Sexual": 3.805,
        "Violence": 1.656
    }
}

{
    "text": "شكراً لك على المساعدة",
    "labels": ["NOT"],
    "top_label": "NOT",
    "is_toxic": False,
    "scores": {
        "NOT": 94.747,
        "Hatred": 4.771,
        "Violence": 3.802,
        "Sexual": 3.508,
        "Cussing": 2.306,
        "Appearance": 2.286,
        "Racial": 2.183
    }
}

{
    "text": "يلعن امك يا ابن الحمار",
    "labels": ["Cussing", "Hatred"],
    "top_label": "Cussing",
    "is_toxic": True,
    "scores": {
        "Cussing": 99.999,
        "Hatred": 94.404,
        "Sexual": 2.938,
        "Racial": 0.762,
        "Appearance": 0.033,
        "Violence": 0.027,
        "NOT": 0.001
    }
}

{
    "text": "الشكل والتفكير حمار",
    "labels": ["Hatred"],
    "top_label": "Hatred",
    "is_toxic": True,
    "scores": {
        "Hatred": 98.338,
        "Racial": 12.536,
        "Appearance": 12.058,
        "NOT": 8.266,
        "Cussing": 2.405,
        "Sexual": 1.201,
        "Violence": 0.909
    }
}

{
    "text": "ليبرالية شوية صعاليك",
    "labels": ["Hatred", "Racial"],
    "top_label": "Racial",
    "is_toxic": True,
    "scores": {
        "Racial": 70.922,
        "Hatred": 65.004,
        "Cussing": 22.948,
        "Appearance": 5.112,
        "NOT": 4.478,
        "Sexual": 2.064,
        "Violence": 1.858
    }
}

---

Probability Scores

from arabic_toxic import predict_proba

scores = predict_proba("يلعن امك يا ابن الحمار")

print(scores)

---

Supported Labels

Label	Description
Appearance	Appearance-based insults
Cussing	Profanity and offensive language
Hatred	Hate and hostility
NOT	Non-toxic content
Racial	Ethnic or racial attacks
Sexual	Sexual harassment or abuse
Violence	Violent threats or content

---

Model Architecture

The model uses:

• Arabic text normalization
• TF-IDF word n-grams
• TF-IDF character n-grams
• Multi-label classification
• One-vs-Rest strategy
• SGD Logistic Classifier

Designed for:

• Content Moderation
• Social Media Monitoring
• Arabic NLP Research
• Online Community Safety
• AI Safety Applications

---

Evaluation Results

Test-set performance:

Metric	Score
Macro F1	0.9109
Micro F1	0.9205
Subset Accuracy	0.8433
Hamming Loss	0.0307

---

Notes

This package is intended for Arabic content moderation, research, and production prototyping.

Performance may vary depending on dialect, domain, and writing style. Always validate the model on your own data before using it in production decision systems.

---

Author

Dr. Faisal Alshargi

AI Researcher • NLP Engineer • Data Scientist

GitHub: https://github.com/alshargi

PyPI: https://pypi.org/project/arabic-toxic/