Dataset Viewer
Auto-converted to Parquet Duplicate

The dataset viewer should be available soon. Please retry later.

Davlan/sib200(複製)

データの内容

Flores-200 に基づく、205 の言語と方言のトピック分類のデータ。1 行が 1 文で、text、トピックを表す category、Flores-200 の文番号 index_id を含む。トピックは科学・技術、旅行、政治、スポーツ、健康、エンターテインメント、地理の 7 つである。元の文は新聞社や通信社の記事で、アノテーターは Masakhane から募集された。報道の文章だけなので、他の分野への適用は限られる。

データの概要

項目 値
行数 205,820(205 構成・615 分割の合計)
列数 3
容量 43.87 MB(ファイル 822 本)
ファイルの形式 tsv 148、txt 50
言語 ace, acm, acq, aeb, af, ajp, ak, als, am, apc, ar, ars, ary, arz, as, ast, awa, ayr, azb, azj, ba, bm, ban, be, bem, bn,
公開 public

構成と分割

構成(config) 分割(split)ごとの行数 列数
ace_Arab train 701、validation 99、test 204 3
ace_Latn train 701、validation 99、test 204 3
acm_Arab train 701、validation 99、test 204 3
acq_Arab train 701、validation 99、test 204 3
aeb_Arab train 701、validation 99、test 204 3
afr_Latn train 701、validation 99、test 204 3
ajp_Arab train 701、validation 99、test 204 3
aka_Latn train 701、validation 99、test 204 3
als_Latn train 701、validation 99、test 204 3
amh_Ethi train 701、validation 99、test 204 3
apc_Arab train 701、validation 99、test 204 3
arb_Arab train 701、validation 99、test 204 3
arb_Latn train 701、validation 99、test 204 3
ars_Arab train 701、validation 99、test 204 3
ary_Arab train 701、validation 99、test 204 3
arz_Arab train 701、validation 99、test 204 3
asm_Beng train 701、validation 99、test 204 3
ast_Latn train 701、validation 99、test 204 3
awa_Deva train 701、validation 99、test 204 3
ayr_Latn train 701、validation 99、test 204 3

(ほか 185 構成。全体は HF のデータセットビューアで見られる)

列(構成 ace_Arab・分割 train。どの構成も同じ列)

列 型 先頭の行の値
index_id int64 431
category string geography
text string توركي د كليلڠ لى لاؤة بك تلو لاݢي: لاؤة ايݢيا د …

行数と列は HF のデータセットビューアの集計(元のデータセットの現在の版で数えたもの)。全データセットの一覧は HYUKI5/dataset-catalog にある。

複製について

Davlan/sib200 をリビジョン 38977a667f6fc264d5c26ec57a01e16db040b358 の時点でそのまま複製したもの。中身は変えていない。判断モデル Laya(convaiinnovations/laya)の微調整に使う候補として、調査メモ(Laya 学習用データセット調査)に載せたデータを集めている。

出典

論文

アノテーションの方法

元データ

データの元は報道の分野であり、詳細は **** を参照。

初期のデータ収集と正規化

記事は word-tokenized されているが、正確な前処理のパイプラインに関する情報は利用できない。

元の言語の作成者は誰か?

元の言語は、上記の通信社および新聞社に所属するジャーナリストと作家が作成した。

アノテーションの過程

詳細は ** を参照。

微調整に使うときの注意

  • 調査メモの台帳での区分: B(継承条件・gated・モデル由来の条件・上流の規約などが付く)。区分は記載の文言で機械的に付けたもので、法的な判定ではない。学習に入れる前に元のカードとライセンスの原文で確かめる。
  • Laya の入力は 2,048 token まで(ModernBERT-large / mmBERT)。長い本文は先頭から黙って切られるので、切れる行は事前に除くか、要約・切り出しをしてから入れる。
  • 分割は記事・文書単位で行い、同じ本文から作った別の問いを学習側と較正側に分けない。温度と閾値は学習に使っていない分割で別々に当てる。

調査メモの記述

メモ 節 ライセンス(記載) 確認 注記
multilingual_cjk.md 1. 大規模多言語データセット(XNLI, PAWS-X, MASSIVE, SIB-200, Belebele ほか) (L21) cc-by-sa-4.0(card 本文に "CC 4.0 Commercial" の記載もあり、タグと不整合→要確認) 確認済 確認印なし(各ファイルの凡例により確認済扱い) / 原文注記: 論文 arXiv:2309.07445。1文〜短文でニュース領域のみ。mteb/sib200 もあり(dl=11,604、cc-by-sa-4.0)

元のカード(日本語訳)

以下は元のデータセットのカードの本文を日本語に訳したもの(機械翻訳)。訳に疑問があるときは、末尾の原文を確かめる。


データセットカード: SIB-200

目次

  • 目次
  • データセットの説明
    • データセットの要約
    • 対応タスクとリーダーボード
    • 言語
  • データセットの構造
    • データインスタンス
    • データフィールド
    • データ分割
  • データセットの作成
    • 選定の根拠
    • 元データ
    • アノテーション
    • 個人情報と機微な情報
  • データの利用における考慮事項
    • データセットの社会的影響
    • バイアスに関する議論
    • その他の既知の制限
  • 追加情報
    • データセットの管理者
    • ライセンス情報
    • 引用情報
    • 貢献

データセットの説明

データセットの要約

SIB-200 は、Flores-200 に基づく、205 の言語と方言を対象とした、公開されているトピック分類データセットとして最大のものである。

train/validation/test の各分割は、205 言語すべてについて利用できる。

対応タスクとリーダーボード

  • topic classification: Wikipedia の文を、例えば science/technology、sports、politics などのトピックに分類する。

