bert-base-cased-NER-reranker

A BERT model trained on the synthetic literary NER context retrieval dataset Amalvy et. al, 2023.

To use this model, construct a text of the form NER-sentence [SEP] context-sentence. The model should predict the positive class if context-sentence is useful to predict NER-sentence, and the negative class otherwise.

Performance Metrics

The model obtains 98.34 F1 on the synthetic test set. See Amalvy et. al, 2023 for details about NER performance gains when using this retriever model to assit a NER model at inference.

How to Reproduce Training

See the training script here.

Citation

If you use this model in your research, please cite:

@InProceedings{Amalvy2023,
  title	       = {Learning to Rank Context for Named Entity Recognition Using a Synthetic Dataset},
  author       = {Amalvy, A. and Labatut, V. and Dufour, R.},
  booktitle    = {2023 Conference on Empirical Methods in Natural Language Processing},
  year	       = {2023},
  doi	       = {10.18653/v1/2023.emnlp-main.642},
  pages	       = {10372-10382},
}