public-knowledge-project/ref-annotation-benchmark

RenoBench: A Citation Parsing Benchmark

RenoBench (Reference Annotation Benchmark) is a standardized evaluation benchmark for citation parsing—the task of annotating plain-text bibliographic references with structured components following the JATS (Journal Article Tag Suite) standard.

Dataset Description

RenoBench contains 10,000 plain-text citations paired with their corresponding JATS XML annotations. The dataset was assembled by extracting plain-text references from public domain PDFs and matching them to publisher-provided structured annotations.

Data Sources

Citations are sourced from four scholarly publishing platforms:

Source	Description	Percentage
SciELO	Scientific Electronic Library Online	47%
Redalyc	Red de Revistas Científicas de América Latina	24%
Open Research Europe	European Commission open access platform	14%
PKP	Public Knowledge Project OJS journals	14%

Dataset Composition

• 59% of citations include a persistent identifier (DOI)
• 14% of citing articles are preprints
• Languages: English (32%), Portuguese (30%), Spanish (23%), French (7%), German (3%), Italian (2%), Russian (2%), Chinese (1%)
• Publication types: Journal articles (53%), books (30%), webpages (8%), theses (5%), conference proceedings (4%)

Data Fields

Field	Type	Description
citing_article_doi	string	DOI of the article containing the citation (may be null)
plaintext	string	The plain-text citation as extracted from the PDF
xml	string	JATS XML annotation with structured bibliographic fields
source	string	Publishing platform source (scielo, redalyc, ore, pkp)

Example

Plain-text citation:

Stone NJ, Robinson JG, Lichtenstein AH, et al. 2013 ACC/AHA Guideline on the
Treatment of Blood Cholesterol to Reduce Atherosclerotic Cardiovascular Risk
in Adults. Circulation 2014;129(25 Suppl 2):S1-S45. doi:10.1161/01.cir.0000437738.63853.7a.

JATS XML annotation:

<mixed-citation publication-type="journal">
  <person-group person-group-type="author">
    <string-name><surname>Stone</surname> <given-names>NJ</given-names></string-name>,
    <string-name><surname>Robinson</surname> <given-names>JG</given-names></string-name>,
    <string-name><surname>Lichtenstein</surname> <given-names>AH</given-names></string-name>,
    <etal>et al</etal>
  </person-group>.
  <article-title>2013 ACC/AHA Guideline on the Treatment of Blood Cholesterol...</article-title>.
  <source>Circulation</source>
  <year>2014</year>;<volume>129</volume>(<issue>25</issue>):<fpage>S1</fpage>-<lpage>S45</lpage>.
  <pub-id pub-id-type="doi">10.1161/01.cir.0000437738.63853.7a</pub-id>.
</mixed-citation>

JATS XML Elements

The annotations use standard JATS reference elements:

Element	Description
<surname>	Author family name
<given-names>	Author given name(s) or initials
<article-title>	Title of the cited article
<source>	Journal name, book title, or publisher
<year>	Publication year
<volume>	Journal volume
<issue>	Journal issue
<fpage>, <lpage>	First and last page numbers
<pub-id pub-id-type="doi">	Digital Object Identifier

Data Collection

1. PDF Extraction: Article PDFs were converted to markdown using markitdown
2. Citation Extraction: Plain-text citations were extracted using Llama-3.1-8B-Instruct, with programmatic verification that extracted text appeared in the source document
3. Matching: Plain-text citations were matched to JATS XML annotations using normalized edit distance (threshold ≥ 0.75)
4. Filtering: Automated quality checks removed citations with structural errors, malformed fields, or annotation inconsistencies
5. Sampling: Balanced sampling across languages, publication types, and sources using learned sampling weights

Intended Use

RenoBench is designed for:

• Benchmarking citation parsing systems (GROBID, neural parsers, LLMs)
• Training sequence labeling or text-to-text models for citation parsing
• Evaluating multilingual and cross-domain generalization

Limitations

• Annotations reflect publisher practices, which may vary in completeness
• Some citation styles (legal, patents) are underrepresented
• Language distribution reflects source platform demographics

Citation

Coming soon!