Span Marker Gelectra Large Germeval14

🚀 SpanMarker for GermEval 2014 NER

This is a SpanMarker model fine-tuned on the GermEval 2014 NER Dataset. It aims to solve named entity recognition tasks in German texts, providing accurate identification of named entities in the dataset.

✨ Features

• Fine-tuned on GermEval 2014: The model is specifically fine-tuned on the GermEval 2014 NER Dataset, which is based on German Wikipedia and News Corpora, covering over 31,000 sentences and 590,000 tokens.
• 12 Named Entity Classes: It can recognize 12 classes of named entities, including four main classes (PER, LOC, ORG, OTH) and their sub - classes with fine - grained labels.
• Accurate Evaluation: Evaluation is performed using both SpanMarker's internal evaluation code with seqeval and the official GermEval 2014 Evaluation Script.

📦 Installation

The installation steps are not provided in the original document, so this section is skipped.

💻 Usage Examples

Basic Usage

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 .")

📚 Documentation

Dataset Introduction

The GermEval 2014 NER Shared Task uses a new dataset with German Named Entity annotation. The data was sampled from German Wikipedia and News Corpora as a collection of citations. The NER annotation follows the NoSta - D guidelines, which extend the Tübingen Treebank guidelines. It uses four main NER categories with sub - structure and annotates embeddings among NEs, such as [ORG FC Kickers [LOC Darmstadt]].

Named Entity Classes

12 classes of Named Entities are annotated and recognized. There are four main classes: PERson, LOCation, ORGanisation, and OTHer. Sub - classes are introduced with two fine - grained labels: -deriv marks derivations from NEs (e.g., "englisch"), and -part marks compounds including a NE as a subsequence (e.g., deutschlandweit).

Fine - Tuning

We use the same hyper - parameters as in the "German's Next Language Model" paper, with the GELECTRA Large model as the backbone.

Evaluation is carried out using SpanMarker's internal evaluation code with seqeval and the official GermEval 2014 Evaluation Script. A backup of the nereval.py script can be found here.

We fine - tune 5 models and upload the model with the best F1 - Score on the development set. The results on the development set are as follows:

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

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

Scripts for training and evaluation are also available.

🔧 Technical Details

The model uses the SpanMarker architecture and fine - tunes on the GELECTRA Large model. The hyper - parameters are set according to the "German's Next Language Model" paper. The evaluation is based on seqeval and the official GermEval 2014 Evaluation Script.

📄 License

This project is licensed under the MIT license.

📊 Model Information