The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
shunk031/JGLUE(複製)
データの内容
JGLUE(日本語の言語理解ベンチマーク)を Hugging Face で読み込むための、スクリプトと設定ファイルのデータセット。ビューアで読める行はなく、読み込み時に MARC-ja、JCoLA、JSTS などのタスクの形に整える。MARC-ja は外部からの取得が要ると調査メモにある。
データの概要
| 項目 | 値 |
|---|---|
| 行数・列数 | 不明(データセットビューアで読めない形式。ファイルの内訳だけを示す) |
| 容量 | 548.91 KB(ファイル 14 本) |
| ファイルの形式 | yaml 3、py 3、yml 1、md 1、(拡張子なし) 1、toml 1、ほか |
| 言語 | ja |
| 公開 | public |
全データセットの一覧は HYUKI5/dataset-catalog にある。
複製について
shunk031/JGLUE をリビジョン 41cc99f1c01d41b1ab13435ae97b7076c2199f4c 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
- 元のデータセット: https://huggingface.co/datasets/shunk031/JGLUE
- 複製したリビジョン:
41cc99f1c01d41b1ab13435ae97b7076c2199f4c(2025-03-31 08:17:52+00:00) - 元のライセンス表記:
['cc-by-4.0']
論文
- https://arxiv.org/abs/2309.12676
- https://aclanthology.org/2022.lrec-1.317
- https://aclanthology.org/2020.emnlp-main.369/
- https://aclanthology.org/P16-1168/
- https://aclanthology.org/P18-2124/
- https://aclanthology.org/N19-1421/
- https://aclanthology.org/2022.lrec-1.317/
- https://www.anlp.jp/proceedings/annual_meeting/2022/pdf_dir/E8-4.pdf
- https://www.anlp.jp/proceedings/annual_meeting/2022/pdf_dir/E7-1.pdf
アノテーションの方法
構築の理由
JGLUE の論文 より:
JGLUE は、GLUE および SuperGLUE の幅広いタスクを網羅するよう設計されており、テキスト分類、文対分類、質問応答という 3 種類のタスクからなる。
元の言語の作成者は誰か?
- 元の言語データの作成者は、Amazon の利用者(MARC-ja)、Yahoo! Crowdsourcing のクラウドワーカー(JSTS、JNLI、JCommonsenseQA)、日本語 Wikipedia の執筆者(JSQuAD)、Lancers のクラウドワーカーである。
微調整に使うときの注意
- 調査メモの台帳での区分: D(記載なし・other・unknown・資料間の食い違い)。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 |
2. 日本語データセット (L99) | HF タグは cc-by-4.0。GitHub 側は CC-BY-SA-4.0 → 不一致 | 確認済 | 複数/混在/食い違い / 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: script同梱(JGLUE.py)。MARC-ja は外部取得が必要。dl=2,563、lastModified 2025-03-31。類似: llm-book/JGLUE(同様に script、MARC-ja の前処 |
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。
JGLUE のデータセットカード
このデータセットの読み込みスクリプトは GitHub で開発されている。 お気軽に issue や pull request を送ってほしい。
この読み込みスクリプトのバージョンは、JGLUE のバージョンに対応するよう更新された。 リリース履歴は yahoojapan/JGLUE/releases と shunk031/huggingface-datasets_JGLUE/releases で確認してほしい。
目次
- 目次
- データセットの説明
- データセットの概要
- 対応タスクとリーダーボード
- 言語
- データセットの構造
- データの例
- データのフィールド
- データ分割
- データセットの作成
- 構築の理由
- 元データ
- アノテーション
- 個人情報と機密情報
- データ利用上の留意点
- 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
- データセットの作成者
- ライセンス情報
- 引用情報
- 貢献者
データセットの説明
- ホームページ: https://github.com/yahoojapan/JGLUE
- リポジトリ: https://github.com/shunk031/huggingface-datasets_JGLUE
データセットの概要
JGLUE(Japanese General Language Understanding Evaluation)は、日本語の汎用的な NLU(自然言語理解)能力を測るために構築された。JGLUE は翻訳を介さず、ゼロから構築された。JGLUE が日本語の NLU 研究を促進することを期待している。
JGLUE は、Yahoo Japan Corporation と早稲田大学の Kawahara Lab による共同研究プロジェクトで構築された。
対応タスクとリーダーボード
JGLUE は、テキスト分類、文対分類、QA のタスクからなる。各タスクは複数のデータセットからなる。
対応タスク
MARC-ja
MARC-ja は、テキスト分類タスクのデータセットである。このデータセットは、多言語 Amazon レビューコーパス(MARC)(Keung+, 2020)の日本語部分に基づく。
JCoLA
JCoLA(Japanese Corpus of Linguistic Acceptability)は、日本語の言語モデルを対象とした統語評価のための新しいデータセットであり、言語学者による容認性判断が付いた 10,020 文からなる。文は言語学の学術誌、ハンドブック、教科書から手作業で抽出された。JCoLA は JGLUE ベンチマーク(Kurihara et al., 2022)に含まれている。
JSTS
JSTS は、STS(Semantic Textual Similarity)データセットの日本語版である。STS は、文のペアの意味的類似度を推定するタスクである。JSTS と JNLI(後述)の文は、MS COCO Caption Dataset の日本語版である YJ Captions Dataset(Miyazaki and Shimizu, 2016)から抽出されている。
JNLI
JNLI は、NLI(Natural Language Inference)データセットの日本語版である。NLI は、前提文が仮説文に対してもつ推論関係を認識するタスクである。推論関係は、含意(entailment)、矛盾(contradiction)、中立(neutral)の 3 種類である。
JSQuAD
JSQuAD は、読解のデータセットの一つである SQuAD(Rajpurkar+, 2018)の日本語版である。データセットの各事例は、与えられた文脈(Wikipedia の記事)に関する質問と、その答えからなる。JSQuAD は SQuAD 1.1 に基づいている(答えられない質問はない)。20211101 時点の 日本語 Wikipedia のダンプ を使用した。
JCommonsenseQA
JCommonsenseQA は、常識推論能力を必要とする選択式の質問応答データセットである CommonsenseQA(Talmor+, 2019)の日本語版である。知識ベース ConceptNet から抽出したシードを用いて、クラウドソーシングで構築された。
リーダーボード
リーダーボードは近く公開される予定である。テストセットは、そのときに公開される。
言語
JGLUE の言語データは日本語(BCP-47 ja-JP)である。
データセットの構造
データの例
特定の設定を読み込む場合、利用者はバージョンに依存する接尾辞を付けなければならない:
MARC-ja
from datasets import load_dataset
dataset = load_dataset("shunk031/JGLUE", name="MARC-ja")
print(dataset)
# DatasetDict({
# train: Dataset({
# features: ['sentence', 'label', 'review_id'],
# num_rows: 187528
# })
# validation: Dataset({
# features: ['sentence', 'label', 'review_id'],
# num_rows: 5654
# })
# })
JCoLA
from datasets import load_dataset
dataset = load_dataset("shunk031/JGLUE", name="JCoLA")
print(dataset)
# DatasetDict({
# train: Dataset({
# features: ['uid', 'source', 'label', 'diacritic', 'sentence', 'original', 'translation', 'gloss', 'simple', 'linguistic_phenomenon'],
# num_rows: 6919
# })
# validation: Dataset({
# features: ['uid', 'source', 'label', 'diacritic', 'sentence', 'original', 'translation', 'gloss', 'simple', 'linguistic_phenomenon'],
# num_rows: 865
# })
# validation_out_of_domain: Dataset({
# features: ['uid', 'source', 'label', 'diacritic', 'sentence', 'original', 'translation', 'gloss', 'simple', 'linguistic_phenomenon'],
# num_rows: 685
# })
# validation_out_of_domain_annotated: Dataset({
# features: ['uid', 'source', 'label', 'diacritic', 'sentence', 'original', 'translation', 'gloss', 'simple', 'linguistic_phenomenon'],
# num_rows: 685
# })
# })
JCoLA データセットの例(validation - out of domain annotated)を以下に示す。
{
"uid": 9109,
"source": "Asano_and_Ura_2010",
"label": 1,
"diacritic": "g",
"sentence": "太郎のゴミの捨て方について話した。",
"original": "太郎のゴミの捨て方",
"translation": "‘The way (for Taro) to throw out garbage’",
"gloss": true,
"linguistic_phenomenon": {
"argument_structure": true,
