---
license: mit
library_name: span-marker
base_model: deepset/gelectra-large
tags:
- span-marker
- token-classification
- ner
- named-entity-recognition
pipeline_tag: token-classification
widget:
- text: "Jürgen Schmidhuber studierte ab 1983 Informatik und Mathematik an der TU München ."
  example_title: "Wikipedia"
datasets:
- gwlms/germeval2014
language:
- de
model-index:
  - name: SpanMarker with GELECTRA Large on GermEval 2014 NER Dataset by Stefan Schweter (@stefan-it)
    results:
      - task:
          type: token-classification
          name: Named Entity Recognition
        dataset:
          type: gwlms/germeval2014
          name: GermEval 2014
          split: test
          revision: f3647c56803ce67c08ee8d15f4611054c377b226
        metrics:
          - type: f1
            value: 0.8908
            name: F1
          - type: precision
            value: 0.8901
            name: Precision
          - type: recall
            value: 0.8916
            name: Recall
metrics:
  - f1
  - recall
  - precision
---

# SpanMarker for GermEval 2014 NER

This is a [SpanMarker](https://github.com/tomaarsen/SpanMarkerNER) model that
was fine-tuned on the [GermEval 2014 NER Dataset](https://sites.google.com/site/germeval2014ner/home).

The GermEval 2014 NER Shared Task builds on a new dataset with German Named Entity annotation with the following 
properties:  The data was sampled from German Wikipedia and News Corpora as a collection of citations. The dataset 
covers over 31,000 sentences corresponding to over 590,000 tokens. The NER annotation uses the NoSta-D guidelines, 
which extend the Tübingen Treebank guidelines, using four main NER categories with sub-structure, and annotating 
embeddings among NEs such as `[ORG FC Kickers [LOC Darmstadt]]`.

12 classes of Named Entites are annotated and must be recognized: four main classes `PER`son, `LOC`ation, `ORG`anisation,
and `OTH`er and their subclasses by introducing two fine-grained labels: `-deriv` marks derivations from NEs such as 
"englisch" (“English”), and `-part` marks compounds including a NE as a subsequence deutschlandweit (“Germany-wide”).

# Fine-Tuning

We use the same hyper-parameters as used in the
["German's Next Language Model"](https://aclanthology.org/2020.coling-main.598/) paper using the released
[GELECTRA Large](https://huggingface.co/deepset/gelectra-large) model as backbone.

Evaluation is performed with SpanMarkers internal evaluation code that uses `seqeval`. Additionally we use
the official GermEval 2014 Evaluation Script for double-checking the results. A backup of the `nereval.py` script
can be found [here](https://github.com/bplank/DaNplus/blob/master/scripts/nereval.perl).

We fine-tune 5 models and upload the model with best F1-Score on development set. Results on development set are
in brackets:

| Model                  | Run 1           | Run 2           | Run 3           | Run 4           | Run 5           | Avg.
| ---------------------- | --------------- | --------------- | --------------- | --------------- | --------------- | ---------------
| GELECTRA Large (5e-05) | (89.99) / 89.08 | (89.55) / 89.23 | (89.60) / 89.10 | (89.34) / 89.02 | (89.68) / 88.80 | (89.63) / 89.05

The best model achieves a final test score of 89.08%:

```bash
1. Strict, Combined Evaluation (official):
Accuracy:  99.26%;
Precision:  89.01%;
Recall:  89.16%;
FB1:  89.08
```

Scripts for [training](trainer.py) and [evaluation](evaluator.py) are also available.

# Usage

The fine-tuned model can be used like:

```python
from span_marker import SpanMarkerModel

# Download from the 🤗 Hub
model = SpanMarkerModel.from_pretrained("stefan-it/span-marker-gelectra-large-germeval14")

# Run inference
entities = model.predict("Jürgen Schmidhuber studierte ab 1983 Informatik und Mathematik an der TU München .")
```