Dataset Viewer

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.

webis/Touche23-ValueEval(複製)

データの内容

議論の背後にある人間の価値を、0/1 のラベルで示したデータ。1 行が 1 件の議論で、Conclusion(結論)・Stance(結論への賛否)・Premise(前提)・価値に訴えているかを 0/1 で示す Labels の配列を持つ。IBM-ArgQ-Rank-30kArgs や Zhihu、The New York Times など 6 つの情報源から集め、ラベルは Kiesel らの価値の分類体系に基づくとカードにある。The New York Times 由来の部分は著作権のため直接配布されず、専用のプログラムで取得する。

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 46.97 KB(ファイル 3 本)
ファイルの形式 py 1
言語 en
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

webis/Touche23-ValueEval をリビジョン 66ed7d7e0ca7778b8662bd31af55365c07d6c1cd 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

元のカードにアノテーションの節はあるが、中身は空欄(More Information Needed)だった。

微調整に使うときの注意

  • 調査メモの台帳での区分: A(許諾型のライセンスだけで、疑義の注記が無い)。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
behavior_values_management.md Q2. 価値観・道徳・規範のラベル付きテキストには何があり、Laya の問いにどう落とせるか (L59) cc-by-4.0 確認済 原文注記: HF 版はスクリプト型で、datasets-server は対応しない。nyt は著作権のため直接配布が無く、専用ダウンロード用のプログラムで作る(カード)。議論文(賛否の前提)で、経営者の発言ではない
reasoning_argument_numeric.md Q1. 論証(主張・根拠・立場・説得力・論証の質)に付いたラベルのデータは何があるか (L22) cc-by-4.0 確認済 原文注記: script。出所に Zhihu・NYT・欧州の市民討論が混じる
jev_sources_tasksource_bekko.csv CSV(tasksource-jev-typed-decisions) cc-by-4.0, CC BY 4.0 (DPI) / license_use=commercial 確認済 license_use=commercial / [Jev 元データ source] source_annotation_converted(deterministic recast; no model r

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。


The Touché23-ValueEval データセット

目次

  • 目次
  • データセットの説明
    • データセットの概要
    • データセットの使い方
    • 対象タスクとリーダーボード
    • 言語
  • データセットの構造
    • 議論のインスタンス
    • メタデータのインスタンス
  • 補足情報
    • データセットの作成者
    • ライセンス情報
    • 引用情報

データセットの説明

データセットの概要

Touché23-ValueEval データセットには、6 つの異なる情報源から集めた 9324 件の議論が含まれる。議論の情報源は、その Argument ID の先頭の文字で示される:

アノテーションされたラベルは、ACL'22 で発表された Identifying the Human Values behind Arguments(Kiesel et al. 2022)で示された価値の分類体系に基づいている。

[1] https://language.ml [2] https://en.wikipedia.org/wiki/Nahj_al-Balagha [3] https://en.wikipedia.org/wiki/Ghurar_al-Hikam_wa_Durar_al-Kalim

データセットの使い方

既定の設定名は main である。

from datasets import load_dataset
dataset = load_dataset("webis/Touche23-ValueEval")
print(dataset['train'].info.description)
for argument in iter(dataset['train']):
    print(f"{argument['Argument ID']}: {argument['Stance']} '{argument['Conclusion']}': {argument['Premise']}")

対象タスクとリーダーボード

人間の価値の検出

言語

議論のインスタンス はすべて単一言語であり、英語(主に en-US)の文書のみを含む。 一部のデータ部分の メタデータのインスタンス には、議論を元の言語と表現で記したものも追加されている。

データセットの構造

議論のインスタンス

各議論のインスタンスは次の属性を持つ:

  • Argument ID: データセット内で議論を一意に識別する ID
  • Conclusion: 議論の結論の文
  • Stance: Premise の Conclusion; one of "in favor of", "against" に対する立場
  • Premise: 議論の前提の文
  • Labels: 各例の Labels は、1(議論が価値に訴えている)と 0(議論が価値に訴えていない)からなる配列である。並び順は元のファイルと同じである。