"binding": false,
"control_raising": false,
"ellipsis": false,
"filler_gap": false,
"island_effects": false,
"morphology": false,
"nominal_structure": false,
"negative_polarity_concord_items": false,
"quantifier": false,
"verbal_agreement": false,
"simple": false
}
}
JSTS
from datasets import load_dataset
dataset = load_dataset("shunk031/JGLUE", name="JSTS")
print(dataset)
# DatasetDict({
# train: Dataset({
# features: ['sentence_pair_id', 'yjcaptions_id', 'sentence1', 'sentence2', 'label'],
# num_rows: 12451
# })
# validation: Dataset({
# features: ['sentence_pair_id', 'yjcaptions_id', 'sentence1', 'sentence2', 'label'],
# num_rows: 1457
# })
# })
JSTS データセットの例を以下に示す。
{
"sentence_pair_id": "691",
"yjcaptions_id": "127202-129817-129818",
"sentence1": "街中の道路を大きなバスが走っています。 (A big bus is running on the road in the city.)",
"sentence2": "道路を大きなバスが走っています。 (There is a big bus running on the road.)",
"label": 4.4
}
JNLI
from datasets import load_dataset
dataset = load_dataset("shunk031/JGLUE", name="JNLI")
print(dataset)
# DatasetDict({
# train: Dataset({
# features: ['sentence_pair_id', 'yjcaptions_id', 'sentence1', 'sentence2', 'label'],
# num_rows: 20073
# })
# validation: Dataset({
# features: ['sentence_pair_id', 'yjcaptions_id', 'sentence1', 'sentence2', 'label'],
# num_rows: 2434
# })
# })
JNLI データセットの例を以下に示す。
{
"sentence_pair_id": "1157",
"yjcaptions_id": "127202-129817-129818",
"sentence1": "街中の道路を大きなバスが走っています。 (A big bus is running on the road in the city.)",
"sentence2": "道路を大きなバスが走っています。 (There is a big bus running on the road.)",
"label": "entailment"
}
JSQuAD
from datasets import load_dataset
dataset = load_dataset("shunk031/JGLUE", name="JSQuAD")
print(dataset)
# DatasetDict({
# train: Dataset({
# features: ['id', 'title', 'context', 'question', 'answers', 'is_impossible'],
# num_rows: 62859
# })
# validation: Dataset({
# features: ['id', 'title', 'context', 'question', 'answers', 'is_impossible'],
# num_rows: 4442
# })
# })
JSQuAD の例を以下に示す。
{
"id": "a1531320p0q0",
"title": "東海道新幹線",
"context": "東海道新幹線 [SEP] 1987 年(昭和 62 年)4 月 1 日の国鉄分割民営化により、JR 東海が運営を継承した。西日本旅客鉄道(JR 西日本)が継承した山陽新幹線とは相互乗り入れが行われており、東海道新幹線区間のみで運転される列車にも JR 西日本所有の車両が使用されることがある。2020 年(令和 2 年)3 月現在、東京駅 - 新大阪駅間の所要時間は最速 2 時間 21 分、最高速度 285 km/h で運行されている。",
"question": "2020 年(令和 2 年)3 月現在、東京駅 - 新大阪駅間の最高速度はどのくらいか。",
"answers": {
"text": ["285 km/h"],
"answer_start": [182]
},
"is_impossible": false
}
JCommonsenseQA
from datasets import load_dataset
dataset = load_dataset("shunk031/JGLUE", name="JCommonsenseQA")
print(dataset)
# DatasetDict({
# train: Dataset({
# features: ['q_id', 'question', 'choice0', 'choice1', 'choice2', 'choice3', 'choice4', 'label'],
# num_rows: 8939
# })
# validation: Dataset({
# features: ['q_id', 'question', 'choice0', 'choice1', 'choice2', 'choice3', 'choice4', 'label'],
# num_rows: 1119
# })
# })
JCommonsenseQA の例を以下に示す。
{
"q_id": 3016,
"question": "会社の最高責任者を何というか? (What do you call the chief executive officer of a company?)",
"choice0": "社長 (president)",
"choice1": "教師 (teacher)",
"choice2": "部長 (manager)",
"choice3": "バイト (part-time worker)",
"choice4": "部下 (subordinate)",
"label": 0
}
データのフィールド
MARC-ja
sentence_pair_id: 文対の IDyjcaptions_id: yjcaptions 内の文の ID(後述)sentence1: 1 つ目の文sentence2: 2 つ目の文label: 文の類似度: 5(意味が同等)から 0(意味がまったく異なる)まで
yjcaptions_id の説明
JGLUE の README.md より、次の 2 つの場合がある:
- 1 枚の画像内の文対:
(image id)-(sentence1 id)-(sentence2 id)- 例: 723-844-847
- 先頭が "g" の文 ID は、クラウドワーカーが生成した文を意味する(例: 69501-75698-g103)。JNLI のみ
- 2 枚の画像にまたがる文対:
(image id of sentence1)_(image id of sentence2)-(sentence1 id)-(sentence2 id)- 例: 91337_217583-96105-91680
JCoLA
JCoLA の README.md および JCoLA の論文 より
uid: 文の一意な IDsource: 元の論文の著者と出版年label: 容認性判断のラベル(非容認は 0、容認は 1)diacritic: 元の論文に記載されていた容認性判断の表記sentence: 文(必要に応じて作成者が修正したもの)original: 元の論文に掲載されていた原文translation: 元の論文に掲載されていた文の英訳(ある場合のみ)gloss: 元の論文に掲載されていた文の gloss(ある場合のみ)linguistic_phenomenonargument_structure: 項の語順と格標示に基づく容認性判断binding: 名詞句の束縛に基づく容認性判断control_raising: コントロール述語・繰り上げ述語に分類される述語に基づく容認性判断ellipsis: 文中の要素を省略できるかどうかに基づく容認性判断filler_gap: 移動された要素と空所の依存関係に基づく容認性判断island effects: wh 移動などの filler-gap 依存に対する制限に基づく容認性判断(島効果)morphology: 形態論に基づく容認性判断nominal_structure: 名詞句の内部構造に基づく容認性判断negative_polarity_concord_items: 否定極性項目・否定一致項目(NPIs/NCIs)が現れることのできる位置の制限に基づく容認性判断quantifiers: 浮遊数量詞などの数量詞の分布に基づく容認性判断verbal_agreement: 主語と動詞の依存関係に基づく容認性判断simple: 標示された統語構造をもたない容認性判断
JNLI
sentence_pair_id: 文対の IDyjcaptions_id: yjcaptions 内の文の IDsentence1: 前提文sentence2: 仮説文label: 推論関係
JSQuAD
title: Wikipedia 記事の題名paragraphs: 段落の集合qas: 質問と答えの組の集合question: 質問id: 質問の IDanswers: 答えの集合text: 答えの文字列answer_start: 開始位置(文字インデックス)is_impossible: すべての値が false であるcontext: 題名と段落を連結したもの
JCommonsenseQA
q_id: 質問の IDquestion: 質問choice{0..4}: 選択肢label: 正解の選択肢の ID
データ分割
現在利用できるのは train と dev の分割のみで、test の分割はリーダーボードが公開された後に利用可能になる。
JCoLA の論文 より:
ドメイン内(in-domain)のデータは、訓練データ(6,919 件)、開発データ(865 件)、テストデータ(865 件)に分割される。一方、ドメイン外(out-of-domain)のデータは評価のみに用いられ、開発データ(685 件)とテストデータ(686 件)に分けられる。
| タスク | データセット | 訓練 | 開発 | テスト |
|---|---|---|---|---|
| テキスト分類 | MARC-ja | 187,528 | 5,654 | 5,639 |
| JCoLA | 6,919 | 865† / 685‡ | 865† / 685‡ | |
| 文対分類 | JSTS | 12,451 | 1,457 | 1,589 |
| JNLI | 20,073 | 2,434 | 2,508 | |
| 質問応答 | JSQuAD | 62,859 | 4,442 | 4,420 |
| JCommonsenseQA | 8,939 | 1,119 | 1,118 |
JCoLA: † はドメイン内。‡ はドメイン外。
データセットの作成
構築の理由
JGLUE の論文 より:
JGLUE は、GLUE および SuperGLUE の幅広いタスクを網羅するよう設計されており、テキスト分類、文対分類、質問応答という 3 種類のタスクからなる。
元データ
初期のデータ収集と正規化
[追加情報が必要]
元の言語の作成者は誰か?
