README.md

---
license: cc0-1.0
task_categories:
  - feature-extraction
language:
  - en
tags:
  - datacite
  - crossref
  - citations
  - scholarly-communications
  - bibliometrics
  - doi
  - landing-pages
pretty_name: Crossref DataCite Citations
size_categories:
  - 1M<n<10M
configs:
  - config_name: all
    data_files:
      - split: train
        path: data/all/*.parquet
    default: true
  - config_name: asserted
    data_files:
      - split: train
        path: data/asserted/*.parquet
  - config_name: mined
    data_files:
      - split: train
        path: data/mined/*.parquet
  - config_name: ambiguous_urls
    data_files:
      - split: train
        path: data/ambiguous_urls/*.parquet
  - config_name: url_evidence
    data_files:
      - split: train
        path: data/url_evidence/*.parquet
---

# Crossref DataCite Citations

A dataset of DataCite-registered works and the Crossref-registered works that
cite them, extracted from Crossref reference metadata and confirmed against
the DataCite monthly data file.

## Dataset Description

Each record in the citation configurations is one DataCite DOI together with
every confirmed citing work found in Crossref reference metadata. A reference
is confirmed when it carries a DOI registered in DataCite or an arXiv
identifier whose arXiv DOI is registered in DataCite. Registered landing-page
URLs are evaluated against release metadata before an exact decision joins citations.

Built from the Crossref Metadata Plus monthly snapshot `2026-07` and the
DataCite monthly data file `2026-07` with
[crossref-citation-extraction](https://github.com/cometadata/crossref-arxiv-citation-extraction)
version 2.2.0.

### Provenance and Ambiguity Policy

- `publisher` and `crossref` matches come from references whose structured
  `DOI` field holds the cited DOI with a `doi-asserted-by` value of
  `publisher` or `crossref`. Every other match is `mined`: a DOI, arXiv
  identifier, or registered landing-page URL found in the reference's `DOI`
  (without that assertion), `URL`, `article-title`, `journal-title`, or
  `unstructured` field.
- A landing-page URL registered by more than one DataCite record never
  produces a citation. Such references are published separately in the
  `ambiguous_urls` configuration with every candidate DOI and are excluded
  from every citation count.
- Counts are recomputed per configuration: a citing work appears in a
  configuration only with its matches of that configuration's provenances.

### Dataset Configurations

| Config | Description | Cited Works | Citing Works | Reference Matches |
|--------|-------------|-------------|--------------|-------------------|
| `all` | All confirmed citations (default) | 1,705,826 | 7,339,071 | 7,385,186 |
| `asserted` | Citations whose DOI was asserted by the publisher or matched by Crossref | 582,302 | 1,269,805 | 1,277,754 |
| `mined` | Citations mined from reference metadata without a publisher or Crossref DOI assertion | 1,314,323 | 6,070,488 | 6,107,432 |

The `ambiguous_urls` configuration holds 2,631 unresolved
URL references. It is evidence for downstream resolution, not citations, and
it is excluded from the counts above. The `url_evidence` configuration records
all URL decisions. Status totals are `exact` 38,508,
`related` 396, `uncertain` 34,163,
and `rejected` 282.

### Build Identity

| Field | Value |
|-------|-------|
| Crossref snapshot | `2026-07` |
| DataCite release | `2026-07` |
| Run identity fingerprint | `62d4c5b15deb3053eb082f5a2809e042fae118bef2b1d326947e6a29b6e8e8b7` |
| Extractor version | `2.2.0` |
| Artifact manifest | `metadata/pipeline-manifest.json` |

The manifest lists every published file with its size and SHA-256 digest; the
SHA-256 of the manifest bytes identifies this publication.

### Loading the Dataset

```python
from datasets import load_dataset

dataset = load_dataset("cometadata/crossref-datacite-citations")
asserted = load_dataset("cometadata/crossref-datacite-citations", "asserted")
mined = load_dataset("cometadata/crossref-datacite-citations", "mined")
ambiguous = load_dataset("cometadata/crossref-datacite-citations", "ambiguous_urls")
url_evidence = load_dataset("cometadata/crossref-datacite-citations", "url_evidence")
```

## Data Schema

### Citation Configurations (`all`, `asserted`, `mined`)

Each record contains:

| Field | Type | Description |
|-------|------|-------------|
| `doi` | string | Canonical lowercase DataCite DOI of the cited work |
| `citation_count` | integer | Number of distinct citing works in this configuration |
| `cited_by` | array | List of citing work objects |
| `reference_count` | integer | Number of confirmed reference tuples |

Each citing work object in `cited_by` contains:

| Field | Type | Description |
|-------|------|-------------|
| `doi` | string | DOI of the citing work |
| `provenance` | string | Strongest provenance among this work's matches (`publisher` > `crossref` > `mined`) |
| `matches` | array | Reference matches from this citing work, in reference order |

Each match contains:

| Field | Type | Description |
|-------|------|-------------|
| `provenance` | string | How this match was obtained (see below) |
| `match_type` | string | What was matched: `doi`, `arxiv`, `url` |
| `raw_match` | string | The exact reference text that matched (a substring, or a whole structured value) |
| `reference` | struct | The Crossref reference metadata (string-valued fields; absent keys are null) |

### `ambiguous_urls`

| Field | Type | Description |
|-------|------|-------------|
| `citing_doi` | string | DOI of the citing work |
| `ref_index` | integer | Position of the reference in the citing work's reference list |
| `url` | string | Normalized landing-page URL that matched several DataCite records |
| `raw_match` | string | The exact substring that matched |
| `candidate_dois` | array | Every DataCite DOI registered with this landing-page URL |
| `reference` | struct | The Crossref reference metadata (string-valued fields; absent keys are null) |

### `url_evidence`

Every row records a unique URL candidate and its semantic decision from
selected DataCite metadata. A unique registered URL creates a candidate; it
does not confirm a citation. Only `exact` decisions enter citation
configurations.

