masaki-sakata's picture
Update README.md
d0f5694 verified
metadata
dataset_info:
  features:
    - name: wiki_title
      dtype: string
    - name: qid
      dtype: string
    - name: description
      dtype: string
  splits:
    - name: en
      num_bytes: 1938197
      num_examples: 26205
  download_size: 1256160
  dataset_size: 1938197
configs:
  - config_name: default
    data_files:
      - split: en
        path: data/en-*
license: mit
language:
  - en
tags:
  - Wikidata
  - Wikipedia
  - Description
  - Entity
  - QID
  - Knowledge

Wikidata Descriptions Dataset

wikidata_descriptions pairs English Wikipedia article titles (wiki_title) and their Wikidata IDs (qid) with the English "description" available in Wikidata.
The corpus contains 26 205 entities. Wikidata descriptions are short, one-line summaries that concisely state what an entity is.
They can be used as lightweight contextual information in entity linking, search, question answering, knowledge-graph completion and many other NLP / IR tasks.


Dataset Structure

from datasets import load_dataset
ds = load_dataset("masaki-sakata/wikidata_descriptions", split="en")
print(ds)
# Dataset({
#   features: ['wiki_title', 'qid', 'description'],
#   num_rows: 26205
# })

Field description:

column type description
wiki_title str Title of the corresponding English Wikipedia article
qid str Wikidata identifier, e.g. Q7156
description str English one-line description provided by Wikidata (CC-0)

Example:

{
    "wiki_title": "Michael Jordan",
    "qid": "Q41421",
    "description": "American basketball player and businessman (born 1963)"
}

Quick Usage Example

from datasets import load_dataset
ds = load_dataset("masaki-sakata/wikidata_descriptions", split="en")
# Retrieve the first 3 entities
for record in ds.select(range(3)):
    print(record)

Source & Construction

  1. Seed list
    masaki-sakata/entity_popularity (English split) provides wiki_title and qid.
  2. Enrichment
    For every qid, we queried the Wikidata API (wbgetentities, language = en) and kept the English description if present.
  3. Filtering
    Rows with empty or missing descriptions were dropped, resulting in 26 205 items.
  4. License
    All descriptions originate from Wikidata and are released under CC-0.
    The dataset as a whole is distributed under the MIT License.