- 元の言語データの作成者は、Amazon の利用者(MARC-ja)、Yahoo! Crowdsourcing のクラウドワーカー(JSTS、JNLI、JCommonsenseQA)、日本語 Wikipedia の執筆者(JSQuAD)、Lancers のクラウドワーカーである。
アノテーション
アノテーションの過程
MARC-ja
JGLUE の論文 より:
テキスト分類データセットの一つとして、Multilingual Amazon Reviews Corpus(MARC)(Keung et al., 2020)に基づくデータセットを構築した。MARC は、Amazon の購入サイト上の商品レビューからなる多言語コーパスで、5 段階の星評価(1〜5)が付いている。このコーパスは英語や日本語を含む 6 言語を対象としている。JGLUE では MARC の日本語部分を用いた。また、人間にも計算機にもクラスラベルを判定しやすくするため、テキスト分類タスクを二値分類タスクとした。星 1 と 2 のレビューは “negative” に、星 4 と 5 のレビューは “positive” に変換した。星 3 のレビューは用いていない。
MARC の問題の一つは、評価と本文の内容が食い違うデータが含まれることがある点である。例えば、内容が肯定的なレビューに星 1 や 2 が付いている場合がこれにあたる。こうしたデータはデータセットの質を下げる。評価に用いる開発・テストの事例の質を高めるため、約 12,000 件のレビューについて、ポジティブかネガティブかを判定するタスクをクラウドソーシングで行った。10 人中 7 人以上のワーカーが同じ票を投じたレビューのみを採用し、最多票のラベルをそのレビューに割り当てた。得られたレビューを開発データとテストデータに分割した。
上記の手順により、開発データとテストデータとしてそれぞれ 5,654 件と 5,639 件の事例を得た。訓練データは事例数が多いため、清掃の手順は行わず、187,528 件の事例を MARC から直接抽出した。MARC-ja の統計は表 2 に示す。MARC-ja の評価指標には正解率(accuracy)を用いた。テキストの二値分類タスクだからである。
JCoLA
JCoLA の論文 より:
3 JCoLA
本研究では、JCoLA(Japanese Corpus of Linguistic Acceptability)を紹介する。これは、日本語を対象とした初の大規模な容認性判断タスクのデータセットとなるものである。JCoLA は、日本語統語論の教科書やハンドブックの文、および理論言語学の有力誌の一つである JEAL(Journal of East Asian Linguistics)に掲載された日本語統語論の論文の文からなる。
3.1 データ収集
JCoLA の文は、日本語統語論に焦点を当てた主要な教科書やハンドブックから収集した。本文に加えて、脚注に含まれる例文も収集の対象とした。また、JEAL(Journal of East Asian Linguistics)に掲載された日本語統語論の論文からも容認性判断を収集した。具体的には、2006 年から 2015 年に JEAL に掲載された全論文(計 133 本)を調べ、日本語統語論に関する 26 本の論文から 2,252 件の容認性判断を抽出した(表 2)。容認性判断には付録や脚注の文を含めたが、統語構造の分析のために提示された文(統語構造を示す括弧付きの文など)は含めなかった。その結果、計 11,984 文の例文を収集した。これを基に、後続の節で説明する方法論により JCoLA を構築した。
JSTS と JNLI
JGLUE の論文 より:
文対分類のデータセットとして、意味的テキスト類似度(STS)データセット JSTS と、自然言語推論(NLI)データセット JNLI を構築した。
概要
STS は、文のペアの意味的類似度を推定するタスクである。正解の類似度は通常、クラウドソーシングで複数のワーカーが付けた 0(意味がまったく異なる)から 5(意味が同等)までの整数値の平均として与えられる。
NLI は、前提文が仮説文に対してもつ推論関係を認識するタスクである。推論関係は一般に、“entailment”、“contradiction”、“neutral” の 3 つのラベルで定義される。正解の推論関係は、クラウドソーシングで複数のワーカーから回答を集めた後、多数決で付与されることが多い。
STS タスクと NLI タスクについて、それぞれ GLUE に含まれる STS-B(Cer et al., 2017)と MultiNLI(Williams et al., 2018)がある。日本語のデータセットとしては、NLI データセット SNLI(Stanford NLI)の機械翻訳版である JSNLI(Yoshikoshi et al., 2020)と、STS/NLI データセット SICK(Marelli et al., 2014)の人手翻訳版である JSICK(Yanaka and Mineshima, 2021)がある。第 1 節で述べたとおり、これらは自動翻訳や人手翻訳に起因する問題を抱えている。この問題を解決するため、日本語の STS/NLI データセットをゼロから構築した。JSTS と JNLI の文対は、基本的に MS COCO Caption Dataset(Chen et al., 2015)の日本語版である YJ Captions Dataset(Miyazaki and Shimizu, 2016)から抽出した。JSTS と JNLI の文対の多くは重複しており、SICK と JSICK のように、同じ文対について類似度と推論関係の関係を分析できる。
JSTS の類似度は、STS-B と同様に 0 から 5 までの実数で与えられる。JNLI の推論関係は、SNLI や MultiNLI と同様に上記の 3 つのラベルから与えられる。推論関係の定義も SNLI に基づく。
構築方法
JSTS と JNLI の構築の流れは図 1 に示す。基本的には、YJ Captions の同一画像に対する 2 つのキャプションを文対として用いる。これらの文対について、類似度と含意・中立の NLI 関係をクラウドソーシングで得る。ただし、同一画像のキャプションから、類似度が低い文対や矛盾関係の文対を集めるのは難しい。この問題を解決するため、類似度が低い文対は異なる画像のキャプションから集めた。矛盾関係は、与えられたキャプションに対して矛盾する文を書くようワーカーに依頼して集めた。
JSTS と JNLI の構築手順の詳細は以下のとおりである。
- YJ Captions の同一画像に対する 2 つのキャプションを使い、STS タスクをクラウドソーシングで行った。2 つのキャプションの類似度を 5 人のワーカーに答えてもらい、その平均値を正解の類似度とした。回答のばらつきが大きい文対は回答の質が低いとみなし、削除した。このタスクを 16,000 組の文対について行い、類似度の分散が 1.0 以上の文対を削除した結果、正解の類似度が付いた 10,236 組の文対が得られた。このデータを JSTS-A と呼ぶ。
- 類似度が低い文対を集めるため、異なる画像のキャプションの文対を用いて、手順 1 と同じ STS タスクをクラウドソーシングで行った。このタスクを 4,000 組の文対について行い、正解の類似度が付いた 2,970 組の文対を集めた。このデータを JSTS-B と呼ぶ。
- JSTS-A については、NLI タスクをクラウドソーシングで行った。推論関係には向きがあるため、文対について両方向の推論関係を得た。前述のとおり、同一画像のキャプションから作られた JSTS-A からは矛盾の事例を集めにくいため、この手順では含意と中立の事例を集めた。推論関係の回答は 10 人のワーカーから集めた。6 人以上が同じ回答をした場合、その回答が含意または中立であれば正解のラベルとして採用した。JSTS-A について両方向の推論関係を得るため、JSTS-A の 2 倍にあたる 20,472 組の文対についてこのタスクを行った。その結果、17,501 組の文対について推論関係が得られた。このデータを JNLI-A と呼ぶ。異なる画像のキャプション同士の推論関係は定義・判定が難しいため、JSTS-B は NLI タスクには使わなかった。
- 矛盾の NLI 事例を集めるため、YJCaptions の各キャプションに対して矛盾する文を 4 つ書くタスクをクラウドソーシングで行った。書かれた文のうち、短い文や元の文との関連性が低い文などの低品質な文を除くため、編集距離が 0.75 以上の文対を削除した。さらに、作成した文対が矛盾しているかを確かめるため、10 人のワーカーによる片方向の NLI タスクを行った。6 人以上のワーカーが矛盾と回答した文対のみを採用した。矛盾関係には向きがないため、採用した文対の逆方向にも矛盾関係を自動的に付与した。1,800 件のキャプションから 7,200 組の文対を得て、そのうち片方向の矛盾関係を付与した 3,779 組を集めた。逆方向にも矛盾関係を自動付与したことで、事例数は 2 倍の 7,558 件になった。このデータを JNLI-C と呼ぶ。
