---
language: tr
license: mit
tags:
- ner
- token-classification
- turkish
- crf
- morphology
datasets:
- unimelb-nlp/wikiann
- turkish-nlp-suite/turkish-wikiNER
metrics:
- f1
---

# Extended Turkish NER (Hybrid CRF)

This is a high-performance **Named Entity Recognition (NER)** model for Turkish, using a hybrid approach of **Conditional Random Fields (CRF)**, deep morphological analysis (**Nuve/Zemberek**), and contextual embeddings (**BERTurk**).

## Features
- **6 Extended Categories:** `PER`, `LOC`, `ORG`, `COMPANY`, `GROUP`, `MOVIE`.
- **Hybrid Features:** Combines linguistic morphology with semantic BERT vectors.
- **Gazetteer Support:** Uses 160K+ entity entries for high precision.

## Performance
| Metric | Value |
| :--- | :--- |
| **Best F1-Score** | **%86.66** |
| Precision | %87.42 |
| Recall | %85.91 |

## Available Models (6 Variants)

| Model File | Description | F1 Score |
| :--- | :--- | :--- |
| `ner_crf_model.pkl` | **Best Hybrid (Nuve + BERT)** - Main SOTA model | **0.8666** |
| `final_proper_model.pkl` | Full features without embeddings | 0.8557 |
| `crf_gold_best.pkl` | Best Gold-only trained model | 0.8514 |
| `crf_gold_no_emb.pkl` | Gold without BERT embeddings | 0.8496 |
| `crf_gold_gaz_only.pkl` | Gazetteer-only features (baseline) | 0.8463 |
| `final_model.pkl` | Alternative final configuration | 0.8487 |

## Usage
The models are saved as `.pkl` files (sklearn-crfsuite).
Refer to the source code for feature extraction logic using Nuve and BERTurk.

### Example Inference
```python
import joblib
model = joblib.load("models/ner_crf_model.pkl")
# Use FeatureExtractor from src/features.py to prepare input
```

## Citation
Please cite this work if you use it in your research.
[Akademik Makale](https://github.com/WildGenie/nerextended/blob/master/docs/Akademik_Makale.md)