Candidate fields:

| Field | Type | Description |
|-------|------|-------------|
| `citing_doi` | string | Canonical lowercase DOI of the Crossref work containing the reference |
| `ref_index` | integer | Position of the reference in that work's reference list |
| `target_doi` | string | Canonical lowercase DataCite DOI proposed by the unique URL lookup |
| `url` | string | Rust-normalized registered landing-page URL; Python does not normalize it again |
| `raw_match` | string | Exact reference substring that produced the URL candidate |
| `provenance` | string | Always `mined` for URL candidates |
| `match_type` | string | Always `url` for URL candidates |
| `reference` | struct | Crossref reference metadata, using the safe stringification rules |

Decision fields:

| Field | Type | Description |
|-------|------|-------------|
| `candidate key` | derived | `(citing_doi, ref_index, target_doi)` uniquely identifies a candidate; it is not an additional stored column |
| `status` | string | `exact`, `related`, `uncertain`, or `rejected` semantic result |
| `reason_codes` | array of string | Ordered reasons for the result |
| `rule_id` | string | Identifier of the rule that produced the result |
| `validator_version` | integer | Current URL validator version `1` |
| `metadata_sha256` | string | Digest of the canonical selected target record; absent metadata uses the canonical missing-record digest `canonical_digest({"metadata": "missing"})` = `4a81185bc14f9b8386a19946353ef04716569b0ca44001ec8da71b83eadc490a` |
| `policy_sha256` | string | Digest of the current canonical URL policy `5270040aa63c329d19c9fd61d30ed94eace3dec758b0b6f83f154be7fcde40c9` |

`exact` identifies the same object, version, and granularity as `target_doi`.
`related` identifies a different but related object, version, or granularity.
`uncertain` has insufficient semantic evidence. `rejected` conflicts with the
target or an exclusion rule. The manifest binds the four status totals to this
ledger separately from citation counts and `ambiguous_urls` totals.

### Provenance Values

- `publisher` - the reference's structured `DOI` field holds the DataCite DOI with `doi-asserted-by: publisher`
- `crossref` - the reference's structured `DOI` field holds the DataCite DOI with `doi-asserted-by: crossref`
- `mined` - any other match: a DOI, arXiv identifier, or registered landing-page URL found in the reference's `DOI`, `URL`, `article-title`, `journal-title`, or `unstructured` field (a structured `DOI` counts as mined unless `doi-asserted-by` is `publisher` or `crossref`)

The `asserted` config contains only `publisher` and `crossref` matches; the
`mined` config contains only `mined` matches.

## Extraction Process

References in the Crossref snapshot are scanned for DOIs, arXiv identifiers,
and landing-page URLs. Each candidate is normalized and looked up in an index
built from the DataCite monthly data file `2026-07`: DOIs must be
registered in DataCite, arXiv identifiers must map to a registered arXiv DOI,
and uniquely registered landing-page URLs create candidates. Candidate URLs
are semantically validated against selected release metadata, and only
`exact` decisions enter citation configurations. URLs registered by several
records are recorded as ambiguous evidence instead.

## Data Sources

- Crossref Metadata Plus monthly snapshot `2026-07`
- DataCite monthly data file `2026-07`
- Extraction tools: [cometadata/crossref-arxiv-citation-extraction](https://github.com/cometadata/crossref-arxiv-citation-extraction)

## Limitations

- Only citations registered in Crossref reference metadata are included.
- Only DOIs and landing-page URLs registered in DataCite are matched; works
  registered with other agencies are not represented.
- Landing-page matching uses the URLs registered in the DataCite release
  `2026-07`; earlier or historical URLs of a record are not
  matched.
- Registered landing pages on DOI resolver hosts (`doi.org`, `dx.doi.org`,
  `www.doi.org`), bare site roots, and hosts without a dot are ignored for
  URL matching, so records registered with such a URL are matched only by DOI.
- The dataset reflects one Crossref snapshot and one DataCite release; each
  publication replaces the previous contents rather than accumulating them.
- References whose URL is registered by several DataCite records appear only
  in `ambiguous_urls` and are never counted as citations.
- Reference parsing depends on successful identifier extraction from
  reference metadata; the `mined` subset may contain lower-quality matches
  from noisy reference text.

## License

This dataset is released under
[CC0 1.0 Universal](https://creativecommons.org/publicdomain/zero/1.0/).

## Citation

```bibtex
@dataset{crossref_datacite_citations,
  title = {Crossref DataCite Citations},
  author = {Cometadata},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/cometadata/crossref-datacite-citations}
}
```