- 手順 4 で集めた 3,779 組の文対について、手順 1 と 2 と同じ方法で類似度の付与と絞り込みを行った STS タスクをクラウドソーシングで行った。この方法で、3,779 組から正解の類似度が付いた 2,303 組の文対を集めた。このデータを JSTS-C と呼ぶ。
JSQuAD
JGLUE の論文 より:
質問応答のデータセットとして、読解データセットの一つである SQuAD(Rajpurkar et al., 2016)の日本語版と、次節で説明する CommonsenseQA の日本語版を構築した。
読解とは、文書を読み、それについての質問に答えるタスクである。読解の評価セットは英語で多く構築されており、その後に他言語や多言語のものも構築されてきた。
日本語では、クイズを対象とした読解データセット(Suzukietal.,2018)や、運転分野を対象としたもの(Takahashi et al., 2019)は構築されているが、一般分野のものはない。一般分野のデータセットを構築するため、Wikipedia を用いた。構築の過程は基本的に SQuAD 1.1(Rajpurkar et al., 2016)に従う。
まず、Wikipedia から質の高い記事を抽出するため、Wikipedia 内のハイパーリンクに基づいて記事の質を推定する Nayuki を用いた。上位 10,000 記事から 822 記事を無作為に選んだ。例えば、記事には “熊本県 (Kumamoto Prefecture)” や “フランス料理 (French cuisine)” が含まれる。次に、記事を段落に分け、各段落をクラウドワーカーに提示して、段落を理解すれば答えられる質問と答えを書いてもらった。図 2 に JSQuAD の例を示す。システムの評価を頑健にするため、開発セットとテストセットについては、ワーカーにさらに 2 つの答えを書いてもらった。
JCommonsenseQA
JGLUE の論文 より:
概要
JCommonsenseQA は、常識推論能力を評価するための 5 択の QA からなる CommonsenseQA(Talmor et al., 2019)の日本語版である。図 3 に JCommonsenseQA の例を示す。CommonsenseQA と同様に、JCommonsenseQA は知識ベース ConceptNet(Speer et al., 2017)から抽出したシードを用いて、クラウドソーシングで構築されている。ConceptNet は、2 つの概念とその関係の三つ組からなる多言語の知識ベースである。三つ組には向きがあり、(source concept, relation, target concept) の形で表される。例えば (bullet train, AtLocation, station) がある。
構築方法
JCommonsenseQA の構築の流れは図 4 に示す。まず、ConceptNet から質問セット(QS)を集めた。各 QS は、ある起点概念と、その起点概念と同じ関係をもつ 3 つの目標概念からなる。次に、各 QS について、目標概念のうち 1 つだけを答えとする質問を書くタスクと、2 つの妨害選択肢を加えるタスクをクラウドソーシングで行った。以下では、JCommonsenseQA の構築手順の詳細を、CommonsenseQA との違いを示しながら説明する。
- ConceptNet から日本語の QS を集めた。CommonsenseQA は、“RelatedTo” や “IsA” のような一般的な関係を除いた順方向の関係(source concept, relation, target concept)のみを用いる。JCommonsenseQA も同様に、一般的な関係を除いた 22 種類の関係5 の集合を用いるが、より多様な質問とするため関係の向きを両方向とした。すなわち、逆方向の関係(source concept, relation−1, target concept)6 も用いた。この設定で、日本語の起点概念と目標概念をもつ QS を 43,566 件抽出し、そこから 7,500 件を無作為に選んだ。
- CommonsenseQA の低品質な質問には、答えとみなせる妨害選択肢を含むものがある。妨害選択肢の質を高めるため、CommonsenseQA では採用されていない次の 2 つの処理を加えた。第一に、QS の 3 つの目標概念に互いの表記揺れや同義語が含まれる場合、その QS を削除した。表記揺れの識別には、形態素解析辞書 Juman Dic7 の単語 ID を用いた。第二に、目標概念に同義語が含まれるかを判定するタスクをクラウドソーシングで行った。その結果、7,500 件の QS から 5,920 件を採用した。
- 各 QS について、3 つの目標概念のうち 1 つだけが答えになる質問文を書くタスクをクラウドソーシングで行った。図 4 の例では “駅 (station)” が答えで、残りは妨害選択肢である。低品質な質問文を除くため、次の質問文を削除した。
- 選択肢の語を含む質問文(そのような質問は容易に解けてしまうため)。
- “XX characters” という表現を含む質問文(XX は数字)。8
- “?” で終わらない、形式の誤った質問文。
- その結果、5,920 × 3 = 17,760 個の質問文が作られ、不適切な質問文を削除して 15,310 個を採用した。
- CommonsenseQA では、妨害選択肢を加える際に、一方は ConceptNet から選び、もう一方はクラウドソーシングで作成していた。JCommonsenseQA では、妨害選択肢の種類をより多様にするため、ConceptNet から選ぶ代わりに 2 つともクラウドソーシングで作成した。質問の質を高めるため9、加えられた妨害選択肢が次のいずれかに当てはまる質問を削除した:
- 妨害選択肢が質問文に含まれている。
- 妨害選択肢が既存の選択肢のいずれかと重複している。
- その結果、15,310 個の質問に妨害選択肢が加えられ、そのうち 13,906 個を採用した。
- 各質問に 3 人のクラウドワーカーに回答してもらい、少なくとも 2 人が正解した質問のみを採用した。その結果、13,906 個の質問のうち 11,263 個を採用した。
アノテーターは誰か?
データセットの構築におけるすべてのクラウドソーシングのタスクに Yahoo! Crowdsourcing を使用した。
JCoLA の論文 より:
JCoLA における正解率の上限の目安として、JCoLA データの一部を用いて Lancers2 上で人間による容認性判断の実験を行った。
個人情報と機密情報
[追加情報が必要]
データ利用上の留意点
データセットの社会的影響
JGLUE の論文 より:
翻訳を介さずゼロから、日本語の汎用的な NLU 能力を測る日本語 NLU ベンチマーク JGLUE を構築した。JGLUE が日本語の NLU 研究を促進することを期待する。
バイアスに関する議論
[追加情報が必要]
その他の既知の制限
JCoLA の論文 より:
JCoLA に含まれるすべての文は、理論統語論の教科書、ハンドブック、論文から抽出されたものである。そのため、これらの文は理論的に意味があることが保証されており、JCoLA を難しいデータセットにしている。しかし、言語現象の分布は元の文献の分布を直接反映しており、極端に偏っている。実際、表 3 に見られるように、ほとんどの言語現象では文数が 100 を超えるが、約 10 文しかない言語現象もいくつかある。また、言語モデルに特定の文脈が与えられた文を解釈させるのは難しいため、非容認性が文脈に依存する文は JCoLA から必然的に除かれた。この除去の過程により、一部の言語現象(省略など)では非容認の文が削除され、その結果、容認の文と非容認の文のバランスが偏った(容認の文の割合が高くなった)。
追加情報
- 日本語言語理解ベンチマーク JGLUE の構築 〜 自然言語処理モデルの評価用データセットを公開しました - Yahoo! JAPAN Tech Blog https://techblog.yahoo.co.jp/entry/2022122030379907/
データセットの作成者
MARC-ja
- Keung, Phillip, et al. "The Multilingual Amazon Reviews Corpus." Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2020.
JCoLA
- Someya, Sugimoto, and Oseki. "JCoLA: Japanese Corpus of Linguistic Acceptability." arxiv preprint arXiv:2309.12676 (2023).
JSTS と JNLI
- Miyazaki, Takashi, and Nobuyuki Shimizu. "Cross-lingual image caption generation." Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2016.
