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:

1. MASRIHEADS Probs (13 dims): Emotion(8), Sentiment(3), Sarcasm(2).
2. BIHEADS Probs (13 dims): Emotion(8), Sentiment(3), Sarcasm(2).
3. 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.