# taMASRIBERT – Egyptian Arabic Multi-Task Model

This model jointly performs **sarcasm detection**, **sentiment analysis** (3 classes), and **emotion recognition** (8 classes) for Egyptian Arabic and code-switched text. It combines a BERT‑based transformer (`MASRIBERTv3`) with a BiLSTM+BiGRU sequence encoder fed with FastText embeddings.

## Model Architecture
- **BERT backbone**: `T0KII/MASRIBERTv3` (trained on MASRISET, 3.2M Egyptian Arabic texts)
- **Sequence encoder**: 2‑layer BiLSTM + 2‑layer BiGRU on FastText (300‑dim)
- **Deep fusion**: Concatenation of BERT `[CLS]` and RNN final states → task‑specific heads

## Usage
1. Clone this repository.
2. Install dependencies: `pip install -r requirements.txt`
3. Run the inference script: `python inference.py`