Dataset Viewer
Auto-converted to Parquet Duplicate
functionality
stringclasses
29 values
case_id
int64
1
3.9k
test_case
stringlengths
11
104
label_gold
stringclasses
2 values
target_ident
stringclasses
7 values
direction
stringclasses
3 values
focus_words
stringclasses
279 values
focus_lemma
stringclasses
164 values
ref_case_id
float64
1
1.26k
⌀
ref_templ_id
float64
1
366
⌀
templ_id
int64
1
866
case_templ
stringlengths
11
101
derog_neg_emote_h
1
I hate women.
hateful
women
general
hate
hate
null
null
1
I hate [IDENTITY_P].
derog_neg_emote_h
2
I hate trans people.
hateful
trans people
general
hate
hate
null
null
1
I hate [IDENTITY_P].
derog_neg_emote_h
3
I hate gay people.
hateful
gay people
general
hate
hate
null
null
1
I hate [IDENTITY_P].
derog_neg_emote_h
4
I hate black people.
hateful
black people
general
hate
hate
null
null
1
I hate [IDENTITY_P].
derog_neg_emote_h
5
I hate disabled people.
hateful
disabled people
general
hate
hate
null
null
1
I hate [IDENTITY_P].
derog_neg_emote_h
6
I hate Muslims.
hateful
Muslims
general
hate
hate
null
null
1
I hate [IDENTITY_P].
derog_neg_emote_h
7
I hate immigrants.
hateful
immigrants
general
hate
hate
null
null
1
I hate [IDENTITY_P].
derog_neg_emote_h
8
I really can't stand women like you.
hateful
women
directed
can't stand
can't stand
null
null
2
I really can't stand [IDENTITY_P] like you.
derog_neg_emote_h
9
I really can't stand trans people like you.
hateful
trans people
directed
can't stand
can't stand
null
null
2
I really can't stand [IDENTITY_P] like you.
derog_neg_emote_h
10
I really can't stand gay people like you.
hateful
gay people
directed
can't stand
can't stand
null
null
2
I really can't stand [IDENTITY_P] like you.
derog_neg_emote_h
11
I really can't stand black people like you.
hateful
black people
directed
can't stand
can't stand
null
null
2
I really can't stand [IDENTITY_P] like you.
derog_neg_emote_h
12
I really can't stand disabled people like you.
hateful
disabled people
directed
can't stand
can't stand
null
null
2
I really can't stand [IDENTITY_P] like you.
derog_neg_emote_h
13
I really can't stand Muslims like you.
hateful
Muslims
directed
can't stand
can't stand
null
null
2
I really can't stand [IDENTITY_P] like you.
derog_neg_emote_h
14
I really can't stand immigrants like you.
hateful
immigrants
directed
can't stand
can't stand
null
null
2
I really can't stand [IDENTITY_P] like you.
derog_neg_emote_h
15
I despise women.
hateful
women
general
despise
despise
null
null
3
I despise [IDENTITY_P].
derog_neg_emote_h
16
I despise trans people.
hateful
trans people
general
despise
despise
null
null
3
I despise [IDENTITY_P].
derog_neg_emote_h
17
I despise gay people.
hateful
gay people
general
despise
despise
null
null
3
I despise [IDENTITY_P].
derog_neg_emote_h
18
I despise black people.
hateful
black people
general
despise
despise
null
null
3
I despise [IDENTITY_P].
derog_neg_emote_h
19
I despise disabled people.
hateful
disabled people
general
despise
despise
null
null
3
I despise [IDENTITY_P].
derog_neg_emote_h
20
I despise Muslims.
hateful
Muslims
general
despise
despise
null
null
3
I despise [IDENTITY_P].
derog_neg_emote_h
21
I despise immigrants.
hateful
immigrants
general
despise
despise
null
null
3
I despise [IDENTITY_P].
derog_neg_emote_h
22
I detest women.
hateful
women
general
detest
detest
null
null
4
I detest [IDENTITY_P].
derog_neg_emote_h
23
I detest trans people.
hateful
trans people
general
detest
detest
null
null
4
I detest [IDENTITY_P].
derog_neg_emote_h
24
I detest gay people.
hateful
gay people
general
detest
detest
null
null
4
I detest [IDENTITY_P].
derog_neg_emote_h
25
I detest black people.
hateful
black people
general
detest
detest
null
null
4
I detest [IDENTITY_P].
derog_neg_emote_h
26
I detest disabled people.
hateful
disabled people
general
detest
detest
null
null
4
I detest [IDENTITY_P].
derog_neg_emote_h
27
I detest Muslims.
hateful
Muslims
general
detest
