---
language:
- he
license: mit
tags:
- named-entity-recognition
- token-classification
- hebrew
- historical-manuscripts
- joint-learning
- multi-task-learning
- k-fold-validation
datasets:
- custom
metrics:
- f1
- accuracy
library_name: transformers
pipeline_tag: token-classification
---

# Hebrew Manuscript Joint NER Model v2

## Model Description

This model performs **joint entity recognition and role classification** for Hebrew historical manuscripts. It simultaneously:
1. **Named Entity Recognition (NER)**: Identifies person names in Hebrew text
2. **Role Classification**: Classifies each person as Author, Copyist, or Other

**Key Features:**
- Multi-task learning architecture
- Trained on Hebrew manuscript catalog data (MARC records)
- Uses distant supervision from structured metadata
- Optimized for historical Hebrew text
- Robust k-fold cross-validation


## K-Fold Cross-Validation Results

This model was trained using **5-fold cross-validation** for robust evaluation.

### Aggregate Performance

| Metric | Mean | Std Dev | Min | Max |
|--------|------|---------|-----|-----|
| **NER F1** | **0.9080** | ±0.0035 | 0.9011 | 0.9102 |
| **Classification Accuracy** | **1.0000** | ±0.0000 | - | - |

### Per-Fold Results

- **Fold 1**: NER F1 = 0.9011, Class Acc = 1.0000
- **Fold 2**: NER F1 = 0.9089, Class Acc = 1.0000
- **Fold 3**: NER F1 = 0.9096, Class Acc = 1.0000
- **Fold 4**: NER F1 = 0.9102, Class Acc = 1.0000
- **Fold 5**: NER F1 = 0.9100, Class Acc = 1.0000

**Best Model**: Fold 4 (NER F1: 0.9102)


## Model Architecture

- **Base Model**: [dicta-il/dictabert](https://huggingface.co/dicta-il/dictabert)
- **Architecture**: Joint multi-task learning
  - NER head: Token classification (B-PERSON, I-PERSON, O)
  - Role classification head: Sequence classification (AUTHOR, COPYIST, OTHER)
- **Training**: 5-fold cross-validation with early stopping
- **Regularization**: Dropout (0.3), Weight decay (0.01)

## Intended Use

### Primary Use Cases
- Extracting person names from Hebrew manuscript descriptions
- Identifying roles of people mentioned in manuscripts
- Building knowledge graphs of Hebrew manuscript creators
- Digital humanities research on Hebrew manuscripts

### Example Usage

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

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("alexgoldberg/hebrew-manuscript-joint-ner-v2")
model = AutoModel.from_pretrained("alexgoldberg/hebrew-manuscript-joint-ner-v2")

# Example text (Hebrew)
text = "הספר נכתב על ידי רבי משה בן מימון"
tokens = tokenizer(text, return_tensors="pt")

# Get predictions
with torch.no_grad():
    outputs = model(**tokens)
    # outputs contains both NER and role classification logits
```

## Training Data

- **Source**: Hebrew manuscript catalog records (MARC format)
- **Size**: ~10,000 samples
- **Annotation**: Distant supervision from structured metadata fields
- **Languages**: Hebrew (historical and modern)
- **Domain**: Manuscript descriptions, colophons, catalog records

## Training Procedure

### Hyperparameters
- **Epochs**: 10 (with early stopping, patience=3)
- **Batch Size**: 4
- **Learning Rate**: 2e-5
- **Optimizer**: AdamW
- **Dropout**: 0.3
- **Weight Decay**: 0.01
- **Lambda Weight**: 0.5 (for multi-task loss balancing)

### Data Split
- **K-Fold**: 5-fold stratified cross-validation
- **Stratification**: By number of persons per sample
- **Train/Val per fold**: 90/10 split

## Evaluation

### Metrics
- **NER**: Precision, Recall, F1 (seqeval)
- **Classification**: Accuracy
- **Combined**: Geometric mean of NER F1 and Classification Accuracy

### Validation Strategy
5-fold cross-validation ensures robust performance estimates and reduces overfitting to a single train/test split.

## Limitations

- Optimized for Hebrew manuscript descriptions (may not generalize to other Hebrew text types)
- Person names must follow historical Hebrew naming conventions
- Limited to three role categories (Author, Copyist, Other)
- Trained on catalog data (may not work well on manuscript images/OCR)

## Ethical Considerations

- Model trained on historical cultural heritage data
- Should be used to assist, not replace, expert manuscript catalogers
- Potential biases from historical naming conventions and catalog practices

## Citation

If you use this model, please cite:

```bibtex
@misc{hebrew-manuscript-joint-ner-v2,
  author = {Goldberg, Alexander},
  title = {Hebrew Manuscript Joint NER Model v2},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/alexgoldberg/hebrew-manuscript-joint-ner-v2}}
}
```

## Model Card Authors

Alexander Goldberg

## Model Card Contact

For questions or issues, please open an issue on the model repository.