rayliuca/WikidataLabels

Wikidata Labels

Large parallel corpus for machine translation

• Entity label data extracted from Wikidata (2022-01-03), filtered for item entities only
• Only download the languages you need with datasets>=2.14.0
• Similar dataset: https://huggingface.co/datasets/wmt/wikititles (18 Wikipedia titles pairs instead of all Wikidata entities)

Dataset Details

Dataset Sources

• Wikidata JSON dump (wikidata-20220103-all.json.gz) https://www.wikidata.org/wiki/Wikidata:Database_download

Uses

You can generate parallel text examples from this dataset like below:

from datasets import load_dataset
import pandas as pd

def parallel_labels(lang_codes: list, how="inner", repo_id="rayliuca/wikidata_entity_label", merge_config={}, datasets_config={}) -> pd.DataFrame:
    out_df = None
    for lc in lang_codes:
        dataset = load_dataset(repo_id, lc, **datasets_config)
        dataset_df = dataset['label'].to_pandas().rename(columns={"label":lc}).drop(columns=['lastrevid'])
        if out_df is None:
            out_df = dataset_df
        else:
            out_df = out_df.merge(
                    dataset_df,
                    on='wikidata_id',
                    how=how,
                    **merge_config
                )
    return out_df

# Note: the "en" subset is >4GB
parallel_labels(['en', 'fr', 'ja', 'zh']).head()

Output

	wikidata_id	en	fr	ja	zh
0	Q109739412	SARS-CoV-2 Omicron variant	variant Omicron du SARS-CoV-2	SARSコロナウイルス2-オミクロン株	嚴重急性呼吸道症候群冠狀病毒2型Omicron變異株
1	Q108460606	Ulughbegsaurus	Ulughbegsaurus	ウルグベグサウルス	兀魯伯龍屬
2	Q108556886	AUKUS	AUKUS	AUKUS	AUKUS
3	Q106496152	Claude Joseph	Claude Joseph	クロード・ジョゼフ	克洛德·约瑟夫
4	Q105519361	The World's Finest Assassin Gets Reincarnated in a Different World as an Aristocrat	The World's Finest Assassin Gets Reincarnated in Another World as an Aristocrat	世界最高の暗殺者、異世界貴族に転生する	世界頂尖的暗殺者轉生為異世界貴族

Note: this example table above shows a quirk(?) of the Wiki data. The French Wikipedia page The World's Finest Assassin Gets Reincarnated in Another World as an Aristocrat uses English for its title. While this could be disadvantageous for direct translation training, it also provides insights into how native speakers might call this entity instead of the literal translation on the Wiki page as well

Dataset Structure

Each language has its own subset (aka config), which means you only have to download the languages you need with datasets>=2.14.0

Each subset has these fields:

• wikidata_id
• lastrevid
• label

Dataset Creation

Data Collection and Processing

• Filtered for item entities only
• Ignored the descriptions as those texts are not very parallel

Bias, Risks, and Limitations

• Might be slightly outdated (2022)
• Popular languages have more entries
• Labels are not guaranteed to be literal translations (see examples above)