---
dataset_info:
- config_name: docs
  features:
  - name: docid
    dtype: string
  - name: title
    dtype: string
  - name: subtitle
    dtype: string
  - name: abstract
    dtype: string
  splits:
  - name: fr.test
    num_bytes: 21532392
    num_examples: 16389
  - name: en.test
    num_bytes: 19436528
    num_examples: 16389
  download_size: 22803206
  dataset_size: 40968920
- config_name: qrels
  features:
  - name: qid
    dtype: int64
  - name: docid
    dtype: string
  - name: rel
    dtype: int64
  splits:
  - name: en_fr.test
    num_bytes: 10393138
    num_examples: 360596
  - name: en_en.test
    num_bytes: 10393138
    num_examples: 360596
  download_size: 5660924
  dataset_size: 20786276
- config_name: queries
  features:
  - name: qid
    dtype: int64
  - name: query
    dtype: string
  splits:
  - name: en.test
    num_bytes: 20056464
    num_examples: 357710
  download_size: 7709798
  dataset_size: 20056464
configs:
- config_name: docs
  data_files:
  - split: fr.test
    path: docs/fr.test-*
  - split: en.test
    path: docs/en.test-*
- config_name: qrels
  data_files:
  - split: en_fr.test
    path: qrels/en_fr.test-*
  - split: en_en.test
    path: qrels/en_en.test-*
- config_name: queries
  data_files:
  - split: en.test
    path: queries/en.test-*
license: cc-by-nc-4.0
multilinguality:
- multilingual
- translation
task_categories:
- text-retrieval
task_ids:
- document-retrieval
language:
- en
- fr
pretty_name: CLIRudit
source_datasets:
- original
tags:
- research papers
---

# Dataset Card for CLIRudit

<!-- Provide a quick summary of the dataset. -->

**CLIRudit** is a dataset for **academic Cross-lingual information retrieval** (CLIR), consisting of English queries and French documents, based on [**Érudit**](https://www.erudit.org/en/), a non-profit publishing platform based in Quebec, Canada.

The CLIRudit dataset follows a TREC-style structure with three main components:

* **Queries**: Generated from English keywords of research articles by creating all possible three-keyword combinations.
For example, an article with keywords {A, B, C, D} would generate four queries: "A, B, C", "A, B, D", "A, C, D", and "B, C, D".
* **Relevance judgments (qrels)**: A document is considered relevant to a query if its English keywords metadata contains all keywords present in the query, reflecting the assumption that authors want their articles to be discoverable through these keywords.
* **Document collection**: Each document consists of concatenated French title, subtitle, and abstract, which serves as the retrieval unit.

The dataset includes only Érudit **research articles** containing both abstracts and keywords in both French and English.
The translations between languages were provided by the original authors of the articles.

As an _empirical upper bound_, we also include the actual English translations of the French titles, subtitles, and abstracts as documents.
This represents the best possible performance that can be achieved by a retrieval method with perfect translation.


## Dataset Details

### Dataset Description

<!-- Provide a longer summary of what this dataset is. -->

- **Language(s) (NLP):** English, French
- **License:** CC BY-NC 4.0. The dataset should not be used for any commercial purpose.


### Dataset Sources

<!-- Provide the basic links for the dataset. -->

- **Repository:** TBA
- **Paper:** [CLIRudit: Cross-Lingual Information Retrieval of Scientific Documents](https://arxiv.org/abs/2504.16264)


## Uses

The dataset is meant to be used to evaluate cross-lingual IR models.


## Dataset Structure

<!-- This section provides a description of the dataset fields, and additional information about the dataset structure such as criteria used to create the splits, relationships between data points, etc. -->

### Data Instances

A typical instance of the `docs` subset looks like:

```
{
  'docid': '000200ar',
  'title': 'Le charme féminin chez les Peuls Djeneri du Mali',
  'subtitle': 'Un « objet » de la nature ou de la culture?',
  'abstract': 'La notion de charme soulève une confusion de sens, car elle est souvent utilisée dans un sens commun...',
}
```

A typical instance of the `queries` subset looks like:

```
{
  'qid': 0,
  'query': '"biblioclasts", books with holes, picture books'
}
```

A typical instance of the `qrels` subset looks like:

```
{
  'qid': 0,
  'docid': '1089655ar',
  'rel': 1
}
```




### Data Fields

- `qid`: query id
- `query`: query text
- `docid`: document id in Érudit
- `title`: article title, if any
- `subtitle`: article subtitle,  if any
- `abstract`: article abstract
- `rel`: relevance label (currently only positives are labelled with 1)

<!-- Note that the descriptions can be initialized with the **Show Markdown Data Fields** output of the [Datasets Tagging app](https://huggingface.co/spaces/huggingface/datasets-tagging), you will then only need to refine the generated descriptions. -->



### Data Splits

There is one `subset` per dataset component (documents, queries, qrels). 
`split` is used to represent language and train/test split. 

For qrels, we first indicate query lang and then doc lang in the `split` name; e.g., `en_fr.test` means test split judgments with English queries and French documents. 




## Citation

<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->

**BibTeX:**

```bibtex
@article{valentini2025clirudit,
  title={CLIRudit: Cross-Lingual Information Retrieval of Scientific Documents}, 
  author={Francisco Valentini and Diego Kozlowski and Vincent Larivière},
  year={2025},
  eprint={2504.16264},
  journal={arxiv:2504.16264},
  archivePrefix={arXiv},
  primaryClass={cs.IR},
  url={https://arxiv.org/abs/2504.16264}, 
}
```

**APA:**

Francisco Valentini, Diego Kozlowski, & Vincent Larivière (2025). CLIRudit: Cross-Lingual Information Retrieval of Scientific Documents. arxiv:2504.16264.