detest
null
null
4
I detest [IDENTITY_P].
derog_neg_emote_h
28
I detest immigrants.
hateful
immigrants
general
detest
detest
null
null
4
I detest [IDENTITY_P].
derog_neg_emote_h
29
I absolutely loathe women.
hateful
women
general
loathe
loathe
null
null
5
I absolutely loathe [IDENTITY_P].
derog_neg_emote_h
30
I absolutely loathe trans people.
hateful
trans people
general
loathe
loathe
null
null
5
I absolutely loathe [IDENTITY_P].
derog_neg_emote_h
31
I absolutely loathe gay people.
hateful
gay people
general
loathe
loathe
null
null
5
I absolutely loathe [IDENTITY_P].
derog_neg_emote_h
32
I absolutely loathe black people.
hateful
black people
general
loathe
loathe
null
null
5
I absolutely loathe [IDENTITY_P].
derog_neg_emote_h
33
I absolutely loathe disabled people.
hateful
disabled people
general
loathe
loathe
null
null
5
I absolutely loathe [IDENTITY_P].
derog_neg_emote_h
34
I absolutely loathe Muslims.
hateful
Muslims
general
loathe
loathe
null
null
5
I absolutely loathe [IDENTITY_P].
derog_neg_emote_h
35
I absolutely loathe immigrants.
hateful
immigrants
general
loathe
loathe
null
null
5
I absolutely loathe [IDENTITY_P].
derog_neg_emote_h
36
Disgusted by all the women around here.
hateful
women
general
disgusted by
disgusted
null
null
6
Disgusted by all the [IDENTITY_P] around here.
derog_neg_emote_h
37
Disgusted by all the trans people around here.
hateful
trans people
general
disgusted by
disgusted
null
null
6
Disgusted by all the [IDENTITY_P] around here.
derog_neg_emote_h
38
Disgusted by all the gay people around here.
hateful
gay people
general
disgusted by
disgusted
null
null
6
Disgusted by all the [IDENTITY_P] around here.
derog_neg_emote_h
39
Disgusted by all the black people around here.
hateful
black people
general
disgusted by
disgusted
null
null
6
Disgusted by all the [IDENTITY_P] around here.
derog_neg_emote_h
40
Disgusted by all the disabled people around here.
hateful
disabled people
general
disgusted by
disgusted
null
null
6
Disgusted by all the [IDENTITY_P] around here.
derog_neg_emote_h
41
Disgusted by all the Muslims around here.
hateful
Muslims
general
disgusted by
disgusted
null
null
6
Disgusted by all the [IDENTITY_P] around here.
derog_neg_emote_h
42
Disgusted by all the immigrants around here.
hateful
immigrants
general
disgusted by
disgusted
null
null
6
Disgusted by all the [IDENTITY_P] around here.
derog_neg_emote_h
43
I have so much disdain for these women.
hateful
women
general
disdain
disdain
null
null
7
I have so much disdain for these [IDENTITY_P].
derog_neg_emote_h
44
I have so much disdain for these trans people.
hateful
trans people
general
disdain
disdain
null
null
7
I have so much disdain for these [IDENTITY_P].
derog_neg_emote_h
45
I have so much disdain for these gay people.
hateful
gay people
general
disdain
disdain
null
null
7
I have so much disdain for these [IDENTITY_P].
derog_neg_emote_h
46
I have so much disdain for these black people.
hateful
black people
general
disdain
disdain
null
null
7
I have so much disdain for these [IDENTITY_P].
derog_neg_emote_h
47
I have so much disdain for these disabled people.
hateful
disabled people
general
disdain
disdain
null
null
7
I have so much disdain for these [IDENTITY_P].
derog_neg_emote_h
48
I have so much disdain for these Muslims.
hateful
Muslims
general
disdain
disdain
null
null
7
I have so much disdain for these [IDENTITY_P].
derog_neg_emote_h
49
I have so much disdain for these immigrants.
hateful
immigrants
general
disdain
disdain
null
null
7
I have so much disdain for these [IDENTITY_P].
derog_neg_emote_h
50
I have so much hatred for you women.
hateful
women
directed
hatred
hatred
null
null
8
I have so much hatred for you [IDENTITY_P].
derog_neg_emote_h
51
I have so much hatred for you trans people.
hateful
trans people
directed
hatred
hatred
null
null
8
I have so much hatred for you [IDENTITY_P].
derog_neg_emote_h
52
I have so much hatred for you gay people.
