clinia/CURE-v1

Dataset Card for CURE v1

Clinia’s CURE, Crosslingual Understanding and Retrieval Evaluation for health

Evaluate your retriever’s performance on query-passage pairs curated by medical professionals, across 10 disciplines and 3 cross-lingual settings.

Dataset Details

Uses

Direct Use

You can load the dataset with the following code:

language_setting = "en-en"
domain = "dermatology"
dermatology_queries = load_dataset(
            path="clinia/CURE-v1",
            data_files=f"{domain}/queries-{language_setting}.jsonl",
        )
corpus = load_dataset(
            path="clinia/CURE-v1",
            data_files=f"{domain}/corpus.jsonl",
        )
dermatology_qrels = load_dataset(
            path="clinia/CURE-v1",
            data_files=f"{domain}/qrels.jsonl",
        )

Dataset Structure

The dataset is organized into multiple folders, with each folder representing one of the 10 medical disciplines we currently support. Additionally, there is a folder named all, which holds data from all the disciplines combined.

Each folder contains:

• one queries-{lang}.jsonl for each language setting currently supported. Currently we support en-en,fr-en, and es-en.

  • Each row has:
    • _id : a unique identifier for the query
    • text : the query text
    • expertise : the complexity of the query; either layman or expert

• a corpus.jsonl file, with the following attributes for each row:

  • _id : a unique identifier for the passage
  • text : the passage text
  • title : the title of the article from which the passage was taken

• a qrels.jsonl file containing the relevance judgements. Note that these are shared across languages settings since they are cross-lingual datasets.

  • Each row has a corpus-id, a query-id and a score

Dataset Creation

Curation Rationale

• A lack of datasets for health information retrieval that capture the needs of medical professionals searching for clinical information
• A lack of cross-lingual datasets for information retrieval covering medical literature. A focus on evaluating the cross-lingual capacity is of importance as although much of the published literature is in English, non-native speakers may find it more efficient to seek out knowledge in their native language.
• A lack of datasets covering a broad array of medical disciplines and segmented as such
• A lack of a medically inclined dataset mixing layman vocabulary with expert terminology

Source Data

The corpus is a collection of open-access medical articles segmented into passages.

The queries come from our medical annotators, inspired by their daily information workflows.

Data Collection and Processing

Medical professionals were asked to annotate based on scientific literature they consulted in the course of their duties, express their information needs as queries and identify passages that met these needs. Both queries and passages were used to mine an index of over 7M passages to construct a retrieval corpus. Each query is translated to ensure that it exists in English, French and Spanish, and queries are assigned to one of the supported disciplines.

To ensure the source for retrieval corpus is of high quality, we employ multiple layers of filtering, deduplication, as well as popular heuristics from the literature inspired by Common Crawl.

We plan on keep updating it over time with more disciplines and languages as more anntators join the team and we gather feedback from the community.

Annotation process

The annotation workflow used Argilla. Annotators added queries and filled in the query-passage pairs with relevant metadata, such as the query level.

Dataset Card Authors

Olivier Rousseau

Nadia Sheikh

Daniel Buades Marcos

Dataset Card Contact

For more information or if you wish to contribute to CURE for its future releases, please contact us through email at [email protected]