| # MASRIHEADS: Multi-Task Egyptian Arabic BERT |
|
|
| MASRIHEADS is a multi-task transformer-based model designed for the Egyptian Arabic dialect (Masri). It performs simultaneous classification for Sarcasm, Sentiment, and Emotion by leveraging a shared backbone and specialized task heads. |
|
|
| ## Model Details |
| - **Base Model:** `T0KII/MASRIBERTv2` |
| - **Architecture:** BERT-based Multi-Task Classifier |
| - **Max Sequence Length:** 128 |
| - **Language:** Egyptian Arabic (arz) |
|
|
| ## Task Heads & Labels |
| The model is equipped with three classification heads: |
| - **Sarcasm (2 classes):** `not sarcastic`, `sarcastic` |
| - **Sentiment (3 classes):** `negative`, `neutral`, `positive` |
| - **Emotion (8 classes):** `none`, `anger`, `joy`, `sadness`, `love`, `sympathy`, `surprise`, `fear` |
|
|
| ## Training Data |
| The model was trained on a consolidated dataset of Egyptian dialect text, including: |
| - `ar_sarcasm` |
| - `ArSarcasm_v2` |
| - `iSarcasmEval` |
| - `egyptian-arabic-fake-reviews` |
| - `egyptian_sentiment_analysis_dataset` |
| - `emotone_ar` |
|
|
| ## Performance (Test Set F1) |
| | Task | Macro-F1 | |
| |-----------|----------| |
| | Sarcasm | 0.7882 | |
| | Sentiment | 0.7671 | |
| | Emotion | 0.7649 | |
| | **Mean** | **0.7734**| |