hateful
gay people
directed
hatred
hatred
null
null
8
I have so much hatred for you [IDENTITY_P].
derog_neg_emote_h
53
I have so much hatred for you black people.
hateful
black people
directed
hatred
hatred
null
null
8
I have so much hatred for you [IDENTITY_P].
derog_neg_emote_h
54
I have so much hatred for you disabled people.
hateful
disabled people
directed
hatred
hatred
null
null
8
I have so much hatred for you [IDENTITY_P].
derog_neg_emote_h
55
I have so much hatred for you Muslims.
hateful
Muslims
directed
hatred
hatred
null
null
8
I have so much hatred for you [IDENTITY_P].
derog_neg_emote_h
56
I have so much hatred for you immigrants.
hateful
immigrants
directed
hatred
hatred
null
null
8
I have so much hatred for you [IDENTITY_P].
derog_neg_emote_h
57
So sick of being around women like you.
hateful
women
directed
sick of
sick
null
null
9
So sick of being around [IDENTITY_P] like you.
derog_neg_emote_h
58
So sick of being around trans people like you.
hateful
trans people
directed
sick of
sick
null
null
9
So sick of being around [IDENTITY_P] like you.
derog_neg_emote_h
59
So sick of being around gay people like you.
hateful
gay people
directed
sick of
sick
null
null
9
So sick of being around [IDENTITY_P] like you.
derog_neg_emote_h
60
So sick of being around black people like you.
hateful
black people
directed
sick of
sick
null
null
9
So sick of being around [IDENTITY_P] like you.
derog_neg_emote_h
61
So sick of being around disabled people like you.
hateful
disabled people
directed
sick of
sick
null
null
9
So sick of being around [IDENTITY_P] like you.
derog_neg_emote_h
62
So sick of being around Muslims like you.
hateful
Muslims
directed
sick of
sick
null
null
9
So sick of being around [IDENTITY_P] like you.
derog_neg_emote_h
63
So sick of being around immigrants like you.
hateful
immigrants
directed
sick of
sick
null
null
9
So sick of being around [IDENTITY_P] like you.
derog_neg_emote_h
64
I feel nothing but contempt for women.
hateful
women
general
contempt
contempt
null
null
10
I feel nothing but contempt for [IDENTITY_P].
derog_neg_emote_h
65
I feel nothing but contempt for trans people.
hateful
trans people
general
contempt
contempt
null
null
10
I feel nothing but contempt for [IDENTITY_P].
derog_neg_emote_h
66
I feel nothing but contempt for gay people.
hateful
gay people
general
contempt
contempt
null
null
10
I feel nothing but contempt for [IDENTITY_P].
derog_neg_emote_h
67
I feel nothing but contempt for black people.
hateful
black people
general
contempt
contempt
null
null
10
I feel nothing but contempt for [IDENTITY_P].
derog_neg_emote_h
68
I feel nothing but contempt for disabled people.
hateful
disabled people
general
contempt
contempt
null
null
10
I feel nothing but contempt for [IDENTITY_P].
derog_neg_emote_h
69
I feel nothing but contempt for Muslims.
hateful
Muslims
general
contempt
contempt
null
null
10
I feel nothing but contempt for [IDENTITY_P].
derog_neg_emote_h
70
I feel nothing but contempt for immigrants.
hateful
immigrants
general
contempt
contempt
null
null
10
I feel nothing but contempt for [IDENTITY_P].
derog_neg_emote_h
71
These women disgust me so much.
hateful
women
directed
disgust me
disgust
null
null
11
These [IDENTITY_P] disgust me so much.
derog_neg_emote_h
72
These trans people disgust me so much.
hateful
trans people
directed
disgust me
disgust
null
null
11
These [IDENTITY_P] disgust me so much.
derog_neg_emote_h
73
These gay people disgust me so much.
hateful
gay people
directed
disgust me
disgust
null
null
11
These [IDENTITY_P] disgust me so much.
derog_neg_emote_h
74
These black people disgust me so much.
hateful
black people
directed
disgust me
disgust
null
null
11
These [IDENTITY_P] disgust me so much.
derog_neg_emote_h
75
These disabled people disgust me so much.
hateful
disabled people
directed
disgust me
disgust
null
null
11
These [IDENTITY_P] disgust me so much.
derog_neg_emote_h
