README.md

---
license: cc0-1.0
task_categories:
  - text-classification
  - feature-extraction
language:
  - en
tags:
  - arxiv
  - crossref
  - citations
  - preprints
  - scholarly-communications
  - bibliometrics
  - doi
pretty_name: Crossref arXiv Citations
size_categories:
  - 100K<n<1M
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
---

# Crossref arXiv Citations

A dataset of arXiv preprints and their citations extracted from Crossref metadata, validated against DataCite records and DOI resolution.

## Dataset Description

This dataset maps arXiv works to the works in Crossref that cite them. Each record represents an arXiv preprint with all known citations from Crossref-registered works.

### Dataset Configurations

| Config | Description | arXiv Works | Citations |
|--------|-------------|-------------|-----------|
| `all` | All validated citations (default) | 923,302 | 5,121,642 |
| `asserted` | Citations where DOI was explicitly provided by publisher or matched by Crossref | 110,030 | 308,474 |
| `mined` | Citations where DOI was extracted from unstructured text | 891,035 | 4,813,168 |

### Loading the Dataset

```python
from datasets import load_dataset

# Load default (all citations)
dataset = load_dataset("cometadata/crossref-arxiv-citations")

# Load specific config
asserted = load_dataset("cometadata/crossref-arxiv-citations", "asserted")
mined = load_dataset("cometadata/crossref-arxiv-citations", "mined")
```

## Data Schema

Each record contains:

| Field | Type | Description |
|-------|------|-------------|
| `arxiv_doi` | string | The DOI for the arXiv work (format: `10.48550/arXiv.{id}`) |
| `arxiv_id` | string | The arXiv identifier (e.g., `1412.6980`) |
| `reference_count` | integer | Total number of reference instances |
| `citation_count` | integer | Number of unique citing works |
| `cited_by` | array | List of citing work objects |

Each citing work object in `cited_by` contains:

| Field | Type | Description |
|-------|------|-------------|
| `doi` | string | DOI of the citing work |
| `provenance` | string | How the citation was obtained (see below) |
| `matches` | array | List of reference matches from this citing work |

### Provenance Values

Each citation includes a `provenance` field indicating how the DOI was obtained:

- `publisher` - DOI was explicitly provided by the publisher in the reference metadata
- `crossref` - DOI was matched/validated by Crossref
- `mined` - DOI was extracted from unstructured text or other fields

The `asserted` config contains only `publisher` and `crossref` provenance citations (higher confidence). The `mined` config contains only `mined` provenance citations.

### Example Record

```json
{
  "arxiv_doi": "10.48550/arXiv.1412.6980",
  "arxiv_id": "1412.6980",
  "reference_count": 40040,
  "citation_count": 40040,
  "cited_by": [
    {
      "doi": "10.1016/j.jhydrol.2025.134709",
      "provenance": "mined",
      "matches": [
        {
          "provenance": "mined",
          "raw_match": "arXiv.1412.6980",
          "reference": {
            "key": "10.1016/j.jhydrol.2025.134709_b0175",
            "unstructured": "Kingma, D.P., & Ba, J. (2014), Adam: A Method for Stochastic Optimization..."
          }
        }
      ]
    }
  ]
}
```

## Extraction Process

### 1. Reference Extraction

arXiv references were extracted from the Crossref January 2025 Public Data File by scanning reference metadata for arXiv identifiers. The extraction detects multiple reference formats:

- arXiv IDs, including legacy formats: `arXiv:1412.6980`, `arXiv: 2206.15325`, `arXiv:cs.DM/9910013`
- arXiv DOIs: `10.48550/arXiv.1412.6980`
- arXiv URLs: `arxiv.org/abs/1412.6980`

Identifiers are normalized (lowercase, version suffixes removed) and deduplicated per citing work.

### 2. Provenance Tracking

For each extracted citation, the provenance is determined by checking:

1. If the reference has an explicit `DOI` field with `doi-asserted-by: publisher` -> `publisher`
2. If the reference has an explicit `DOI` field with `doi-asserted-by: crossref` -> `crossref`
3. Otherwise (DOI extracted from unstructured text) -> `mined`

### 3. Validation

The extracted arXiv DOIs were validated through a two-stage process:

1. DOIs were checked against the DataCite metadata for the arXiv record set (~2.8M records)
2. For DOIs not found in DataCite, HTTP HEAD requests to `doi.org` were used to verify resolution

Only arXiv works with valid, resolvable DOIs are included in this dataset.

## Data Sources

- January 2025 Crossref Public Data File
- January 2025 DataCite Public Data File
- Extraction tools: [cometadata/crossref-arxiv-citation-extraction](https://github.com/cometadata/crossref-arxiv-citation-extraction)

## Limitations

- Only includes citations registered in Crossref metadata
- Reference parsing depends on successful identifier extraction from unstructured references
- 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/). The underlying Crossref metadata is available under similar open terms.

## Citation

If you use this dataset, please cite:

```bibtex
@dataset{crossref_arxiv_citations_2025,
  title = {Crossref arXiv Citations},
  author = {Cometadata},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/cometadata/crossref-arxiv-citations}
}
```