言語

利用可能な言語は 205 言語である。

データセットの構造

データインスタンス

英語の例は次のようになる。

from datasets import load_dataset
data = load_dataset('Davlan/sib200', 'eng_Latn') 

# Please, specify the language code

# A data point example is below:

{
'label': 0, 
'index_id': 1523, 
'text': 'Mutation adds new genetic variation, and selection removes it from the pool of expressed variation.'
}

データフィールド

  • label: トピックの ID
  • index_id: flores-200 における文の ID
  • text: テキスト

トピックは次のリストに対応する。

"science/technology", "travel", "politics", "sports", "health", "entertainment", "geography"

データ分割

すべての言語について、三つの分割がある。

元の分割名は train、dev、test で、それぞれ train、validation、test の分割に対応する。

分割のサイズは次のとおり。

Language train validation test
英語 701 99 204

データセットの作成

選定の根拠

本データセットは、自然言語処理において十分に扱われてこなかった多くの言語を含む 205 言語について、新しい資源を提供するために導入された。

[More Information Needed]

元データ

データの元は報道の分野であり、詳細は **** を参照。

初期のデータ収集と正規化

記事は word-tokenized されているが、正確な前処理のパイプラインに関する情報は利用できない。

元の言語の作成者は誰か?

元の言語は、上記の通信社および新聞社に所属するジャーナリストと作家が作成した。

アノテーション

アノテーションの過程

詳細は ** を参照。

アノテーターは誰か?

アノテーターは Masakhane から募集された。

個人情報と機微な情報

データは新聞社から得られたものであり、公的な立場の人物や個人への言及のみを含む。

データの利用における考慮事項

データセットの社会的影響

[More Information Needed]

バイアスに関する議論

[More Information Needed]

その他の既知の制限

利用者は、本データセットには報道の文章しか含まれないことを念頭に置くべきである。そのため、開発されたシステムの他の分野への適用は制限される可能性がある。

追加情報

データセットの管理者

ライセンス情報

データのライセンス上の状態は CC 4.0 Commercial である。

引用情報

データセットの参考文献は BibTeX 形式で記載すること。例えば次のとおり。

@misc{adelani2023sib200,
      title={SIB-200: A Simple, Inclusive, and Big Evaluation Dataset for Topic Classification in 200+ Languages and Dialects}, 
      author={David Ifeoluwa Adelani and Hannah Liu and Xiaoyu Shen and Nikita Vassilyev and Jesujoba O. Alabi and Yanke Mao and Haonan Gao and Annie En-Shiun Lee},
      year={2023},
      eprint={2309.07445},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

貢献

本データセットを追加してくれた @dadelani に感謝する。

原文(英語)

Dataset Card for SIB-200

Dataset Summary

SIB-200 is the largest publicly available topic classification dataset based on Flores-200 covering 205 languages and dialects.

The train/validation/test sets are available for all the 205 languages.

Supported Tasks and Leaderboards

  • topic classification: categorize wikipedia sentences into topics e.g science/technology, sports or politics.

Languages

There are 205 languages available :

Dataset Structure

Data Instances

The examples look like this for English:

from datasets import load_dataset
data = load_dataset('Davlan/sib200', 'eng_Latn') 

# Please, specify the language code

# A data point example is below:

{
'label': 0, 
'index_id': 1523, 
'text': 'Mutation adds new genetic variation, and selection removes it from the pool of expressed variation.'
}

Data Fields

  • label: topic id
  • index_id: sentence id in flores-200
  • text: text

The topics correspond to this list:

"science/technology", "travel", "politics", "sports", "health", "entertainment", "geography"

Data Splits

For all languages, there are three splits.

The original splits were named train, dev and test and they correspond to the train, validation and test splits.

The splits have the following sizes :

Language train validation test
English 701 99 204

Dataset Creation

Curation Rationale

The dataset was introduced to introduce new resources for 205 languages, many are under-served for natural language processing.

[More Information Needed]

Source Data

The source of the data is from the news domain, details can be found here ****

Initial Data Collection and Normalization

The articles were word-tokenized, information on the exact pre-processing pipeline is unavailable.

Who are the source language producers?

The source language was produced by journalists and writers employed by the news agency and newspaper mentioned above.

Annotations

Annotation process

Details can be found here **

Who are the annotators?

Annotators were recruited from Masakhane

Personal and Sensitive Information

The data is sourced from newspaper source and only contains mentions of public figures or individuals

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

[More Information Needed]

Other Known Limitations

Users should keep in mind that the dataset only contains news text, which might limit the applicability of the developed systems to other domains.

Additional Information

Dataset Curators

Licensing Information

The licensing status of the data is CC 4.0 Commercial

Citation Information

Provide the BibTex-formatted reference for the dataset. For example:

@misc{adelani2023sib200,
      title={SIB-200: A Simple, Inclusive, and Big Evaluation Dataset for Topic Classification in 200+ Languages and Dialects}, 
      author={David Ifeoluwa Adelani and Hannah Liu and Xiaoyu Shen and Nikita Vassilyev and Jesujoba O. Alabi and Yanke Mao and Haonan Gao and Annie En-Shiun Lee},
      year={2023},
      eprint={2309.07445},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Contributions

Thanks to @dadelani for adding this dataset.

Downloads last month
53

Paper for HYUKI5/Davlan__sib200