AffilGood Contrastive Dataset
This dataset is designed for training entity linking models, specifically for organizations and their representations. It includes 52,900 samples of organization queries with various representations and negative examples, generated using the AffilGood framework. The dataset is formatted to support contrastive learning tasks and can be used to train models for identifying and linking organizations to their official representations.
Dataset Details
- Number of Samples: 52,900
- Source: Generated using the AffilGood framework, based on the Research Organization Registry (ROR) data and various entity representations.
- Data Format: JSONL (line-by-line JSON)
- Query Types:
- Canonical names
- Aliases
- Acronyms
- Labels
- Location-based variations (e.g., city, country)
- Parent organizations
Each entry in the dataset contains:
- query: A query with an organization name and optional variations (e.g., city, acronym).
- positive: The corresponding correct entity representation with full details (e.g., parent organization, location).
- hard_negatives: A list of semantically similar but incorrect entities to provide challenging examples for training.
- ror_id: The ROR identifier for the organization.
- lang: The language of the query (e.g., "en" for English).
- query_type: The type of query (e.g.,
wikidata_label_city_only).
Dataset Structure
Each sample in the dataset is structured as follows:
{
"query": "[MENTION] Department for Child Protection [CITY] Perth",
"positive": "[MENTION] Department for Child Protection and Family Support [PARENT] Government of Western Australia [CITY] Perth [COUNTRY] Australia",
"hard_negatives": [
"[MENTION] Princess Margaret Hospital for Children [ACRONYM] PMH [PARENT] Government of Western Australia Department of Health [CITY] Perth [COUNTRY] Australia"
],
"ror_id": "https://ror.org/00420et70",
"lang": "en",
"query_type": "wikidata_label_city_only"
}