JSQuAD
JGLUE の作成者は、JSQuAD の元データを日本語 Wikipedia のダンプから整理した。
JCommonsenseQA
CommonsenseQA と同様に、JCommonsenseQA は知識ベース ConceptNet から抽出したシードを用いて、クラウドソーシングで構築されている。
ライセンス情報
JGLUE
本作品は Creative Commons Attribution-ShareAlike 4.0 International License の下で提供される。
JCoLA
本コーパスのテキストは既刊の著作物から抜粋したものであり、著作権(該当する場合)は元の著者または出版社に帰属する。日本国内での研究目的の利用はフェアユースとして適法であると考えるが、その保証はしない。
引用情報
JGLUE
@inproceedings{kurihara-lrec-2022-jglue,
title={JGLUE: Japanese general language understanding evaluation},
author={Kurihara, Kentaro and Kawahara, Daisuke and Shibata, Tomohide},
booktitle={Proceedings of the Thirteenth Language Resources and Evaluation Conference},
pages={2957--2966},
year={2022},
url={https://aclanthology.org/2022.lrec-1.317/}
}
@inproceedings{kurihara-nlp-2022-jglue,
title={JGLUE: 日本語言語理解ベンチマーク},
author={栗原健太郎 and 河原大輔 and 柴田知秀},
booktitle={言語処理学会第 28 回年次大会},
pages={2023--2028},
year={2022},
url={https://www.anlp.jp/proceedings/annual_meeting/2022/pdf_dir/E8-4.pdf},
note={in Japanese}
}
MARC-ja
@inproceedings{marc_reviews,
title={The Multilingual Amazon Reviews Corpus},
author={Keung, Phillip and Lu, Yichao and Szarvas, György and Smith, Noah A.},
booktitle={Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing},
year={2020}
}
JCoLA
@article{someya-arxiv-2023-jcola,
title={JCoLA: Japanese Corpus of Linguistic Acceptability},
author={Taiga Someya and Yushi Sugimoto and Yohei Oseki},
year={2023},
eprint={2309.12676},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
@inproceedings{someya-nlp-2022-jcola,
title={日本語版 CoLA の構築},
author={染谷 大河 and 大関 洋平},
booktitle={言語処理学会第 28 回年次大会},
pages={1872--1877},
year={2022},
url={https://www.anlp.jp/proceedings/annual_meeting/2022/pdf_dir/E7-1.pdf},
note={in Japanese}
}
JSTS と JNLI
@inproceedings{miyazaki2016cross,
title={Cross-lingual image caption generation},
author={Miyazaki, Takashi and Shimizu, Nobuyuki},
booktitle={Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
pages={1780--1790},
year={2016}
}
貢献者
JGLUE データセットを作成してくれた Kentaro Kurihara、Daisuke Kawahara、Tomohide Shibata に感謝する。 JCoLA データセットを作成してくれた Taiga Someya に感謝する。
原文(英語)
Dataset Card for JGLUE
This dataset loading script is developed on GitHub. Please feel free to open an issue or pull request.
The version of this loading script has been updated to correspond to the version of JGLUE. Please check the release history at yahoojapan/JGLUE/releases and shunk031/huggingface-datasets_JGLUE/releases.
Dataset Summary
From JGLUE's README.md:
JGLUE, Japanese General Language Understanding Evaluation, is built to measure the general NLU ability in Japanese. JGLUE has been constructed from scratch without translation. We hope that JGLUE will facilitate NLU research in Japanese.
JGLUE has been constructed by a joint research project of Yahoo Japan Corporation and Kawahara Lab at Waseda University.
Supported Tasks and Leaderboards
From JGLUE's README.md:
JGLUE consists of the tasks of text classification, sentence pair classification, and QA. Each task consists of multiple datasets.
Supported Tasks
MARC-ja
From JGLUE's README.md:
MARC-ja is a dataset of the text classification task. This dataset is based on the Japanese portion of Multilingual Amazon Reviews Corpus (MARC) (Keung+, 2020).
JCoLA
From JCoLA's README.md
JCoLA (Japanese Corpus of Linguistic Accept010 ability) is a novel dataset for targeted syntactic evaluations of language models in Japanese, which consists of 10,020 sentences with acceptability judgments by linguists. The sentences are manually extracted from linguistics journals, handbooks and textbooks. JCoLA is included in JGLUE benchmark (Kurihara et al., 2022).
JSTS
From JGLUE's README.md:
JSTS is a Japanese version of the STS (Semantic Textual Similarity) dataset. STS is a task to estimate the semantic similarity of a sentence pair. The sentences in JSTS and JNLI (described below) are extracted from the Japanese version of the MS COCO Caption Dataset, the YJ Captions Dataset (Miyazaki and Shimizu, 2016).
JNLI
From JGLUE's README.md:
JNLI is a Japanese version of the NLI (Natural Language Inference) dataset. NLI is a task to recognize the inference relation that a premise sentence has to a hypothesis sentence. The inference relations are entailment, contradiction, and neutral.
JSQuAD
From JGLUE's README.md:
JSQuAD is a Japanese version of SQuAD (Rajpurkar+, 2018), one of the datasets of reading comprehension. Each instance in the dataset consists of a question regarding a given context (Wikipedia article) and its answer. JSQuAD is based on SQuAD 1.1 (there are no unanswerable questions). We used the Japanese Wikipedia dump as of 20211101.
JCommonsenseQA
From JGLUE's README.md:
JCommonsenseQA is a Japanese version of CommonsenseQA (Talmor+, 2019), which is a multiple-choice question answering dataset that requires commonsense reasoning ability. It is built using crowdsourcing with seeds extracted from the knowledge base ConceptNet.
Leaderboard
From JGLUE's README.md:
A leaderboard will be made public soon. The test set will be released at that time.
Languages
The language data in JGLUE is in Japanese (BCP-47 ja-JP).
Dataset Structure
Data Instances
When loading a specific configuration, users has to append a version dependent suffix:
MARC-ja
from datasets import load_dataset
dataset = load_dataset("shunk031/JGLUE", name="MARC-ja")
print(dataset)
# DatasetDict({
# train: Dataset({
# features: ['sentence', 'label', 'review_id'],
# num_rows: 187528
# })
# validation: Dataset({
# features: ['sentence', 'label', 'review_id'],
# num_rows: 5654
# })
# })
JCoLA
from datasets import load_dataset
dataset = load_dataset("shunk031/JGLUE", name="JCoLA")
print(dataset)
# DatasetDict({
# train: Dataset({
# features: ['uid', 'source', 'label', 'diacritic', 'sentence', 'original', 'translation', 'gloss', 'simple', 'linguistic_phenomenon'],
# num_rows: 6919
# })
# validation: Dataset({
# features: ['uid', 'source', 'label', 'diacritic', 'sentence', 'original', 'translation', 'gloss', 'simple', 'linguistic_phenomenon'],
# num_rows: 865
# })
# validation_out_of_domain: Dataset({
# features: ['uid', 'source', 'label', 'diacritic', 'sentence', 'original', 'translation', 'gloss', 'simple', 'linguistic_phenomenon'],
# num_rows: 685
# })
# validation_out_of_domain_annotated: Dataset({
# features: ['uid', 'source', 'label', 'diacritic', 'sentence', 'original', 'translation', 'gloss', 'simple', 'linguistic_phenomenon'],
# num_rows: 685
# })
# })
An example of the JCoLA dataset (validation - out of domain annotated) looks as follows:
{
"uid": 9109,
"source": "Asano_and_Ura_2010",
"label": 1,
"diacritic": "g",
"sentence": "太郎のゴミの捨て方について話した。",
"original": "太郎のゴミの捨て方",
"translation": "‘The way (for Taro) to throw out garbage’",
"gloss": true,
"linguistic_phenomenon": {
"argument_structure": true,
"binding": false,
"control_raising": false,
"ellipsis": false,
"filler_gap": false,
"island_effects": false,
"morphology": false,
"nominal_structure": false,
"negative_polarity_concord_items": false,
"quantifier": false,
"verbal_agreement": false,
"simple": false
}
}
JSTS
from datasets import load_dataset
dataset = load_dataset("shunk031/JGLUE", name="JSTS")
print(dataset)
# DatasetDict({
# train: Dataset({
# features: ['sentence_pair_id', 'yjcaptions_id', 'sentence1', 'sentence2', 'label'],
# num_rows: 12451
# })
# validation: Dataset({
# features: ['sentence_pair_id', 'yjcaptions_id', 'sentence1', 'sentence2', 'label'],
# num_rows: 1457
# })
# })
An example of the JSTS dataset looks as follows:
{
"sentence_pair_id": "691",
"yjcaptions_id": "127202-129817-129818",
