Edit model card

CamemBERT pretrained on french trade directories from the XIXth century

This mdoel is part of the material of the paper

Abadie, N., Carlinet, E., Chazalon, J., Duménieu, B. (2022). A Benchmark of Named Entity Recognition Approaches in Historical Documents Application to 19𝑡ℎ Century French Directories. In: Uchida, S., Barney, E., Eglin, V. (eds) Document Analysis Systems. DAS 2022. Lecture Notes in Computer Science, vol 13237. Springer, Cham. https://doi.org/10.1007/978-3-031-06555-2_30

The source code to train this model is available on the GitHub repository of the paper as a Jupyter notebook in src/ner/10-camembert_pretraining.ipynb.

Model description

This model pre-train the model Jean-Baptiste/camembert-ner on a set of ~845k entries from Paris trade directories from the XIXth century extracted with OCR. Trade directory entries are short and strongly structured texts that giving the name, activity and location of a person or business, e.g:

Peynaud, R. de la Vieille Bouclerie, 18. Richard, Joullain et comp., (commission- —Phéâtre Français. naire, (entrepôt), au port de la Rapée-

Intended uses & limitations

This model is intended for reproducibility of the NER evaluation published in the DAS2022 paper. Several derived models trained for NER on trade directories are available on HuggingFace, each trained on a different dataset :

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 5e-05
  • train_batch_size: 8
  • eval_batch_size: 8
  • seed: 42
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • num_epochs: 3.0

Training results

Training Loss Epoch Step Validation Loss
1.9603 1.0 100346 1.8005
1.7032 2.0 200692 1.6460
1.5879 3.0 301038 1.5570

Framework versions

  • Transformers 4.16.0.dev0
  • Pytorch 1.10.1+cu102
  • Datasets 1.17.0
  • Tokenizers 0.10.3
Downloads last month
8
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.