File size: 7,984 Bytes
ace5e58
 
 
 
 
b21cbea
f810cac
 
 
ec3f809
ace5e58
 
1b287dd
ace5e58
feaad7c
 
b185023
 
b338c39
ace5e58
b338c39
b185023
 
b338c39
 
b185023
 
b338c39
2030665
b185023
154fdc1
b185023
f506268
b185023
 
7beb6a1
b185023
f5f934f
 
48fa0ba
e13e3cd
f5f934f
 
 
 
 
5b5a62b
f5f934f
 
 
 
 
 
 
6f9bf75
f5f934f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
086f0d0
f5f934f
 
 
086f0d0
 
 
f5f934f
 
 
b185023
ace5e58
b185023
ace5e58
9ccc732
 
49e5a16
9ccc732
49e5a16
 
 
 
 
 
 
 
 
9ccc732
49e5a16
9ccc732
 
49e5a16
 
 
9ccc732
0660c09
23bddc1
0660c09
23bddc1
 
 
 
 
 
 
071cfd9
 
 
 
 
 
 
 
 
 
23bddc1
 
071cfd9
23bddc1
071cfd9
 
23bddc1
9ccc732
f5f934f
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
---
task_categories:
- text-classification
language:
- en
task_ids:
- sentiment-classification
- hate-speech-detection
size_categories:
- 10K<n<100K
---

# Tweet Annotation Sensitivity Experiment 2: Annotations in Five Experimental Conditions 

***<font color= red>Attention: This repository contains cases that might be offensive or upsetting. We do not support the views expressed in these hateful posts.</font>***

## Description

The dataset contains tweet data annotations of **hate speech** (HS) and **offensive language** (OL) in five experimental conditions. The tweet data was sampled from the corpus created by [Davidson et al. (2017)](https://ojs.aaai.org/index.php/ICWSM/article/view/14955). We selected 3,000 Tweets for our annotation. We developed five experimental conditions that varied the annotation task structure, as shown in the following figure. All tweets were annotated in each condition.

- **<font color= #871F78>Condition A</font>** presented the tweet and three options on a single screen: hate speech, offensive language, or neither. Annotators could select one or both of hate speech, offensive language, or indicate that neither applied.

- Conditions B and C split the annotation of a single tweet across two screens.
  + For **<font color= Blue>Condition B</font>**, the first screen prompted the annotator to indicate whether the tweet contained hate speech. On the following screen, they were shown the tweet again and asked whether it contained offensive language.
  + **<font color= red>Condition C</font>** was similar to Condition B, but flipped the order of hate speech and offensive language for each tweet. 

- In Conditions D and E, the two tasks are treated independently with annotators being asked to first annotate all tweets for one task, followed by annotating all tweets again for the second task.
  + Annotators assigned **<font color=green>Condition D</font>** were first asked to annotate hate speech for all their assigned tweets, and then asked to annotate offensive language for the same set of tweets.
  + **Condition E** worked the same way, but started with the offensive language annotation task followed by the hate speech annotation task.

We recruited US-based annotators from the crowdsourcing platform [Prolific](https://www.prolific.com/) during November and December 2022. Each annotator annotated up to 50 tweets. The dataset also contains demographic information about the annotators. Annotators received a fixed hourly wage in excess of the US federal minimum wage after completing the task.

<img src="https://raw.githubusercontent.com/chkern/tweet-annotation-sensitivity/main/fig/exp_conditions.png" width = "300" height = "200" alt="" align=center />


