metadata

license: cc-by-4.0
dataset_info:
  features:
    - name: original
      dtype: string
    - name: tokens
      sequence: string
    - name: labels
      sequence: string
    - name: qid
      sequence: string
    - name: language
      dtype: string
    - name: url
      dtype: string
  splits:
    - name: train
      num_bytes: 5669993
      num_examples: 6764
  download_size: 1906917
  dataset_size: 5669993
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
language:
  - nl
  - en
  - es
  - pt
  - el
  - fr
  - de
pretty_name: winnl
task_categories:
  - token-classification

WiNNL

WikiNews Named entity recognition and Linking (WiNNL) is a multilingual news NER & NEL benchmark based on Wikinews articles. The dataset was created by automatically scraping and tagging news articles, and manually corrected by native speakers to ensure accuracy.

You can find more information in the paper: https://aclanthology.org/2024.dlnld-1.3.pdf

The dataset includes the following NER classes in IOB format (labels):

PER (Person): person names
LOC (Location): geographical locations
ORG (Organisation): organisations
AMB (Ambiguous): entities that had an ambigous wikidata link in the article, and could be classified as multiple NER classes
DATE (Date): dates (e.g. "2022-01-01", "5th of January 2022")
DISEASE (Disease): diseases (e.g. "cancer", "COVID-19")
EVT (Event): events (e.g. "2024 US elections")
SPE (Sport Event): sports events (e.g. "World Cup", "Olympics")
OTH (Other): other entities that do not fit into any of the above categories

Please note that only the PER, ORG and LOC classes have been corrected manually.

The qid column contains the Wikidata entity identifiers for the entities in the dataset, also in IOB format.