# Arabic End-of-Turn (EOU) Detection Model — MARBERT Fine-Tuned

This model fine-tunes **MARBERT** for detecting **end-of-turn (EOU)** boundaries in Arabic dialogue.  
It predicts whether a given user message represents a **continuation** or an **end of turn**.

- **Repository:** `nihad-ask/Arabert-EOU-detection-model`
- **Task:** Binary End-of-Utterance Classification
- **Language:** Arabic (MSA + saudi dilect)
- **Base Model:** `UBC-NLP/MARBERT`

---

## 🚦 Task Definition

This is a **binary classification** task:

| Label | Meaning |
|-------|----------|
| **0** | Speaker will continue (NOT end of turn) |
| **1** | End of turn (EOU detected) |

This helps conversational agents determine if the user has finished typing or is likely to continue.

---

## 📌 Use Cases

- Conversational AI / Chatbots  
- Dialogue Systems  
- Turn-taking prediction  
- Speech-to-text segmentation  
- Customer support automation  

---


## 📊 Evaluation

### **Balanced Validation Set**

**Accuracy:** `0.9098`

| Class | Precision | Recall | F1-score | Support |
|-------|-----------|--------|----------|---------|
| **0 – Continue** | 0.9058 | 0.9148 | 0.9103 | 1702 |
| **1 – End of Turn** | 0.9139 | 0.9048 | 0.9094 | 1702 |

**Overall:**

| Metric | Score |
|--------|--------|
| Accuracy | 0.9098 |
| Macro Avg F1 | 0.9098 |
| Weighted Avg F1 | 0.9098 |
| Total Samples | 3404 |



## 🧪 How to Use

### **Python (PyTorch)**

```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "nihad-ask/marbert-EOU-detection-model"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

text = "تمام و بعدين؟"

inputs = tokenizer(text, return_tensors="pt")
outputs = model(**inputs)
prediction = torch.argmax(outputs.logits, dim=1).item()

if prediction == 1:
    print("End of turn")
else:
    print("Speaker will continue")


@misc{marbert_eou_2025,
  author = {Nihad Askri},
  title = {MARBERT Arabic End-of-Utterance Detection},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/nihad-ask/marbert-arabic-EOU-detection-model}}
}