## Codebook

| Column Name     | Description       | Type       |
| -------------- | ------------------ |---------------- |
| case_id  | case ID | integer |
| duration_seconds  | duration of connection to task in seconds | integer |
| last_screen  | last question answered  | factor |
| device  | device type  | factor |
| ethn_hispanic  | Hispanic race/ethnicity  | binary |
| ethn_white  | White race/ethnicity  | binary |
| ethn_afr_american | African-American race/ethnicity  | binary |
| ethn_asian | Asian race/ethnicity  | binary |
| ethn_sth_else | race/ethnicity something else  | binary |
| ethn_prefer_not | race/ethnicity prefer not to say  | binary |
| age | age  | integer |
| education | education attainment  <br>1: Less than high school <br>2: High school <br>3: Some college <br>4: College graduate <br>5: Master's degree or professional degree (law, medicine, MPH, etc.) <br>6: Doctoral degree (PhD, DPH, EdD, etc.)| factor |
| english_fl | English as first language   | binary |
| twitter_use | Twitter use frequency  <br>1: Most days <br>2: Most weeks, but not every day <br>3: A few times a month <br>4: A few times a year <br>5: Less often <br>6: Never  | factor |
| socmedia_use | social media use frequency  <br>1: Most days <br>2: Most weeks, but not every day <br>3: A few times a month <br>4: A few times a year <br>5: Less often <br>6: Never  | factor |
| prolific_hours  | workload on the platform prolific in hours in the last month | integer |
| task_fun  | task perception: fun | binary |
| task_interesting  | task perception: interesting | binary |
| task_boring  | task perception: boring | binary |
| task_repetitive  | task perception: repetitive | binary |
| task_important  | task perception: important | binary |
| task_depressing  | task perception: depressing | binary |
| task_offensive  | task perception: offensive | binary |
| repeat_tweet_coding  | likelihood for another tweet task <br>1: Not at all likely <br>2: Somewhat likely <br>3: Very likely | factor |
| repeat_hs_coding  | likelihood for another hate speech task <br>1: Not at all likely <br>2: Somewhat likely <br>3: Very likely | factor |
| target_online_harassment  | targeted by hateful online behavior | binary |
| target_other_harassment  | targeted by other hateful behavior | binary |
| party_affiliation  | party identification <br>1: Republican <br>2: Democrat <br>3: Independent  | factor |
| societal_relevance_hs  | relevance perception of hate speech  <br>1: Not at all likely <br>2: Somewhat likely <br>3: Very likely | factor |
| annotator_id  | annotator ID | integer |
| condition  | experimental conditions (A-E)  | factor |
| tweet_batch  | tweet ID in batch  | factor |
| hate_speech  | hate speech annotation  | logical |
| offensive_language  | offensive language annotation  | logical |
| tweet_id  | tweet ID  | integer |
| orig_label_hs  | number of persons who annotated the tweet as hate speech in the original dataset from [Davidson et al. (2017)](https://ojs.aaai.org/index.php/ICWSM/article/view/14955)  | integer |
| orig_label_ol  | number of persons who annotated the tweet as offensive language in the original dataset from [Davidson et al. (2017)](https://ojs.aaai.org/index.php/ICWSM/article/view/14955)  | integer |
| orig_label_ne  | number of persons who annotated the tweet as neither in the original dataset from [Davidson et al. (2017)](https://ojs.aaai.org/index.php/ICWSM/article/view/14955)  | integer |
| tweet_hashed  | tweet with usernames hashed  | character |




## Citation

If you find the dataset useful, please cite:
```
@inproceedings{kern-etal-2023-annotation,
    title = "Annotation Sensitivity: Training Data Collection Methods Affect Model Performance",
    author = "Kern, Christoph  and
              Eckman, Stephanie  and
              Beck, Jacob  and
              Chew, Rob  and
              Ma, Bolei  and
              Kreuter, Frauke",
    editor = "Bouamor, Houda  and
              Pino, Juan  and
              Bali, Kalika",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2023",
    month = dec,
    year = "2023",
    address = "Singapore",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.findings-emnlp.992",
    pages = "14874--14886",
}
```

```
@inproceedings{beck-etal-2024-order,
    title = "Order Effects in Annotation Tasks: Further Evidence of Annotation Sensitivity",
    author = "Beck, Jacob  and
              Eckman, Stephanie  and
              Ma, Bolei  and
              Chew, Rob  and
              Kreuter, Frauke",
    editor = {V{\'a}zquez, Ra{\'u}l  and
              Celikkanat, Hande  and
              Ulmer, Dennis  and
              Tiedemann, J{\"o}rg  and
              Swayamdipta, Swabha  and
              Aziz, Wilker  and
              Plank, Barbara  and
              Baan, Joris  and
              de Marneffe, Marie-Catherine},
    booktitle = "Proceedings of the 1st Workshop on Uncertainty-Aware NLP (UncertaiNLP 2024)",
    month = mar,
    year = "2024",
    address = "St Julians, Malta",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.uncertainlp-1.8",
    pages = "81--86",
}
```