| # 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. |
