Datasets:

IVN-RIN
/

BioBERT_Italian

pretty_name: BioBERT_Italian
license: cc-by-sa-4.0
dataset_info:
  features:
    - name: text
      dtype: string
  splits:
    - name: train
      num_bytes: 27319024484
      num_examples: 17203146
  download_size: 14945984639
  dataset_size: 27319024484
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
task_categories:
  - text-generation
language:
  - it
tags:
  - medical
  - biology

From this repository you can download the BioBERT_Italian dataset.

BioBERT_Italian is the Italian translation of the original BioBERT dataset, composed by millions of abstracts of PubMed papers.

Due to the unavailability of an Italian equivalent for the millions of abstracts and full-text scientific papers used by English, BERT-based biomedical models, we leveraged machine translation to obtain an Italian biomedical corpus based on PubMed abstracts and train BioBIT.

BioBIT has been evaluated on 3 downstream tasks: NER (Named Entity Recognition), extractive QA (Question Answering), RE (Relation Extraction). Here are the results, summarized:

NER:
- BC2GM = 82.14%
- BC4CHEMD = 80.70%
- BC5CDR(CDR) = 82.15%
- BC5CDR(DNER) = 76.27%
- NCBI_DISEASE = 65.06%
- SPECIES-800 = 61.86%
QA:
- BioASQ 4b = 68.49%
- BioASQ 5b = 78.33%
- BioASQ 6b = 75.73%
RE:
- CHEMPROT = 38.16%
- BioRED = 67.15%

We also fine-tuned BioBIT on PsyNIT (Psychiatric Ner for ITalian), a native Italian NER (Named Entity Recognition) dataset, composed by Italian Research Hospital Centro San Giovanni Di Dio Fatebenefratelli.

It was created starting from 100 electronic medical reports, manually anonymized (removing personal patient data, physicians’ references, dates, and locations). The anonymized documents were annotated by a psychologist with 10 years of experience. The electronic medical reports contained various information about patients: demographic variables, medical history, results of tests and medical examinations, reports from medical exams, and more. Four sections of such documents were extracted:

Pharmacological history, usually a structured list of medications that the patient is taking and their dosages.
Remote pathologic history and active disease, usually a list of past and current relevant diseases.
Cognitive proximate pathological history, typically unstructured, includes medical examinations the patient has undergone. It also includes information about the patient’s personal life, such as marital status, daily habits, sleep disorders, and any relevant aspects of his/her behavior.
Psychological evaluation, typically unstructured, reports the result of (neuro)psychological examinations, together with comments from the attending physician.

The class of entities in PsyNIT are:

Diagnosis and comorbidities (779 examples, 13.23% of the dataset)
Cognitive symptoms (2386 examples, 40.52% of the dataset)
Neuropsychiatric symptoms (707 examples, 12.01% of the dataset)
Drug treatment (162 examples, 2.75% of the dataset)
Medical assessment (1854 examples, 31.49% of the dataset)

Correspondence to

Claudio Crema (ccrema@fatebenefratelli.eu), Tommaso Mario Buonocore (tommaso.buonocore@unipv.it)

Citation

@article{BUONOCORE2023104431,
title = {Localizing in-domain adaptation of transformer-based biomedical language models},
journal = {Journal of Biomedical Informatics},
volume = {144},
pages = {104431},
year = {2023},
issn = {1532-0464},
doi = {https://doi.org/10.1016/j.jbi.2023.104431},
url = {https://www.sciencedirect.com/science/article/pii/S1532046423001521},
author = {Tommaso Mario Buonocore and Claudio Crema and Alberto Redolfi and Riccardo Bellazzi and Enea Parimbelli},
keywords = {Natural language processing, Deep learning, Language model, Biomedical text mining, Transformer}
}

@article{CREMA2023104557,
title = {Advancing Italian biomedical information extraction with transformers-based models: Methodological insights and multicenter practical application},
journal = {Journal of Biomedical Informatics},
volume = {148},
pages = {104557},
year = {2023},
issn = {1532-0464},
doi = {https://doi.org/10.1016/j.jbi.2023.104557},
url = {https://www.sciencedirect.com/science/article/pii/S1532046423002782},
author = {Claudio Crema and Tommaso Mario Buonocore and Silvia Fostinelli and Enea Parimbelli and Federico Verde and Cira Fundarò and Marina Manera and Matteo Cotta Ramusino and Marco Capelli and Alfredo Costa and Giuliano Binetti and Riccardo Bellazzi and Alberto Redolfi},
keywords = {Natural language processing, Deep learning, Biomedical text mining, Language model, Transformer}
}