geraldamasi/medmentions-ner

MedMentions BioNER (Custom Processed)

This dataset is a custom preprocessed version of the MedMentions dataset for biomedical Named Entity Recognition (NER) tasks.
It is compatible with Hugging Face Datasets and can be used directly for fine-tuning BERT-based models such as BERT or Bio_ClinicalBERT.

---

Dataset Summary

• Task: Named Entity Recognition (NER) in biomedical text
• Source: MedMentions
• Language: English
• Entity Types: 128 entity classes derived from UMLS semantic types
• Format: BIO tagging scheme

Each sentence is represented as a list of tokens and their corresponding BIO labels.

---

Example

Tokens:

'Event', 'Activity', 'Behavior', 'Social Behavior', 'Individual Behavior', 'Daily or Recreational Activity', 'Occupational Activity', 'Health Care Activity', 'Laboratory Procedure', 'Diagnostic Procedure

---

Dataset Structure

• Columns:

  • tokens: List of words in a sentence
  • ner_tags: List of BIO-formatted labels for each token

• Splits:

  • train: 80%
  • validation: 10%
  • test: 10%

Example Row:

{
  "tokens": ["DCTN4", "as", "a", "modifier", "of", "chronic", "Pseudomonas", "aeruginosa", "infection"],
  "ner_tags": ["B-Biologically Active Substance", "O", "O", "O", "O", "B-Disease or Syndrome", "B-Disease or Syndrome", "B-Disease or Syndrome", "B-Disease or Syndrome"]
}

Citation

@article{medmentions,
  title={MedMentions: A Large Biomedical Annotated Corpus with UMLS Concepts},
  author={Mohan, Sunil and Li, Donghui},
  journal={arXiv preprint arXiv:1902.09476},
  year={2019}
}