"sentence1": "街中の道路を大きなバスが走っています。 (A big bus is running on the road in the city.)",
"sentence2": "道路を大きなバスが走っています。 (There is a big bus running on the road.)",
"label": 4.4
}
JNLI
from datasets import load_dataset
dataset = load_dataset("shunk031/JGLUE", name="JNLI")
print(dataset)
# DatasetDict({
# train: Dataset({
# features: ['sentence_pair_id', 'yjcaptions_id', 'sentence1', 'sentence2', 'label'],
# num_rows: 20073
# })
# validation: Dataset({
# features: ['sentence_pair_id', 'yjcaptions_id', 'sentence1', 'sentence2', 'label'],
# num_rows: 2434
# })
# })
An example of the JNLI dataset looks as follows:
{
"sentence_pair_id": "1157",
"yjcaptions_id": "127202-129817-129818",
"sentence1": "街中の道路を大きなバスが走っています。 (A big bus is running on the road in the city.)",
"sentence2": "道路を大きなバスが走っています。 (There is a big bus running on the road.)",
"label": "entailment"
}
JSQuAD
from datasets import load_dataset
dataset = load_dataset("shunk031/JGLUE", name="JSQuAD")
print(dataset)
# DatasetDict({
# train: Dataset({
# features: ['id', 'title', 'context', 'question', 'answers', 'is_impossible'],
# num_rows: 62859
# })
# validation: Dataset({
# features: ['id', 'title', 'context', 'question', 'answers', 'is_impossible'],
# num_rows: 4442
# })
# })
An example of the JSQuAD looks as follows:
{
"id": "a1531320p0q0",
"title": "東海道新幹線",
"context": "東海道新幹線 [SEP] 1987 年(昭和 62 年)4 月 1 日の国鉄分割民営化により、JR 東海が運営を継承した。西日本旅客鉄道(JR 西日本)が継承した山陽新幹線とは相互乗り入れが行われており、東海道新幹線区間のみで運転される列車にも JR 西日本所有の車両が使用されることがある。2020 年(令和 2 年)3 月現在、東京駅 - 新大阪駅間の所要時間は最速 2 時間 21 分、最高速度 285 km/h で運行されている。",
"question": "2020 年(令和 2 年)3 月現在、東京駅 - 新大阪駅間の最高速度はどのくらいか。",
"answers": {
"text": ["285 km/h"],
"answer_start": [182]
},
"is_impossible": false
}
JCommonsenseQA
from datasets import load_dataset
dataset = load_dataset("shunk031/JGLUE", name="JCommonsenseQA")
print(dataset)
# DatasetDict({
# train: Dataset({
# features: ['q_id', 'question', 'choice0', 'choice1', 'choice2', 'choice3', 'choice4', 'label'],
# num_rows: 8939
# })
# validation: Dataset({
# features: ['q_id', 'question', 'choice0', 'choice1', 'choice2', 'choice3', 'choice4', 'label'],
# num_rows: 1119
# })
# })
An example of the JCommonsenseQA looks as follows:
{
"q_id": 3016,
"question": "会社の最高責任者を何というか? (What do you call the chief executive officer of a company?)",
"choice0": "社長 (president)",
"choice1": "教師 (teacher)",
"choice2": "部長 (manager)",
"choice3": "バイト (part-time worker)",
"choice4": "部下 (subordinate)",
"label": 0
}
Data Fields
MARC-ja
sentence_pair_id: ID of the sentence pairyjcaptions_id: sentence ids in yjcaptions (explained below)sentence1: first sentencesentence2: second sentencelabel: sentence similarity: 5 (equivalent meaning) - 0 (completely different meaning)
Explanation for yjcaptions_id
From JGLUE's README.md, there are the following two cases:
- sentence pairs in one image:
(image id)-(sentence1 id)-(sentence2 id)- e.g., 723-844-847
- a sentence id starting with "g" means a sentence generated by a crowdworker (e.g., 69501-75698-g103): only for JNLI
- sentence pairs in two images:
(image id of sentence1)_(image id of sentence2)-(sentence1 id)-(sentence2 id)- e.g., 91337_217583-96105-91680
JCoLA
From JCoLA's README.md and JCoLA's paper
uid: unique id of the sentencesource: author and the year of publication of the source articlelabel: acceptability judgement label (0 for unacceptable, 1 for acceptable)diacritic: acceptability judgement as originally notated in the source articlesentence: sentence (modified by the author if needed)original: original sentence as presented in the source articletranslation: English translation of the sentence as presentend in the source article (if any)gloss: gloss of the sentence as presented in the source article (if any)linguistic_phenomenonargument_structure: acceptability judgements based on the order of arguments and case markingbinding: acceptability judgements based on the binding of noun phrasescontrol_raising: acceptability judgements based on predicates that are categorized as control or raisingellipsis: acceptability judgements based on the possibility of omitting elements in the sentencesfiller_gap: acceptability judgements based on the dependency between the moved element and the gapisland effects: acceptability judgements based on the restrictions on filler-gap dependencies such as wh-movementsmorphology: acceptability judgements based on the morphologynominal_structure: acceptability judgements based on the internal structure of noun phrasesnegative_polarity_concord_items: acceptability judgements based on the restrictions on where negative polarity/concord items (NPIs/NCIs) can appearquantifiers: acceptability judgements based on the distribution of quantifiers such as floating quantifiersverbal_agreement: acceptability judgements based on the dependency between subjects and verbssimple: acceptability judgements that do not have marked syntactic structures
JNLI
sentence_pair_id: ID of the sentence pairyjcaptions_id: sentence ids in the yjcaptionssentence1: premise sentencesentence2: hypothesis sentencelabel: inference relation
JSQuAD
title: title of a Wikipedia articleparagraphs: a set of paragraphsqas: a set of pairs of a question and its answerquestion: questionid: id of a questionanswers: a set of answerstext: answer textanswer_start: start position (character index)is_impossible: all the values are falsecontext: a concatenation of the title and paragraph
JCommonsenseQA
q_id: ID of the questionquestion: questionchoice{0..4}: choicelabel: correct choice id
Data Splits
From JGLUE's README.md:
Only train/dev sets are available now, and the test set will be available after the leaderboard is made public.
From JCoLA's paper:
The in-domain data is split into training data (6,919 instances), development data (865 instances), and test data (865 instances). On the other hand, the out-of-domain data is only used for evaluation, and divided into development data (685 instances) and test data (686 instances).
| Task | Dataset | Train | Dev | Test |
|---|---|---|---|---|
| Text Classification | MARC-ja | 187,528 | 5,654 | 5,639 |
| JCoLA | 6,919 | 865† / 685‡ | 865† / 685‡ | |
| Sentence Pair Classification | JSTS | 12,451 | 1,457 | 1,589 |
| JNLI | 20,073 | 2,434 | 2,508 | |
| Question Answering | JSQuAD | 62,859 | 4,442 | 4,420 |
| JCommonsenseQA | 8,939 | 1,119 | 1,118 |
JCoLA: † in domain. ‡ out of domain.
Dataset Creation
Curation Rationale
From JGLUE's paper:
JGLUE is designed to cover a wide range of GLUE and SuperGLUE tasks and consists of three kinds of tasks: text classification, sentence pair classification, and question answering.
Source Data
Initial Data Collection and Normalization
[More Information Needed]
Who are the source language producers?
- The source language producers are users of Amazon (MARC-ja), crowd-workers of Yahoo! Crowdsourcing (JSTS, JNLI and JCommonsenseQA), writers of the Japanese Wikipedia (JSQuAD), crowd-workers of Lancers.
Annotations
Annotation process
MARC-ja
From JGLUE's paper:
As one of the text classification datasets, we build a dataset based on the Multilingual Amazon Reviews Corpus (MARC) (Keung et al., 2020). MARC is a multilingual corpus of product reviews with 5-level star ratings (1-5) on the Amazon shopping site. This corpus covers six languages, including English and Japanese. For JGLUE, we use the Japanese part of MARC and to make it easy for both humans and computers to judge a class label, we cast the text classification task as a binary classification task, where 1- and 2-star ratings are converted to “negative”, and 4 and 5 are converted to “positive”. We do not use reviews with a 3-star rating.