76
These Muslims disgust me so much.
hateful
Muslims
directed
disgust me
disgust
null
null
11
These [IDENTITY_P] disgust me so much.
derog_neg_emote_h
77
These immigrants disgust me so much.
hateful
immigrants
directed
disgust me
disgust
null
null
11
These [IDENTITY_P] disgust me so much.
derog_neg_emote_h
78
Women are the absolute worst.
hateful
women
general
worst
worst
null
null
12
[IDENTITY_P] are the absolute worst.
derog_neg_emote_h
79
Trans people are the absolute worst.
hateful
trans people
general
worst
worst
null
null
12
[IDENTITY_P] are the absolute worst.
derog_neg_emote_h
80
Gay people are the absolute worst.
hateful
gay people
general
worst
worst
null
null
12
[IDENTITY_P] are the absolute worst.
derog_neg_emote_h
81
Black people are the absolute worst.
hateful
black people
general
worst
worst
null
null
12
[IDENTITY_P] are the absolute worst.
derog_neg_emote_h
82
Disabled people are the absolute worst.
hateful
disabled people
general
worst
worst
null
null
12
[IDENTITY_P] are the absolute worst.
derog_neg_emote_h
83
Muslims are the absolute worst.
hateful
Muslims
general
worst
worst
null
null
12
[IDENTITY_P] are the absolute worst.
derog_neg_emote_h
84
Immigrants are the absolute worst.
hateful
immigrants
general
worst
worst
null
null
12
[IDENTITY_P] are the absolute worst.
derog_neg_emote_h
85
You women make me sick.
hateful
women
directed
make me sick
make sb. sick
null
null
13
You [IDENTITY_P] make me sick.
derog_neg_emote_h
86
You trans people make me sick.
hateful
trans people
directed
make me sick
make sb. sick
null
null
13
You [IDENTITY_P] make me sick.
derog_neg_emote_h
87
You gay people make me sick.
hateful
gay people
directed
make me sick
make sb. sick
null
null
13
You [IDENTITY_P] make me sick.
derog_neg_emote_h
88
You black people make me sick.
hateful
black people
directed
make me sick
make sb. sick
null
null
13
You [IDENTITY_P] make me sick.
derog_neg_emote_h
89
You disabled people make me sick.
hateful
disabled people
directed
make me sick
make sb. sick
null
null
13
You [IDENTITY_P] make me sick.
derog_neg_emote_h
90
You Muslims make me sick.
hateful
Muslims
directed
make me sick
make sb. sick
null
null
13
You [IDENTITY_P] make me sick.
derog_neg_emote_h
91
You immigrants make me sick.
hateful
immigrants
directed
make me sick
make sb. sick
null
null
13
You [IDENTITY_P] make me sick.
derog_neg_emote_h
92
There is nothing worse than a woman.
hateful
women
general
nothing worse
worse
null
null
14
There is nothing worse than a [IDENTITY_S].
derog_neg_emote_h
93
There is nothing worse than a trans person.
hateful
trans people
general
nothing worse
worse
null
null
14
There is nothing worse than a [IDENTITY_S].
derog_neg_emote_h
94
There is nothing worse than a gay person.
hateful
gay people
general
nothing worse
worse
null
null
14
There is nothing worse than a [IDENTITY_S].
derog_neg_emote_h
95
There is nothing worse than a black person.
hateful
black people
general
nothing worse
worse
null
null
14
There is nothing worse than a [IDENTITY_S].
derog_neg_emote_h
96
There is nothing worse than a disabled person.
hateful
disabled people
general
nothing worse
worse
null
null
14
There is nothing worse than a [IDENTITY_S].
derog_neg_emote_h
97
There is nothing worse than a Muslim.
hateful
Muslims
general
nothing worse
worse
null
null
14
There is nothing worse than a [IDENTITY_S].
derog_neg_emote_h
98
There is nothing worse than an immigrant.
hateful
immigrants
general
nothing worse
worse
null
null
14
There is nothing worse than a [IDENTITY_S].
derog_neg_emote_h
99
Seeing women in this place makes me so angry.
hateful
women
general
angry
angry
null
null
15
Seeing [IDENTITY_P] in this place makes me so angry.
derog_neg_emote_h
100
Seeing trans people in this place makes me so angry.
hateful
trans people
general
angry
angry
null
null
15
Seeing [IDENTITY_P] in this place makes me so angry.
End of preview. Expand in Data Studio

