vexccz/manglish-nlp-sentiment

Manglish NLP Sentiment Model (v3.2.0)

XLM-Roberta multi-task model fine-tuned on 14,384 labeled Manglish examples for sentiment, emotion, and intent classification.

Performance

Task	Accuracy	Labels
Sentiment	95.0%	positive, negative, neutral
Emotion	90.3%	happy, sad, angry, fear, surprise, disgust, love, neutral
Intent	97.5%	question, statement, request, complaint, greeting, opinion
Average	94.3%	-

Comparison with v3.1.0

Task	v3.1.0 (distilbert)	v3.2.0 (xlm-roberta)	Improvement
Sentiment	88.5%	95.0%	+6.5%
Emotion	83.6%	90.3%	+6.7%
Intent	94.5%	97.5%	+3.0%
Average	88.9%	94.3%	+5.4%

Architecture

• Base model: xlm-roberta-base
• Architecture: Multi-task with task-specific heads + uncertainty-weighted loss
• Training: Focal loss, cosine annealing, FP16, gradient accumulation (effective batch 32)
• Ensemble: Confidence-based fallback (< 60% uses rule-based)
• Model size: ~1.1GB

Usage

# Install
# pip install malaysian-manglish-nlp[transformers]

from malaysian_manglish_nlp.transformers.manglish_model import load_model, predict

model = load_model()
result = predict("weh best gila makanan dia")
# {
#   "sentiment": {"label": "positive", "confidence": 0.95},
#   "emotion": {"label": "happy", "confidence": 0.91},
#   "intent": {"label": "opinion", "confidence": 0.89}
# }

Training Details

• Dataset: 14,384 examples (7,884 original + 6,500 augmented)
• Epochs: 8 (best at epoch 8)
• Batch size: 4 x 8 gradient accumulation = 32 effective
• Learning rate: 2e-5 with cosine annealing
• Max sequence length: 96 tokens
• Label smoothing: 0.1
• Early stopping: patience 3
• Mixed precision: FP16

Dataset

vexccz/manglish-nlp-dataset

Links

• PyPI: pip install malaysian-manglish-nlp
• GitHub: https://github.com/ZafranYusof/malaysia-manglish-nlp
• Docs: https://manglish-nlp.readthedocs.io
• Demo: https://huggingface.co/spaces/vexccz/manglish-nlp-demo

License

MIT