Reference Graph (refcat)

In Summer 2021, the first version of a reference graph dataset, named "refcat", was released and integrated into the web interface. The dataset contains billions of references between papers in the fatcat catalog, as well as partial coverage of references from papers to books, to websites, and from Wikipedia articles to papers. This is a first step towards identifying links and references between scholarly works of all types preserved in

The refcat dataset can be downloaded in JSON lines format from the "Fatcat Database Snapshots and Bulk Metadata Exports" collection, and is released under a CC-0 license for broad reuse. Acknowledgement and attribution for both the aggregated dataset and the original metadata sources is strongly encouraged (see below for provenance notes).

References can be browsed on on an "outbound" ("References") and "inbound" ("Cited By") basis for individual release entities. There are also special pages for Wikipedia articles ("outbound", such as Internet) and Open Library books ("inbound", such as The Gift). JSON versions of these pages are available, but do not yet represent a stable API. The backend reference graph is available via the Elasticsearch API under the fatcat_ref index, but these schema and semantics of this index are also not yet stable.

How It Works

Raw reference data comes from multiple sources (see "provenance" below), but has the common structure of a "source" entity (which could be a paper, Wikipedia article, etc) and a list of raw references. There might be duplicate references for a single "source" work coming from different providers (eg, both Pubmed and Crossref reference lists). The goal is to match as many references as possible to the "target" work being referenced, creating a link from source to target. If a robust match is not found, the "unmatched" reference is retained and displayed in a human readable fashion if possible.

Depending on the source, raw references may be a simple "raw" string in an arbitrary citation style; may have been parsed or structured in fields like "title", "year", "volume", "issue"; might include a URL or identifier like an identifier; or may have already been matched to a specific target work by another party. It is also possible the reference is vague, malformed, mis-parsed, or not even a reference to a specific work (eg, "personal communication"). Based on the available structure, we might be able to do a simple identifier lookup, or may need to parse a string, or do "fuzzy" matching against various catalogs of known works. As a final step we take all original and potential matches, verify the matches, and attempt to de-duplicate references coming from different providers into a list of matched and unmatched references as output. The refcat corpus is the output of this process.

Two dominant modes of reference matching are employed: identifier based matching and fuzzy matching. Identifier based matching currently works with DOI, Arxiv ids, PMID and PMCID and ISBN. Fuzzy matching employs a scalable way to cluster documents (with pluggable clustering algorithms). For each cluster of match candidates we run a more extensive verification process, which yields a match confidence category, ranging from weak over strong to exact. Strong and exact matches are included in the graph.

All the code for this process is available open source:

  • refcat: batch processing and matching pipeline, in Python and Go
  • fuzzycat: Python verification code and "live" fuzzy matching

Metadata Provenance

The provenance for each reference in the index is tracked and exposed via the match_provenance field. A fatcat- prefix to the field means that the reference came through the refs metadata field stored in the fatcat catalog, but originally came from the indicated source. In the absence of fatcat-, the reference was found, updated, or extracted at indexing time and is not recorded in the release entity metadata.

Specific sources:

  • crossref (and fatcat-crossref): citations deposited by publishers as part of DOI registration. Crossref is the largest single source of citation metadata in refcat. These references may be linked to a specific DOI; contain structured metadata fields; or be in the form of a raw citation string. Sometimes they are "complete" for the given work, and sometimes they only include references which could be matched/linked to a target work with a DOI.
  • fatcat-datacite: same as crossref, but for the Datacite DOI registrar.
  • fatcat-pubmed: references, linked or not linked, from Pubmed/MEDLINE metadata
  • fatcat: references in fatcat where the original provenance can't be infered (but could be manually found by inspecting the release edit history)
  • grobid: references parsed out of full-text PDFs using GROBID
  • wikipedia: citations extracted from Wikipedia (see below for details)

Note that sources of reference metadata which have formal licensing restrictions, even CC-BY or ODC-BY licenses as used by several similar datasets, are not included in refcat.

Current Limitations and Known Issues

The initial Summer 2021 version of the index has a number of limitations. Feedback on features and coverage are welcome! We expect this dataset to be iterated over regularly as there are a few dimensions along which the dataset can be improved and extended.

The reference matching process is designed to eventually operate in both "batch" and "live" modes, but currently only "batch" output is in the index. This means that references from newly published papers are not added to the index in an ongoing fashion.

Fatcat "release" entities (eg, papers) are matched from a Spring 2021 snapshot. References to papers published after this time will not be linked.

Wikipedia citations come from the dataset Wikipedia Citations: A comprehensive dataset of citations with identifiers extracted from English Wikipedia, by Singh, West, and Colavizza. This is a one-time corpus based on a May 2020 snapshot of English Wikipedia only, and is missing many current references and citations. Additionally, only direct identifier lookups (eg, DOI matches) are used, not fuzzy metadata matching.

Open Library "target" matches are based on a snapshot of Open Library works, and are matched either ISBN (extracted from citation string) or fuzzy metadata matching.

Crossref references are extracted from a January 2021 snapshot of Crossref metadata, and do not include many updates to existing works.

Hundreds of millions of raw citation strings ("unstructured") have not been parsed into a structured for fuzzy matching. We plan to use GROBID to parse these citation strings, in addition to the current use of GROBID parsing for references from fulltext documents.

The current GROBID parsing used version v0.6.0. Newer versions of GROBID have improved citation parsing accuracy, and we intend to re-parse all PDFs over time. Additional manually-tagged training datasets could improve GROBID performance even further.

In a future update, we intend to add Wayback (web archive) capture status and access links for references to websites (distinct from references to online journal articles or books). For example, references to an online news article or blog post would indicate the closest (in time, to the "source" publication date) Wayback captures to that web page, if available.

References are only displayed on, not yet on

There is no current or planned mechanism for searching, sorting, or filtering article search results by (inbound) citation count. This would require resource-intensive transformations and continuous re-indexing of search indexes.

It is unclear how the batch-generated refcat dataset and API-editable release refs metadata will interact in the future. The original refs may eventually be dropped from the fatcat API, or at some point the refcat corpus may stabilize and be imported in to fatcat refs instead of being maintained as a separate dataset and index. It would be good to retain a mechanism for human corrections and overrides to the machine-generated reference graph.