SateCincau/nmt-gaul-formal

NMT Bahasa Gaul E-Commerce → Bahasa Indonesia Formal

Model Neural Machine Translation berbasis Transformer untuk mengonversi teks bahasa gaul dari ulasan e-commerce ke bahasa Indonesia formal.

Informasi Model

• Arsitektur: T5-Small Indonesian (fine-tuned)
• Task: Text Style Transfer — Informal → Formal
• Domain: Ulasan produk e-commerce Indonesia
• Dataset: 3.000 pasang kalimat gaul ↔ formal
• Sumber Data: PRDECT-ID (Sutoyo et al., 2025) + anotasi manual

Hasil Evaluasi

Metrik	Skor
BLEU	45.23
ROUGE-1	0.7910
ROUGE-2	0.6414
ROUGE-L	0.7829
METEOR	0.6767

Cara Pakai

from transformers import T5Tokenizer, T5ForConditionalGeneration

tokenizer = T5Tokenizer.from_pretrained("SateCincau/nmt-gaul-formal")
model     = T5ForConditionalGeneration.from_pretrained("SateCincau/nmt-gaul-formal")

teks_gaul = "barangnya kece bgt, recommended pokoknya!"

inputs = tokenizer(
    "formalkan: " + teks_gaul,
    return_tensors="pt",
    max_length=128
)

output = model.generate(**inputs, max_length=128, num_beams=4)
print(tokenizer.decode(output[0], skip_special_tokens=True))
# Output: Barang yang diterima sangat bagus, sangat direkomendasikan.

Contoh

Input (Gaul)	Output (Formal)
barangnya kece bgt, recommended!	Barang yang diterima sangat bagus.
seller fast respon, thx udh sabar	Penjual merespons cepat, terima kasih.
worth it bgt harganya	Harga sangat sepadan.
ongkirnya mahal tp barang cepet nyampe	Biaya pengiriman mahal namun barang cepat tiba.

Referensi

• Vaswani et al. (2017). Attention Is All You Need. NeurIPS.
• Cahyawijaya et al. (2021). IndoNLG Benchmark. EMNLP.
• Sutoyo et al. (2025). PRDECT-ID Dataset.