Biomedical Named Entity Recognition via Dictionary-based Synonym Generalization

Introduction

BioNER-SynGen is a cutting-edge Synonym Generalization (SynGen) framework designed to enhance the task of biomedical named entity recognition. This work aims to solve the synonym generalization problem associated with dictionary-based methods, which often miss out on recognizing biomedical concept synonyms not listed in their respective dictionaries.

Through span-based predictions and the incorporation of two pivotal regularization terms, the SynGen framework sets a new standard in the field:

• Synonym Distance Regularizer: Ensures concepts closely related in meaning are also closely related in the model's internal representation.

• Noise Perturbation Regularizer: Adds a degree of noise resistance to the model, ensuring more robust synonym detection.

Install

git clone https://github.com/fuzihaofzh/BioNER-SynGen.git
cd BioNER-SynGen
pip install -r requirements.txt
./scripts/setup.sh

Run the Experiments

We have conducted our experiments on 6 datasets including NCBI, BC5CDR-D, BC5CDR-D, BC4CHEMD, Species-800, and LINNAEUS. To replicate our experiments, please run each corresponding script. For example, to run SynGen on the NCBI dataset:

./scripts/run_ncbi.sh

You should expect output similar to the following:

Evaluating ncbi__edge,pos=umlsdi,neg=sap,ppreg,pndreg=0.1,ptm=pubmedbert
{
  "P": 0.685,
  "R": 0.631,
  "F1": 0.657
}

Do note, the scores might differ slightly from the results in the paper due to:

1. The inherent randomness when training neural networks.
2. We run each experiment 10 times with different random seeds to mitigate this randomness and report the average score.

Use Your Own Dictionay

Kindly create a dictionary file that matches the format of "datasets/umls/umlsdi.txt". Once done, replace the original file with your newly-created dictionary and execute the "scripts/run_ncbi.sh" script.

Cite

If you find our research and code beneficial, please consider citing our work:

@inproceedings{fu2023biomedical,
  title={Biomedical Named Entity Recognition via Dictionary-based Synonym Generalization},
  author={Fu, Zihao and Su, Yixuan and Meng, Zaiqiao and Collier, Nigel},
  booktitle={Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year={2023}
}

License

This project is licensed under the MIT License - refer to the LICENSE file for details.

Support

For questions, issues, or any other form of communication, please reach out to the authors or open an issue on this repository.

Acknowledgments

We would like to express our gratitude to the creators and contributors of the following projects, which have played an instrumental role in our research:

• SapBERT from Cambridge LTL

• PURE from Princeton NLP