Multilingual named entity recognition for medieval charters. Datasets and models

Annotated dataset for training named entities recognition models for medieval charters in Latin, French and Spanish.

The original raw texts for all charters were collected from four charters collections

- HOME-ALCAR corpus : https://zenodo.org/record/5600884

- CBMA : http://www.cbma-project.eu

- Diplomata Belgica : https://www.diplomata-belgica.be

- CODEA corpus : https://corpuscodea.es/

We include (i) the annotated training datasets, (ii) the contextual and static embeddings trained on medieval multilingual texts and (iii) the named entity recognition models trained using two architectures: Bi-LSTM-CRF + stacked embeddings and fine-tuning on Bert-based models (mBert and RoBERTa)

Codes, datasets and notebooks used to train models can be consulted in our gitlab repository: https://gitlab.com/magistermilitum/ner_medieval_multilingual

Our best RoBERTa model is also available in the HuggingFace library: https://huggingface.co/magistermilitum/roberta-multilingual-medieval-ner