| # BIHEADS: Multi-Task BiLSTM+BiGRU for Egyptian Arabic |
|
|
| BIHEADS is a recurrent neural network ensemble designed to provide a complementary perspective to transformer-based models for Egyptian Arabic. It uses FastText embeddings and a parallel architecture to capture temporal nuances in dialectal speech. |
|
|
| ## Model Details |
| - **Architecture:** Parallel BiLSTM + BiGRU shared backbone |
| - **Embeddings:** FastText `facebook/fasttext-arz-vectors` (300-dim) |
| - **Hidden Size:** 256 per direction |
| - **Layers:** 2-layer LSTM, 2-layer GRU |
|
|
| ## Task Heads & Labels |
| - **Emotion (8 classes):** `none`, `anger`, `joy`, `sadness`, `love`, `sympathy`, `surprise`, `fear` |
| - **Sentiment (3 classes):** `negative`, `neutral`, `positive` |
| - **Sarcasm (2 classes):** `not sarcastic`, `sarcastic` |
|
|
| ## Training Data |
| The model utilizes the same Egyptian-filtered training splits as MASRIHEADS to ensure feature alignment for downstream ensemble learners. |
|
|
| ## Performance (Test Set F1) |
| | Task | Macro-F1 | |
| |-----------|----------| |
| | Sarcasm | 0.6293 | |
| | Sentiment | 0.6284 | |
| | Emotion | 0.5514 | |
| | **Mean** | **0.6030**| |
