---
license: apache-2.0
tags:
  - relation-extraction
  - gliner
  - ner
  - information-extraction
language:
  - en
library_name: gliner2
pipeline_tag: text-classification
---

# GLiNER2 Relation Verifier

Semantic verifier for filtering false positive relations extracted by GLiNER2.

## Overview

GLiNER2's relation extraction can produce false positives when entities co-occur without expressing the claimed relation. This verifier uses a lightweight MLP trained on relation extraction datasets to filter these false positives.

**Performance:**
- Reduces false positive rate from ~50% to ~14% at threshold 0.55
- Precision: 75%, Recall: 44% at recommended threshold
- Trained on ZeroRel + Re-DocRED datasets

## Installation

```bash
pip install gliner2
```

## Usage

### Standalone

```python
from gliner2.verifiers import RelationVerifier

# Load verifier
verifier = RelationVerifier.from_pretrained("oneryalcin/gliner2-relation-verifier")

# Verify relations
text = "Steve Jobs founded Apple in 1976."
relations = {
    "founded": [
        {"head": {"text": "Steve Jobs", "start": 0, "end": 10},
         "tail": {"text": "Apple", "start": 19, "end": 24}}
    ]
}
verified = verifier.verify(text, relations)
print(verified)
# {'founded': [{'head': {...}, 'tail': {...}, 'verifier_score': 0.85}]}
```

### With GLiNER2

```python
from gliner2 import GLiNER2
from gliner2.verifiers import RelationVerifier

# Load models
model = GLiNER2.from_pretrained("fastino/gliner2-base-v1")
verifier = RelationVerifier.from_pretrained("oneryalcin/gliner2-relation-verifier")

# Extract and verify
text = "John read a book about Apple."
results = model.extract_relations(text, ["works_for"], return_dict=True)
verified = verifier.verify(text, results.get("relation_extraction", {}))
# Empty - correctly filtered the false positive!
```

## Threshold Selection

| Threshold | Precision | Recall | FPR | Use Case |
|-----------|-----------|--------|-----|----------|
| 0.50 | 70% | 58% | 25% | Balanced |
| **0.55** | 75% | 44% | 14% | **Recommended** |
| 0.60 | 80% | 36% | 9% | High precision |
| 0.75 | 87% | 21% | 3% | Very conservative |

```python
# Adjust threshold
verifier.set_threshold(0.60)  # More conservative
```

## Model Architecture

- **Encoder**: DeBERTa-v3-base (768-dim)
- **Features**: Head/tail/between-span embeddings + distance + order (3105-dim)
- **Classifier**: 3-layer MLP with LayerNorm and GELU

## Training

Trained on:
- ZeroRel: Zero-shot relation extraction dataset
- Re-DocRED: Revised DocRED with cleaned annotations

See [training details](https://github.com/oneryalcin/GLiNER2/issues/1).

## Citation

```bibtex
@misc{gliner2-verifier,
  author = {Oner Yalcin},
  title = {GLiNER2 Relation Verifier},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/oneryalcin/gliner2-relation-verifier}
}
```

## License

Apache 2.0