加えて、ラベルは value-categories(価値の分類体系の第 2 階層のラベル。Kiesel et al. 2022b)と、human values(同じく第 1 階層のラベル)に分けられる。 この区別は設定名にも反映されている:

  • <config>: Task は主に value-categories の検出に焦点を当てているため、各基本設定(下に挙げる)は、20 個の value-categories をラベルとして持つ:
    labels = ["Self-direction: thought", "Self-direction: action", "Stimulation", "Hedonism", "Achievement", "Power: dominance", "Power: resources", "Face", "Security: personal", "Security: societal", "Tradition", "Conformity: rules", "Conformity: interpersonal", "Humility", "Benevolence: caring", "Benevolence: dependability", "Universalism: concern", "Universalism: nature", "Universalism: tolerance", "Universalism: objectivity"]
    
  • <config>-level1: 価値の分類体系の第 1 階層にある 54 個の human values は、2023 年のタスクでは使わない(アノテーションには使われている)が、value-categories を理解する助けになると考える人もいるため、ここにも挙げておく。その並び順も元のファイルと同じである。詳細は value-categories の設定を参照のこと。

このデータセットの設定名(<config> の置き換え)は次のとおり:

  • main: 8865 件の議論(情報源: A、D、E)。分割は train、validation、test(既定の設定名)
    dataset_main_train = load_dataset("webis/Touche23-ValueEval", split="train")
    dataset_main_validation = load_dataset("webis/Touche23-ValueEval", split="validation")
    dataset_main_test = load_dataset("webis/Touche23-ValueEval", split="test")
    
  • nahjalbalagha: 279 件の議論(情報源: F)。分割は test
    dataset_nahjalbalagha_test = load_dataset("webis/Touche23-ValueEval", name="nahjalbalagha", split="test")
    
  • nyt: 80 件の議論(情報源: G)。分割は test
    dataset_nyt_test = load_dataset("webis/Touche23-ValueEval", name="nyt", split="test")
    
  • zhihu: 100 件の議論(情報源: C)。分割は validation
    dataset_zhihu_validation = load_dataset("webis/Touche23-ValueEval", name="zhihu", split="validation")
    

著作権上の理由から、現在は New york Times データセットに含まれる議論への直接のダウンロードリンクは存在しない。そのため、nyt または nyt-level1 の設定にアクセスすると、専用に作られた nyt-downloader program を使って議論をローカルに作成し、アクセスする。 詳細はプログラムの README を参照のこと。

メタデータのインスタンス

以下はメタデータのすべての設定名である。各設定には meta という分割が 1 つだけある。

  • ibm-meta: 各行は IBM-ArgQ-Rank-30kArgs の 1 件の議論(ID が A で始まるもの)に対応する
    • Argument ID: 議論を一意に識別する ID
    • WA: 加重平均のスコアリング関数による品質ラベル
    • MACE-P: MACE-P のスコアリング関数による品質ラベル
    • stance_WA: 加重平均のスコアリング関数による立場のラベル
    • stance_WA_conf: 加重平均のスコアリング関数による立場のラベルの信頼度
    dataset_ibm_metadata = load_dataset("webis/Touche23-ValueEval", name="ibm-meta", split="meta")
    
  • zhihu-meta: 各行は、中国語の質問応答サイト Zhihu の 1 件の議論(ID が C で始まるもの)に対応する
    • Argument ID: 議論を一意に識別する ID
    • Conclusion Chinese: 元の中国語の結論の文
    • Premise Chinese: 元の中国語の前提の文
    • URL: 議論の元になった発言へのリンク
    dataset_zhihu_metadata = load_dataset("webis/Touche23-ValueEval", name="zhihu-meta", split="meta")
    
  • gdi-meta: 各行は GD IDEAS の 1 件の議論(ID が D で始まるもの)に対応する
    • Argument ID: 議論を一意に識別する ID
    • URL: 議論の元になったトピックへのリンク
    dataset_gdi_metadata = load_dataset("webis/Touche23-ValueEval", name="gdi-meta", split="meta")
    
