Wailnsar4/hespress-arabic-hierarchical-classifier

Hespress Arabic Hierarchical Text Classifier

This model is a fine-tuned Arabic AraBERT model for hierarchical classification of Arabic news articles.

Model

The model is based on aubmindlab/bert-base-arabertv2.

The model was fine-tuned on Arabic articles collected from Hespress.

Dataset

After preprocessing and hierarchical classification:

• Total articles: 70,782
• Main categories: 16
• Hierarchical labels: 148
• Training samples: 56,625
• Validation samples: 7,078
• Test samples: 7,079

Hierarchical labels use the following format:

main_category__subcategory

Example:

اقتصاد__الاستثمار

Main Categories

• اقتصاد
• خارج الحدود
• جهات
• مجتمع
• سياسة
• رياضة
• حوادث
• فن وثقافة
• منوعات
• مغاربة العالم
• بيئة وعلوم
• 24 ساعة
• مصدر مطلع
• تمازيغت
• عالم التقنية
• سيارات

Training

• Base model: AraBERT v2
• Maximum sequence length: 512
• Number of labels: 148
• Learning rate: 2e-5
• Effective batch size: 20
• Epochs: 5
• Weight decay: 0.01
• FP16 mixed precision
• Best checkpoint: checkpoint-13000

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_id = "Wailnsar4/hespress-arabic-hierarchical-classifier"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

text = "شهدت أسعار النفط ارتفاعًا في الأسواق العالمية اليوم."

inputs = tokenizer(
    text,
    return_tensors='pt',
    truncation=True,
    max_length=512
)

outputs = model(**inputs)
prediction = outputs.logits.argmax(dim=-1).item()

print(model.config.id2label[prediction])

Intended Use

This model is intended for educational and research purposes, particularly for Arabic NLP and hierarchical text classification.

Limitations

The model was trained on articles collected from a specific Arabic news source. Its performance may vary when used with articles from other sources, writing styles, or domains.

The model should not be considered a general-purpose Arabic news classifier.

Project

This model was developed as part of an academic artificial intelligence project involving:

1. Web scraping
2. Data preprocessing
3. Hierarchical classification
4. Arabic NLP
5. AraBERT fine-tuning

Author

Wail Ali (Wailnsar4)