One of the problems with MARC is that it sometimes contains data where the rating diverges from the review text. This happens, for example, when a review with positive content is given a rating of 1 or 2. These data degrade the quality of our dataset. To improve the quality of the dev/test instances used for evaluation, we crowdsource a positive/negative judgment task for approximately 12,000 reviews. We adopt only reviews with the same votes from 7 or more out of 10 workers and assign a label of the maximum votes to these reviews. We divide the resulting reviews into dev/test data.
We obtained 5,654 and 5,639 instances for the dev and test data, respectively, through the above procedure. For the training data, we extracted 187,528 instances directly from MARC without performing the cleaning procedure because of the large number of training instances. The statistics of MARC-ja are listed in Table 2. For the evaluation metric for MARC-ja, we use accuracy because it is a binary classification task of texts.
JCoLA
From JCoLA's paper:
3 JCoLA
In this study, we introduce JCoLA (Japanese Corpus of Linguistic Acceptability), which will be the first large-scale acceptability judgment task dataset focusing on Japanese. JCoLA consists of sentences from textbooks and handbooks on Japanese syntax, as well as from journal articles on Japanese syntax that are published in JEAL (Journal of East Asian Linguistics), one of the prestigious journals in theoretical linguistics.
3.1 Data Collection
Sentences in JCoLA were collected from prominent textbooks and handbooks focusing on Japanese syntax. In addition to the main text, example sentences included in the footnotes were also considered for collection. We also collected acceptability judgments from journal articles on Japanese syntax published in JEAL (Journal of East Asian Linguistics): one of the prestigious journals in the-oretical linguistics. Specifically, we examined all the articles published in JEAL between 2006 and 2015 (133 papers in total), and extracted 2,252 acceptability judgments from 26 papers on Japanese syntax (Table 2). Acceptability judgments include sentences in appendices and footnotes, but not sentences presented for analyses of syntactic structures (e.g. sentences with brackets to show their syntactic structures). As a result, a total of 11,984 example. sentences were collected. Using this as a basis, JCoLA was constructed through the methodology explained in the following sections.
JSTS and JNLI
From JGLUE's paper:
For the sentence pair classification datasets, we construct a semantic textual similarity (STS) dataset, JSTS, and a natural language inference (NLI) dataset, JNLI.
Overview
STS is a task of estimating the semantic similarity of a sentence pair. Gold similarity is usually assigned as an average of the integer values 0 (completely different meaning) to 5 (equivalent meaning) assigned by multiple workers through crowdsourcing.
NLI is a task of recognizing the inference relation that a premise sentence has to a hypothesis sentence. Inference relations are generally defined by three labels: “entailment”, “contradiction”, and “neutral”. Gold inference relations are often assigned by majority voting after collecting answers from multiple workers through crowdsourcing.
For the STS and NLI tasks, STS-B (Cer et al., 2017) and MultiNLI (Williams et al., 2018) are included in GLUE, respectively. As Japanese datasets, JSNLI (Yoshikoshi et al., 2020) is a machine translated dataset of the NLI dataset SNLI (Stanford NLI), and JSICK (Yanaka and Mineshima, 2021) is a human translated dataset of the STS/NLI dataset SICK (Marelli et al., 2014). As mentioned in Section 1, these have problems originating from automatic/manual translations. To solve this problem, we construct STS/NLI datasets in Japanese from scratch. We basically extract sentence pairs in JSTS and JNLI from the Japanese version of the MS COCO Caption Dataset (Chen et al., 2015), the YJ Captions Dataset (Miyazaki and Shimizu, 2016). Most of the sentence pairs in JSTS and JNLI overlap, allowing us to analyze the relationship between similarities and inference relations for the same sentence pairs like SICK and JSICK.
The similarity value in JSTS is assigned a real number from 0 to 5 as in STS-B. The inference relation in JNLI is assigned from the above three labels as in SNLI and MultiNLI. The definitions of the inference relations are also based on SNLI.
Method of Construction
Our construction flow for JSTS and JNLI is shown in Figure 1. Basically, two captions for the same image of YJ Captions are used as sentence pairs. For these sentence pairs, similarities and NLI relations of entailment and neutral are obtained by crowdsourcing. However, it is difficult to collect sentence pairs with low similarity and contradiction relations from captions for the same image. To solve this problem, we collect sentence pairs with low similarity from captions for different images. We collect contradiction relations by asking workers to write contradictory sentences for a given caption.
The detailed construction procedure for JSTS and JNLI is described below.
- We crowdsource an STS task using two captions for the same image from YJ Captions. We ask five workers to answer the similarity between two captions and take the mean value as the gold similarity. We delete sentence pairs with a large variance in the answers because such pairs have poor answer quality. We performed this task on 16,000 sentence pairs and deleted sentence pairs with a similarity variance of 1.0 or higher, resulting in the collection of 10,236 sentence pairs with gold similarity. We refer to this collected data as JSTS-A.
- To collect sentence pairs with low similarity, we crowdsource the same STS task as Step 1 using sentence pairs of captions for different images. We conducted this task on 4,000 sentence pairs and collected 2,970 sentence pairs with gold similarity. We refer to this collected data as JSTS-B.
- For JSTS-A, we crowdsource an NLI task. Since inference relations are directional, we obtain inference relations in both directions for sentence pairs. As mentioned earlier,it is difficult to collect instances of contradiction from JSTS-A, which was collected from the captions of the same images,and thus we collect instances of entailment and neutral in this step. We collect inference relation answers from 10 workers. If six or more people give the same answer, we adopt it as the gold label if it is entailment or neutral. To obtain inference relations in both directions for JSTS-A, we performed this task on 20,472 sentence pairs, twice as many as JSTS-A. As a result, we collected inference relations for 17,501 sentence pairs. We refer to this collected data as JNLI-A. We do not use JSTS-B for the NLI task because it is difficult to define and determine the inference relations between captions of different images.
- To collect NLI instances of contradiction, we crowdsource a task of writing four contradictory sentences for each caption in YJCaptions. From the written sentences, we remove sentence pairs with an edit distance of 0.75 or higher to remove low-quality sentences, such as short sentences and sentences with low relevance to the original sentence. Furthermore, we perform a one-way NLI task with 10 workers to verify whether the created sentence pairs are contradictory. Only the sentence pairs answered as contradiction by at least six workers are adopted. Finally,since the contradiction relation has no direction, we automatically assign contradiction in the opposite direction of the adopted sentence pairs. Using 1,800 captions, we acquired 7,200 sentence pairs, from which we collected 3,779 sentence pairs to which we assigned the one-way contradiction relation.By automatically assigning the contradiction relation in the opposite direction, we doubled the number of instances to 7,558. We refer to this collected data as JNLI-C.
- For the 3,779 sentence pairs collected in Step 4, we crowdsource an STS task, assigning similarity and filtering in the same way as in Steps1 and 2. In this way, we collected 2,303 sentence pairs with gold similarity from 3,779 pairs. We refer to this collected data as JSTS-C.
JSQuAD
From JGLUE's paper:
As QA datasets, we build a Japanese version of SQuAD (Rajpurkar et al., 2016), one of the datasets of reading comprehension, and a Japanese version ofCommonsenseQA, which is explained in the next section.
Reading comprehension is the task of reading a document and answering questions about it. Many reading comprehension evaluation sets have been built in English, followed by those in other languages or multilingual ones.
In Japanese, reading comprehension datasets for quizzes (Suzukietal.,2018) and those in the drivingdomain (Takahashi et al., 2019) have been built, but none are in the general domain. We use Wikipedia to build a dataset for the general domain. The construction process is basically based on SQuAD 1.1 (Rajpurkar et al., 2016).
First, to extract high-quality articles from Wikipedia, we use Nayuki, which estimates the quality of articles on the basis of hyperlinks in Wikipedia. We randomly chose 822 articles from the top-ranked 10,000 articles. For example, the articles include “熊本県 (Kumamoto Prefecture)” and “フランス料理 (French cuisine)”. Next, we divide an article into paragraphs, present each paragraph to crowdworkers, and ask them to write questions and answers that can be answered if one understands the paragraph. Figure 2 shows an example of JSQuAD. We ask workers to write two additional answers for the dev and test sets to make the system evaluation robust.