  • cofe-meta: 各行は the Conference for the Future of Europe の 1 件の議論(ID が E で始まるもの)に対応する
    • Argument ID: 議論を一意に識別する ID
    • URL: 議論の元になったコメントへのリンク
    dataset_cofe_metadata = load_dataset("webis/Touche23-ValueEval", name="cofe-meta", split="meta")
    
  • nahjalbalagha-meta: 各行は 1 件の議論(ID が F で始まるもの)に対応する。このファイルには、nahjalbalagha(または nahjalbalagha-level1)の 279 件の議論に関する情報と、これまでラベルが付けられていない 1047 件の追加の議論が含まれる。このデータは language.ml ラボが寄稿したものである。
    • Argument ID: 議論を一意に識別する ID
    • Conclusion Farsi: 議論の結論の文(ペルシア語)
    • Stance Farsi: Premise の Conclusion に対する立場(ペルシア語)
    • Premise Farsi: 議論の前提の文(ペルシア語)
    • Conclusion English: 議論の結論の文(英語。ペルシア語からの翻訳)
    • Stance English: Premise の Conclusion に対する立場。"in favor of" または "against" のいずれか
    • Premise English: 議論の前提の文(英語。ペルシア語からの翻訳)
    • Source: 議論の出典のテキスト。"Nahj al-Balagha" または "Ghurar al-Hikam wa Durar ak-Kalim" のいずれか。それぞれのペルシア語訳を使用した
    • Method: 前提を出典からどのように抽出したか。"extracted"(直接取られたもの)または "deduced"(推論されたもの)のいずれか。結論は推論されたものである
    dataset_nahjalbalagha_metadata = load_dataset("webis/Touche23-ValueEval", name="nahjalbalagha-meta", split="meta")
    
  • nyt-meta: 各行は The New York Times の 1 件の議論(ID が G で始まるもの)に対応する
    • Argument ID: 議論を一意に識別する ID
    • URL: 議論の元になった記事へのリンク
    • Internet Archive timestamp: 使用した、Internet Archive 上の記事の版のタイムスタンプ
    dataset_nyt_metadata = load_dataset("webis/Touche23-ValueEval", name="nyt-meta", split="meta")
    
  • value-categories: 価値の分類体系に関する第 2 階層と第 1 階層の値の構造を持つ、単一の JSON エントリを含む:
    {
      "<value category>": {
        "<level 1 value>": [
          "<exemplary effect a corresponding argument might target>",
          ...
        ], ...
      }, ...
    }
    
    この設定は単一のエントリしか含まないため、使用例は次のとおり:
    value_categories = load_dataset("webis/Touche23-ValueEval", name="value-categories", split="meta")[0]
    

補足情報

データセットの作成者

[More Information Needed]

ライセンス情報

Creative Commons Attribution 4.0 International (CC BY 4.0)

引用情報

@Article{mirzakhmedova:2023a,
  author    = {Nailia Mirzakhmedova and Johannes Kiesel and Milad Alshomary and Maximilian Heinrich and Nicolas Handke\
and Xiaoni Cai and Valentin Barriere and Doratossadat Dastgheib and Omid Ghahroodi and {Mohammad Ali} Sadraei\
and Ehsaneddin Asgari and Lea Kawaletz and Henning Wachsmuth and Benno Stein},
  doi       = {10.48550/arXiv.2301.13771},
  journal   = {CoRR},
  month     = jan,
  publisher = {arXiv},
  title     = {{The Touch{\'e}23-ValueEval Dataset for Identifying Human Values behind Arguments}},
  volume    = {abs/2301.13771},
  year      = 2023
}
原文(英語)

The Touché23-ValueEval Dataset

Dataset Summary

The Touché23-ValueEval Dataset comprises 9324 arguments from six different sources. An arguments source is indicated with the first letter of its Argument ID:

The annotated labels are based on the value taxonomy published in Identifying the Human Values behind Arguments (Kiesel et al. 2022) at ACL'22.

