ClinText-SP Dataset Card

Dataset Description

ClinText-SP is the largest publicly available Spanish clinical corpus designed to support research in clinical natural language processing. It aggregates a rich collection of clinical texts from diverse open sources, including medical journals, annotated corpora from shared tasks, and supplementary sources like Wikipedia and medical textbooks.

The dataset contains:

• 35,996 samples with an average of ~700 tokens per sample
• Approximately 25.62M tokens in total

ClinText-SP offers a balanced mix of long, well-structured clinical case reports and shorter, schematic texts, making it ideal for a variety of clinical NLP tasks.

Data Sources

The corpus is built from three primary source types:

• Medical Journals: Clinical case reports from specialized Spanish-language journals.
• Annotated Corpora: Datasets from shared tasks.
• Other Sources: Additional clinical knowledge extracted from Wikipedia and select medical textbooks to complement the dataset.

Data Preprocessing

• Cleaning & Extraction: Texts were parsed and cleaned from PDFs, HTMLs, and other formats. Extraneous formatting, HTML artifacts, and non-essential metadata (e.g., author names) were removed.
• Customized Strategies: Specific regex-based heuristics and LLM-assisted methods (using Qwen2.5) were employed to accurately extract clinical case information.
• Deduplication & Language Filtering: Fuzzy deduplication (using MinHash) ensured unique entries, and non-Spanish texts were removed using Python Langdetect.

Intended Use

ClinText-SP is ideal for:

• Training and Benchmarking: Facilitating the development of Spanish clinical NLP models, including encoder-based models such as RigoBERTa Clinical.
• Domain-Adaptive Pretraining: Serving as a robust resource for adapting language models to the clinical domain.
• Research and Application: Advancing clinical language understanding and supporting applications in healthcare AI.

Limitations and Biases

• Biases: The dataset may reflect biases inherent to the selected sources and may not cover every clinical specialty.
• Coverage: While comprehensive, the dataset might not fully encapsulate the entirety of clinical nuances across all medical fields.
• Data Quality: Variations in data quality exist due to the diversity of sources and extraction methods.

For more detailed information, please check the original paper.

Citation

If you use ClinText-SP in your research, please cite the work as follows:

BibTeX:

@article{SUBIES2026114998,
title = {Advancing Spanish clinical language understanding through domain-adaptive pretraining and new open clinical resources},
journal = {Engineering Applications of Artificial Intelligence},
volume = {178},
pages = {114998},
year = {2026},
issn = {0952-1976},
doi = {https://doi.org/10.1016/j.engappai.2026.114998},
url = {https://www.sciencedirect.com/science/article/pii/S0952197626012819},
author = {Guillem García Subies and Álvaro Barbero Jiménez and Paloma Martínez Fernández},
}

APA:

Subies, G. G., Barbero Jiménez, Á., & Martínez Fernández, P. (2026). Advancing Spanish clinical language understanding through domain-adaptive pretraining and new open clinical resources. *Engineering Applications of Artificial Intelligence, 178*, 114998. [https://doi.org/10.1016/j.engappai.2026.114998](https://doi.org/10.1016/j.engappai.2026.114998)

Dataset Card Authors and Contact

Guillem García Subies: guillem.garcia@iic.uam.es, 100500844@alumnos.uc3m.es