JCommonsenseQA
From JGLUE's paper:
Overview
JCommonsenseQA is a Japanese version of CommonsenseQA (Talmor et al., 2019), which consists of five choice QA to evaluate commonsense reasoning ability. Figure 3 shows examples of JCommonsenseQA. In the same way as CommonsenseQA, JCommonsenseQA is built using crowdsourcing with seeds extracted from the knowledge base ConceptNet (Speer et al., 2017). ConceptNet is a multilingual knowledge base that consists of triplets of two concepts and their relation. The triplets are directional and represented as (source concept, relation, target concept), for example (bullet train, AtLocation, station).
Method of Construction
The construction flow for JCommonsenseQA is shown in Figure 4. First, we collect question sets (QSs) from ConceptNet, each of which consists of a source concept and three target concepts that have the same relation to the source concept. Next, for each QS, we crowdAtLocation 2961source a task of writing a question with only one target concept as the answer and a task of adding two distractors. We describe the detailed construction procedure for JCommonsenseQA below, showing how it differs from CommonsenseQA.
- We collect Japanese QSs from ConceptNet. CommonsenseQA uses only forward relations (source concept, relation, target concept) excluding general ones such as “RelatedTo” and “IsA”. JCommonsenseQA similarly uses a set of 22 relations5, excluding general ones, but the direction of the relations is bidirectional to make the questions more diverse. In other words, we also use relations in the opposite direction (source concept, relation−1, target concept).6 With this setup, we extracted 43,566 QSs with Japanese source/target concepts and randomly selected 7,500 from them.
- Some low-quality questions in CommonsenseQA contain distractors that can be considered to be an answer. To improve the quality of distractors, we add the following two processes that are not adopted in CommonsenseQA. First, if three target concepts of a QS include a spelling variation or a synonym of one another, this QS is removed. To identify spelling variations, we use the word ID of the morphological dictionary Juman Dic7. Second, we crowdsource a task of judging whether target concepts contain a synonym. As a result, we adopted 5,920 QSs from 7,500.
- For each QS, we crowdsource a task of writing a question sentence in which only one from the three target concepts is an answer. In the example shown in Figure 4, “駅 (station)” is an answer, and the others are distractors. To remove low quality question sentences, we remove the following question sentences.
- Question sentences that contain a choice word(this is because such a question is easily solved).
- Question sentences that contain the expression “XX characters”.8 (XX is a number).
- Improperly formatted question sentences that do not end with “?”.
- As a result, 5,920 × 3 = 17,760question sentences were created, from which we adopted 15,310 by removing inappropriate question sentences.
- In CommonsenseQA, when adding distractors, one is selected from ConceptNet, and the other is created by crowdsourcing. In JCommonsenseQA, to have a wider variety of distractors, two distractors are created by crowdsourcing instead of selecting from ConceptNet. To improve the quality of the questions9, we remove questions whose added distractors fall into one of the following categories:
- Distractors are included in a question sentence.
- Distractors overlap with one of existing choices.
- As a result, distractors were added to the 15,310 questions, of which we adopted 13,906.
- We asked three crowdworkers to answer each question and adopt only those answered correctly by at least two workers. As a result, we adopted 11,263 out of the 13,906 questions.
Who are the annotators?
From JGLUE's README.md:
We use Yahoo! Crowdsourcing for all crowdsourcing tasks in constructing the datasets.
From JCoLA's paper:
As a reference for the upper limit of accuracy in JCoLA, human acceptability judgment experiments were conducted on Lancers2 with a subset of the JCoLA data.
Personal and Sensitive Information
[More Information Needed]
Considerations for Using the Data
Social Impact of Dataset
From JGLUE's paper:
We build a Japanese NLU benchmark, JGLUE, from scratch without translation to measure the general NLU ability in Japanese. We hope that JGLUE will facilitate NLU research in Japanese.
Discussion of Biases
[More Information Needed]
Other Known Limitations
From JCoLA's paper:
All the sentences included in JCoLA have been extracted from textbooks, handbooks and journal articles on theoretical syntax. Therefore, those sentences are guaranteed to be theoretically meaningful, making JCoLA a challenging dataset. However, the distribution of linguistic phenomena directly reflects that of the source literature and thus turns out to be extremely skewed. Indeed, as can be seen in Table 3, while the number of sentences exceeds 100 for most linguistic phenomena, there are several linguistic phenomena for which there are only about 10 sentences. In addition, since it is difficult to force language models to interpret sentences given specific contexts, those sentences whose unacceptability depends on contexts were inevitably removed from JCoLA. This removal process resulted in the deletion of unacceptable sentences from some linguistic phenomena (such as ellipsis), consequently skewing the balance between acceptable and unacceptable sentences (with a higher proportion of acceptable sentences).
Additional Information
- 日本語言語理解ベンチマーク JGLUE の構築 〜 自然言語処理モデルの評価用データセットを公開しました - Yahoo! JAPAN Tech Blog https://techblog.yahoo.co.jp/entry/2022122030379907/
Dataset Curators
MARC-ja
- Keung, Phillip, et al. "The Multilingual Amazon Reviews Corpus." Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2020.
JCoLA
- Someya, Sugimoto, and Oseki. "JCoLA: Japanese Corpus of Linguistic Acceptability." arxiv preprint arXiv:2309.12676 (2023).
JSTS and JNLI
- Miyazaki, Takashi, and Nobuyuki Shimizu. "Cross-lingual image caption generation." Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2016.
JSQuAD
The JGLUE's 'authors curated the original data for JSQuAD from the Japanese wikipedia dump.
JCommonsenseQA
In the same way as CommonsenseQA, JCommonsenseQA is built using crowdsourcing with seeds extracted from the knowledge base ConceptNet
Licensing Information
JGLUE
From JGLUE's README.md':
This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
JCoLA
From JCoLA's README.md':
The text in this corpus is excerpted from the published works, and copyright (where applicable) remains with the original authors or publishers. We expect that research use within Japan is legal under fair use, but make no guarantee of this.
Citation Information
JGLUE
@inproceedings{kurihara-lrec-2022-jglue,
title={JGLUE: Japanese general language understanding evaluation},
author={Kurihara, Kentaro and Kawahara, Daisuke and Shibata, Tomohide},
booktitle={Proceedings of the Thirteenth Language Resources and Evaluation Conference},
pages={2957--2966},
year={2022},
url={https://aclanthology.org/2022.lrec-1.317/}
}
@inproceedings{kurihara-nlp-2022-jglue,
title={JGLUE: 日本語言語理解ベンチマーク},
author={栗原健太郎 and 河原大輔 and 柴田知秀},
booktitle={言語処理学会第 28 回年次大会},
pages={2023--2028},
year={2022},
url={https://www.anlp.jp/proceedings/annual_meeting/2022/pdf_dir/E8-4.pdf},
note={in Japanese}
}
MARC-ja
@inproceedings{marc_reviews,
title={The Multilingual Amazon Reviews Corpus},
author={Keung, Phillip and Lu, Yichao and Szarvas, György and Smith, Noah A.},
booktitle={Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing},
year={2020}
}
JCoLA
@article{someya-arxiv-2023-jcola,
title={JCoLA: Japanese Corpus of Linguistic Acceptability},
author={Taiga Someya and Yushi Sugimoto and Yohei Oseki},
year={2023},
eprint={2309.12676},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
@inproceedings{someya-nlp-2022-jcola,
title={日本語版 CoLA の構築},
author={染谷 大河 and 大関 洋平},
booktitle={言語処理学会第 28 回年次大会},
pages={1872--1877},
year={2022},
url={https://www.anlp.jp/proceedings/annual_meeting/2022/pdf_dir/E7-1.pdf},
note={in Japanese}
}
JSTS and JNLI
@inproceedings{miyazaki2016cross,
title={Cross-lingual image caption generation},
author={Miyazaki, Takashi and Shimizu, Nobuyuki},
booktitle={Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
pages={1780--1790},
year={2016}
}
Contributions
Thanks to Kentaro Kurihara, Daisuke Kawahara, and Tomohide Shibata for creating JGLUE dataset. Thanks to Taiga Someya for creating JCoLA dataset.
- Downloads last month
- 66