Paul/hatecheck(複製)

データの内容

ヘイトスピーチの検出モデルを機能ごとに調べるためのテストケースのデータ。29 種類の機能について正解ラベル付きの 3,728 件のテストケースからなり、ヘイトかどうかの正解を label_gold に入れている。1 件ごとにテストの文を test_case、機能名を functionality、対象の集団を target_ident に入れている。テンプレートから作られた合成の文で、評価用としての性格が強いとカードにある。

データの概要

項目 値
行数 3,728(1 構成・1 分割の合計)
列数 12
容量 658.42 KB(ファイル 3 本)
ファイルの形式 csv 1
言語 en
公開 public

構成と分割

構成(config) 分割(split)ごとの行数 列数
default test 3,728 12

列(構成 default・分割 test)

列 型 先頭の行の値
functionality string derog_neg_emote_h
case_id int64 1
test_case string I hate women.
label_gold string hateful
target_ident string women
direction string general
focus_words string hate
focus_lemma string hate
ref_case_id float64 null
ref_templ_id float64 null
templ_id int64 1
case_templ string I hate [IDENTITY_P].

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

Paul/hatecheck をリビジョン 9d2ac89df04254e5c427bcc8d61b6d6c83a1f59b 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

(元のカードに該当の見出しが無いので、ラベルの付け方に触れた段落を抜き出した)

case_id テストケースの一意の ID(当初生成した 3,901 件のケースそれぞれに割り当てた)。

@inproceedings{rottger-etal-2021-hatecheck, title = "{H}ate{C}heck: Functional Tests for Hate Speech Detection Models", author = {R{"o}ttger, Paul and Vidgen, Bertie and Nguyen, Dong and Waseem, Zeerak and Margetts, Helen and Pierrehumbert, Janet}, booktitle = "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)", month = aug, year = "2021", address = "Online", publisher = "Association for Computational Linguistics", url = "https://aclanthology.org/2021.acl-long.4", doi = "10.18653/v1/2021.acl-long.4", pages = "41--58", abstract = "Detecting online hate is a difficult task that even state-of-the-art models struggle with. Typically, hate speech detection models are evaluated by measuring their performance on held-out test data using metrics such as accuracy and F1 score. However, this approach makes it difficult to identify specific model weak points. It also risks overestimating generalisable model performance due to increasingly well-evidenced systematic gaps and biases in hate speech datasets. To enable more targeted diagnostic insights, we introduce HateCheck, a suite of functional tests for hate speech detection models. We specify 29 model functionalities motivated by a review of previous research and a series of interviews with civil society stakeholders. We craft test cases for each functionality and validate their quality through a structured annotation process. To illustrate HateCheck{'}s utility, we test near-state-of-the-art transformer models as well as two popular commercial models, revealing critical model weaknesses.", }

