boonyhow/plos-citation-contexts

PLOS citation contexts

Full text of PLOS research articles with every in-text citation resolved to its sentence and (where the reference list provides one) the cited DOI. Built from the PLOS JATS XML corpus (allofplos).

Usage

from datasets import load_dataset
ds  = load_dataset("boonyhow/plos-citation-contexts")                    # all journals
one = load_dataset("boonyhow/plos-citation-contexts", "plos_one")        # a single journal

Follow a citation to its reference entry (works whether or not the DOI is known):

paper = ds["train"][0]
refs = {r["ref_id"]: r for r in paper["references"]}
for sec in paper["sections"]:
    for c in sec["citations"]:
        ref = refs.get(c["ref_id"])
        print(c["sentence"], "->", c["cited_doi"] or (ref and ref["raw"]))

Configs

config	papers
all	367,213
plos_one	310,503
plos_neglected_tropical_diseases	11,287
plos_computational_biology	10,930
plos_pathogens	10,258
plos_genetics	10,186
plos_biology	5,446
plos_global_public_health	3,444
plos_medicine	3,372
plos_digital_health	728
plos_climate	329
plos_mental_health	260
plos_water	257
plos_sustainability_and_transformation	104
plos_complex_systems	65
plos_clinical_trials	44

Schema

One row per paper.

field	description
paper_id	DOI of the citing paper
journal	PLOS journal (same as the config name)
title, abstract	article title and main abstract (structured abstracts as Heading: text)
sections	list of sections in document order (subsections are separate entries)
references	the paper's full reference list, in order

Each section: label (introduction / methods / results / discussion / other; subsections inherit their parent's label), heading, text (paragraphs joined by blank lines; figure/table captions excluded), and citations.

Each citation: sentence (the sentence containing the in-text marker), ref_id (key into references), and cited_doi (the reference's DOI, or null). A sentence citing several works appears once per cited work; citation ranges such as [3]–[6] are expanded.

Each reference: ref_id, label (as printed, e.g. 12), doi, pmid, type (journal / book / other ...), authors (Surname Initials), year, title (article or chapter title), source (journal or book title), volume, pages (page range or e-locator), and raw (a one-line formatted citation).

Notes

• cited_doi / doi are null when the reference list doesn't give a DOI. This is common before ~2014, for books, and for web resources. The reference fields (title, first author, year, source) are enough to resolve most of these against Crossref, OpenAlex, or PubMed.
• DOIs are lowercased and stripped of https://doi.org/ prefixes.
• Only research-article items are included (no editorials, synopses, or corrections).
• There is a single train split. If you split for evaluation, split by paper (not by citation row) to avoid leakage.

Citation

If you use this dataset, please cite it:

@misc{plos_citation_contexts_2026,
  author       = {Low, Boon How and Yu, Yucan and Ng, Kin Meng},
  title        = {PLOS Citation Contexts: Sentence-Level Citation Contexts from PLOS Research Articles},
  year         = {2026},
  version      = {v2.1},
  publisher    = {Hugging Face},
  doi          = {10.57967/hf/11359},
  url          = {https://huggingface.co/datasets/boonyhow/plos-citation-contexts},
  note         = {Derived from the PLOS JATS XML corpus (allofplos), CC BY 4.0}
}

Plain text: Low, B. H., Yu, Y., & Ng, K. M. (2026). PLOS Citation Contexts (v2.1) [Data set]. Hugging Face. https://huggingface.co/datasets/boonyhow/plos-citation-contexts

The underlying articles are © their authors and published by PLOS under CC BY 4.0; please also cite the original articles you use.

License

Article text © the authors, published by PLOS under CC BY 4.0. Please cite the original articles.