[1] https://language.ml [2] https://en.wikipedia.org/wiki/Nahj_al-Balagha [3] https://en.wikipedia.org/wiki/Ghurar_al-Hikam_wa_Durar_al-Kalim

Dataset Usage

The default configuration name is main.

from datasets import load_dataset
dataset = load_dataset("webis/Touche23-ValueEval")
print(dataset['train'].info.description)
for argument in iter(dataset['train']):
    print(f"{argument['Argument ID']}: {argument['Stance']} '{argument['Conclusion']}': {argument['Premise']}")

Supported Tasks and Leaderboards

Human Value Detection

Languages

The Argument Instances are all monolingual; it only includes English (mostly en-US) documents. The Metadata Instances for some dataset parts additionally state the arguments in their original language and phrasing.

Dataset Structure

Argument Instances

Each argument instance has the following attributes:

  • Argument ID: The unique identifier for the argument within the dataset
  • Conclusion: Conclusion text of the argument
  • Stance: Stance of the Premise towards the `Conclusion; one of "in favor of", "against"
  • Premise: Premise text of the argument
  • Labels: The Labels for each example is an array of 1s (argument resorts to value) and 0s (argument does not resort to value). The order is the same as in the original files.

Additionally, the labels are separated into value-categories, aka. level 2 labels of the value taxonomy (Kiesel et al. 2022b), and human values, aka. level 1 labels of the value taxonomy. This distinction is also reflected in the configuration names:

  • <config>: As the Task is focused mainly on the detection of value-categories, each base configuration (listed below) has the 20 value-categories as labels:
    labels = ["Self-direction: thought", "Self-direction: action", "Stimulation", "Hedonism", "Achievement", "Power: dominance", "Power: resources", "Face", "Security: personal", "Security: societal", "Tradition", "Conformity: rules", "Conformity: interpersonal", "Humility", "Benevolence: caring", "Benevolence: dependability", "Universalism: concern", "Universalism: nature", "Universalism: tolerance", "Universalism: objectivity"]
    
  • <config>-level1: The 54 human values from the level 1 of the value taxonomy are not used for the 2023 task (except for the annotation), but are still listed here for some might find them useful for understanding the value categories. Their order is also the same as in the original files. For more details see the value-categories configuration.

The configuration names (as replacements for <config>) in this dataset are:

  • main: 8865 arguments (sources: A, D, E) with splits train, validation, and test (default configuration name)
    dataset_main_train = load_dataset("webis/Touche23-ValueEval", split="train")
    dataset_main_validation = load_dataset("webis/Touche23-ValueEval", split="validation")
    dataset_main_test = load_dataset("webis/Touche23-ValueEval", split="test")
    
  • nahjalbalagha: 279 arguments (source: F) with split test
    dataset_nahjalbalagha_test = load_dataset("webis/Touche23-ValueEval", name="nahjalbalagha", split="test")
    
  • nyt: 80 arguments (source: G) with split test
    dataset_nyt_test = load_dataset("webis/Touche23-ValueEval", name="nyt", split="test")
    
  • zhihu: 100 arguments (source: C) with split validation
    dataset_zhihu_validation = load_dataset("webis/Touche23-ValueEval", name="zhihu", split="validation")
    

Please note that due to copyright reasons, there currently does not exist a direct download link to the arguments contained in the New york Times dataset. Accessing any of the nyt or nyt-level1 configurations will therefore use the specifically created nyt-downloader program to create and access the arguments locally. See the program's README for further details.

Metadata Instances

The following lists all configuration names for metadata. Each configuration only has a single split named meta.