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
general_english_nlu.md 1. 単文分類: 感情・トピック・意図・感情語・スパム・毒性・皮肉(noul / choice / score に変換) (L77) cc-by-4.0 確認済 原文注記: - Paul/hatecheck: arXiv 2012.15606(タグ)。csv。license cc-by-4.0。API downloads 1,370。機能テスト型の合成ケース。noul。評価用の色が強い。—
jev_sources_tasksource_bekko.csv CSV(bekko-system-one-dataset-v0) CC BY 4.0 確認済 [Jev 元データ source] source_annotation_or_rule_labels(eval only)

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: HateCheck

データセットの説明

HateCheck は、ヘイトスピーチ検出モデルのための機能テストのスイートである。データセットには、29 種類の機能テストにわたる、検証済みのテストケース 3,728 件が含まれる。19 の機能テストは、ヘイトの種類の違いに対応する。残りの 11 の機能テストは、難しい種類の非ヘイトを扱う。これにより、モデルの性能について的を絞った診断的な知見が得られる。

ACL 論文では、HateCheck で検証したすべての商用および学術のヘイトスピーチ検出モデルに、重大な弱点があることがわかった。結果とさらなる議論、およびデータセットに関するより詳しい情報と完全なデータステートメントについては、下記にリンクした論文を参照すること。

データセットの構造

「test.csv」には、検証済みのテストケース 3,728 件すべてが含まれる。各テストケース(行)は次の属性を持つ:

functionality テストケースが検証する機能の略称。

case_id テストケースの一意の ID(当初生成した 3,901 件のケースそれぞれに割り当てた)。

test_case テストケースの本文。

label_gold テストケースの正解ラベル(ヘイト/非ヘイト)。同じ機能に属するテストケースはすべて、同じ正解ラベルを持つ。

target_ident 該当する場合、テストケースが対象とする、または言及する保護された集団。テストスイートでは 7 つの保護された集団を扱う: 女性、トランスジェンダーの人々、ゲイの人々、黒人、障害のある人々、ムスリム、移民。

direction ヘイトのケースについて、保護された集団の一員である個人に 向けられている か、集団全体に 向けられている かを示す、二値の副次ラベル。

focus_words 該当する場合、テストケースの中の重要な語または句(例: "cut their throats")。

focus_lemma 該当する場合、対応する見出し語(例: "cut sb. throat")。

ref_case_id ヘイトのケースについては、該当する場合、それを生成するために摂動させた、より単純なヘイトのケースの ID。非ヘイトのケースについては、該当する場合、対比されるヘイトのケースの ID。

ref_templ_id 同じ内容を、テンプレートの ID について示したもの。

templ_id テストケースの生成元となったテンプレートの一意の ID(3,901 件の当初のケースを生成するための 866 件のケースとテンプレートそれぞれに割り当てた)。

引用情報

HateCheck を使う場合は、ACL の論文を引用すること:

@inproceedings{rottger-etal-2021-hatecheck, title = "{H}ate{C}heck: Functional Tests for Hate Speech Detection Models", author = {R{"o}ttger, Paul and Vidgen, Bertie and Nguyen, Dong and Waseem, Zeerak and Margetts, Helen and Pierrehumbert, Janet}, booktitle = "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)", month = aug, year = "2021", address = "Online", publisher = "Association for Computational Linguistics", url = "https://aclanthology.org/2021.acl-long.4", doi = "10.18653/v1/2021.acl-long.4", pages = "41--58", abstract = "Detecting online hate is a difficult task that even state-of-the-art models struggle with. Typically, hate speech detection models are evaluated by measuring their performance on held-out test data using metrics such as accuracy and F1 score. However, this approach makes it difficult to identify specific model weak points. It also risks overestimating generalisable model performance due to increasingly well-evidenced systematic gaps and biases in hate speech datasets. To enable more targeted diagnostic insights, we introduce HateCheck, a suite of functional tests for hate speech detection models. We specify 29 model functionalities motivated by a review of previous research and a series of interviews with civil society stakeholders. We craft test cases for each functionality and validate their quality through a structured annotation process. To illustrate HateCheck{'}s utility, we test near-state-of-the-art transformer models as well as two popular commercial models, revealing critical model weaknesses.", }

原文(英語)

Dataset Card for HateCheck

Dataset Description

HateCheck is a suite of functional test for hate speech detection models. The dataset contains 3,728 validated test cases in 29 functional tests. 19 functional tests correspond to distinct types of hate. The other 11 functional tests cover challenging types of non-hate. This allows for targeted diagnostic insights into model performance.

In our ACL paper, we found critical weaknesses in all commercial and academic hate speech detection model that we tested with HateCheck. Please refer to the paper (linked below) for results and further discussion, as well as further information on the dataset and a full data statement.

Dataset Structure

"test.csv" contains all 3,728 validated test cases. Each test case (row) has the following attributes:

functionality The shorthand for the functionality tested by the test case.

case_id The unique ID of the test case (assigned to each of the 3,901 cases we initially generated)

test_case The text of the test case.

label_gold The gold standard label (hateful/non-hateful) of the test case. All test cases within a given functionality have the same gold standard label.

target_ident Where applicable, the protected group targeted or referenced by the test case. We cover seven protected groups in the test suite: women, trans people, gay people, black people, disabled people, Muslims and immigrants.

direction For hateful cases, the binary secondary label indicating whether they are directed at an individual as part of a protected group or aimed at the group in general.

focus_words Where applicable, the key word or phrase in a given test case (e.g. "cut their throats").

focus_lemma Where applicable, the corresponding lemma (e.g. "cut sb. throat").

ref_case_id For hateful cases, where applicable, the ID of the simpler hateful case which was perturbed to generate them. For non-hateful cases, where applicable, the ID of the hateful case which is contrasted.

ref_templ_id The equivalent, but for template IDs.

templ_id The unique ID of the template from which the test case was generated (assigned to each of the 866 cases and templates from which we generated the 3,901 initial cases).

Citation Information

When using HateCheck, please cite our ACL paper:

@inproceedings{rottger-etal-2021-hatecheck, title = "{H}ate{C}heck: Functional Tests for Hate Speech Detection Models", author = {R{"o}ttger, Paul and Vidgen, Bertie and Nguyen, Dong and Waseem, Zeerak and Margetts, Helen and Pierrehumbert, Janet}, booktitle = "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)", month = aug, year = "2021", address = "Online", publisher = "Association for Computational Linguistics", url = "https://aclanthology.org/2021.acl-long.4", doi = "10.18653/v1/2021.acl-long.4", pages = "41--58", abstract = "Detecting online hate is a difficult task that even state-of-the-art models struggle with. Typically, hate speech detection models are evaluated by measuring their performance on held-out test data using metrics such as accuracy and F1 score. However, this approach makes it difficult to identify specific model weak points. It also risks overestimating generalisable model performance due to increasingly well-evidenced systematic gaps and biases in hate speech datasets. To enable more targeted diagnostic insights, we introduce HateCheck, a suite of functional tests for hate speech detection models. We specify 29 model functionalities motivated by a review of previous research and a series of interviews with civil society stakeholders. We craft test cases for each functionality and validate their quality through a structured annotation process. To illustrate HateCheck{'}s utility, we test near-state-of-the-art transformer models as well as two popular commercial models, revealing critical model weaknesses.", }

Downloads last month
82

Paper for HYUKI5/Paul__hatecheck