German NER example using Transformers.

NER classifier for 3 classes on a custom German dataset with ~150K samples. Using huggingface pre-trained transformers and simpletransformers library. Training and final results visualization with W&B.

Jupyter notebook files

File name	Description
Data Parsing.ipynb	Parsing, fixing and preparing the data to train models.
Training the model.ipynb	Training of different transormer architectures.
Summary of all models.ipynb	Shows the summary of all trained models (loss, f1 score, etc). Also runs and evaluates the best trained model.

How to run

• Change the path_working_dir variable at the beginning of each notebook to your own working directory.
• This repository contains only the best trained transformer model pytorch_model.bin in /models/deepset-gbert-base-epochs3due to Github storage constrains.

Weights & biases

Dashboard of all experiments

First overview report

Best model general information

Property	Description
model type	Transformer
model name used for finetuning	deepset/gbert-base
model directory	../deepset-gbert-base-epochs3
model info	https://huggingface.co/deepset/gbert-base
paper	https://arxiv.org/pdf/2010.10906.pdf
release date	Oct 2020

Best model training information

Parameter	Value
number of training epochs	3
training time	~40 min on google colab
learning_rate	1e-4
batch_size	8

Results of the best model:

Metric	Value
f1_score	0.8666327741060837
precision	0.8322213181448332
recall	0.9040127275941312
eval_loss	0.17615252890027014
training_loss	0.14476392756443882