rmtariq/malay_classification

🇲🇾 Malay Claim Classification Model

This is a fine-tuned BERT model built to classify claims in Malay (and English) into 21 categories.

📊 Categories

The model classifies claims into the following categories:

• Politik (Politics)
• Perpaduan (Unity)
• Keluarga (Family)
• Belia (Youth)
• Perumahan (Housing)
• Internet (Internet)
• Pengguna (Consumer)
• Makanan (Food)
• Pekerjaan (Employment)
• Pengangkutan (Transportation)
• Sukan (Sports)
• Ekonomi (Economy)
• Hiburan (Entertainment)
• Jenayah (Crime)
• Alam Sekitar (Environment)
• Teknologi (Technology)
• Pendidikan (Education)
• Agama (Religion)
• Sosial (Social)
• Kesihatan (Health)
• Halal (Halal)

🧠 Base Model

Fine-tuned from bert-base-multilingual-cased, which supports both Malay and English text.

🧪 Example Usage

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

# Load model and tokenizer
model_name = "rmtariq/malay_classification"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Function to classify a claim
def classify_claim(claim):
    # Prepare the input
    inputs = tokenizer(claim, return_tensors="pt", truncation=True, max_length=128)
    
    # Get the prediction
    with torch.no_grad():
        outputs = model(**inputs)
    
    # Get the predicted class
    logits = outputs.logits
    predicted_class_id = logits.argmax().item()
    
    # Get the confidence score
    probabilities = torch.nn.functional.softmax(logits, dim=1)[0]
    confidence = probabilities[predicted_class_id].item()
    
    # Map to category
    category = model.config.id2label[predicted_class_id]
    
    return category, confidence

# Example claims
examples = [
    "Projek mega kerajaan penuh dengan ketirisan.",
    "Harga barang keperluan naik setiap bulan.",
    "Program vaksinasi tidak mencakupi golongan luar bandar.",
    "Makanan di hotel lima bintang tidak jelas status halalnya."
]

# Classify each example
for claim in examples:
    category, confidence = classify_claim(claim)
    print(f"Claim: {claim}")
    print(f"Category: {category}")
    print(f"Confidence: {confidence:.4f}")
    print("-" * 50)

📚 Dataset

Fine-tuned on a custom dataset with 3,675 claims labeled by category, with an 80/20 train/test split.

🔍 Evaluation

The model achieves high accuracy on the test set, with most predictions having confidence scores above 0.95.

🎯 Specific Claim Patterns

The model includes special handling for specific claim patterns:

1. Police-related claims: Claims about the police chief, summons, or threats

   • Example: "Ketua Polis Negara (KPN) Tan Sri Razarudin Husain hantar e-mel berkaitan saman dan berbaur ugutan kepada orang awam"
   • Category: Jenayah (Crime)

2. Zakat-related claims: Claims about zakat fitrah, rice types, or payment validity

   • Example: "Zakat fitrah tidak sah jika dibayar tidak mengikut jenis beras yang dimakan"
   • Category: Agama (Religion)

3. Tax-related claims: Claims about government taxes, especially on palm oil

   • Example: "Kerajaan akan memperkenalkan cukai khas minyak sawit mentah"
   • Category: Ekonomi (Economy)

4. Consumer product claims: Claims about contact lenses or online sales

   • Example: "Kanta lekap tidak boleh dijual secara dalam talian"
   • Category: Pengguna (Consumer)

5. National security claims: Claims about ammunition, colonization, or enemies

   • Example: "Penemuan 50 tan kelongsong dan peluru petanda negara bakal dijajah musuh"
   • Category: Politik (Politics)

📋 License

MIT License