  • ibm-meta: Each row corresponds to one argument (IDs starting with A) from the IBM-ArgQ-Rank-30kArgs
    • Argument ID: The unique identifier for the argument
    • WA: the quality label according to the weighted-average scoring function
    • MACE-P: the quality label according to the MACE-P scoring function
    • stance_WA: the stance label according to the weighted-average scoring function
    • stance_WA_conf: the confidence in the stance label according to the weighted-average scoring function
    dataset_ibm_metadata = load_dataset("webis/Touche23-ValueEval", name="ibm-meta", split="meta")
    
  • zhihu-meta: Each row corresponds to one argument (IDs starting with C) from the Chinese question-answering website Zhihu
    • Argument ID: The unique identifier for the argument
    • Conclusion Chinese: The original chinese conclusion statement
    • Premise Chinese: The original chinese premise statement
    • URL: Link to the original statement the argument was taken from
    dataset_zhihu_metadata = load_dataset("webis/Touche23-ValueEval", name="zhihu-meta", split="meta")
    
  • gdi-meta: Each row corresponds to one argument (IDs starting with D) from GD IDEAS
    • Argument ID: The unique identifier for the argument
    • URL: Link to the topic the argument was taken from
    dataset_gdi_metadata = load_dataset("webis/Touche23-ValueEval", name="gdi-meta", split="meta")
    
  • cofe-meta: Each row corresponds to one argument (IDs starting with E) from the Conference for the Future of Europe
    • Argument ID: The unique identifier for the argument
    • URL: Link to the comment the argument was taken from
    dataset_cofe_metadata = load_dataset("webis/Touche23-ValueEval", name="cofe-meta", split="meta")
    
  • nahjalbalagha-meta: Each row corresponds to one argument (IDs starting with F). This file contains information on the 279 arguments in nahjalbalagha (or nahjalbalagha-level1) and 1047 additional arguments that were not labeled so far. This data was contributed by the language.ml lab.
    • Argument ID: The unique identifier for the argument
    • Conclusion Farsi: Conclusion text of the argument in Farsi
    • Stance Farsi: Stance of the Premise towards the Conclusion, in Farsi
    • Premise Farsi: Premise text of the argument in Farsi
    • Conclusion English: Conclusion text of the argument in English (translated from Farsi)
    • Stance English: Stance of the Premise towards the Conclusion; one of "in favor of", "against"
    • Premise English: Premise text of the argument in English (translated from Farsi)
    • Source: Source text of the argument; one of "Nahj al-Balagha", "Ghurar al-Hikam wa Durar ak-Kalim"; their Farsi translations were used
    • Method: How the premise was extracted from the source; one of "extracted" (directly taken), "deduced"; the conclusion are deduced
    dataset_nahjalbalagha_metadata = load_dataset("webis/Touche23-ValueEval", name="nahjalbalagha-meta", split="meta")
    
  • nyt-meta: Each row corresponds to one argument (IDs starting with G) from The New York Times
    • Argument ID: The unique identifier for the argument
    • URL: Link to the article the argument was taken from
    • Internet Archive timestamp: Timestamp of the article's version in the Internet Archive that was used
    dataset_nyt_metadata = load_dataset("webis/Touche23-ValueEval", name="nyt-meta", split="meta")
    
  • value-categories: Contains a single JSON-entry with the structure of level 2 and level 1 values regarding the value taxonomy:
    {
      "<value category>": {
        "<level 1 value>": [
          "<exemplary effect a corresponding argument might target>",
          ...
        ], ...
      }, ...
    }
    
    As this configuration contains just a single entry, an example usage could be:
    value_categories = load_dataset("webis/Touche23-ValueEval", name="value-categories", split="meta")[0]
    

Additional Information

Dataset Curators

[More Information Needed]

Licensing Information

Creative Commons Attribution 4.0 International (CC BY 4.0)

Citation Information

@Article{mirzakhmedova:2023a,
  author    = {Nailia Mirzakhmedova and Johannes Kiesel and Milad Alshomary and Maximilian Heinrich and Nicolas Handke\
and Xiaoni Cai and Valentin Barriere and Doratossadat Dastgheib and Omid Ghahroodi and {Mohammad Ali} Sadraei\
and Ehsaneddin Asgari and Lea Kawaletz and Henning Wachsmuth and Benno Stein},
  doi       = {10.48550/arXiv.2301.13771},
  journal   = {CoRR},
  month     = jan,
  publisher = {arXiv},
  title     = {{The Touch{\'e}23-ValueEval Dataset for Identifying Human Values behind Arguments}},
  volume    = {abs/2301.13771},
  year      = 2023
}
Downloads last month
98