MASRI-RF-XGB: The Egyptian Arabic Ensemble Judge
MASRI-RF-XGB is the conductor of the Kalamna committee. It is an XGBoost meta-learner that takes the probability outputs from both MASRIHEADS (Transformer) and BIHEADS (RNN) to make a final, highly robust prediction.
Stacking Strategy
The Maestro doesn't just look at the predictions; it analyzes the consensus and conflict between models. It uses a 39-dimensional meta-feature vector per sample:
- MASRIHEADS Probs (13 dims): Emotion(8), Sentiment(3), Sarcasm(2).
- BIHEADS Probs (13 dims): Emotion(8), Sentiment(3), Sarcasm(2).
- Deltas (13 dims): The absolute difference $|Masri - Bi|$ for every class.
Components
- Primary Expert:
T0KII/MASRIHEADS - Secondary Expert:
T0KII/BIHEADS - Conductor: XGBoost Classifier
Final Performance (Validation Splits)
| Task | Macro-F1 | Data Split |
|---|---|---|
| Emotion | 0.9072 | emotone_ar val |
| Sarcasm | 0.7304 | ar_sarcasm committee val |
| Sentiment | 0.7692 | multi-source sentiment val |
Use Case
Use this model when you need the highest possible accuracy for Egyptian dialect analysis, particularly in cases of heavy slang or subtle sarcasm where single models may struggle.