Table 1
Taxonomy of online harm used in manual annotation.
| I. incendiary speech (assaultive speech, extreme speech, dangerous speech, the glorification of violence) |
II. pejorative words and expressions
|
III. insulting/abusive/offensive uses
|
| IV. in/out-group (divisive speech) |
| V. codes |
Figure 1
User statistics per year: newly added users and users active the previous year.
Table 2
Annual message statistics.
| 2016 | 2017 | 2018 | 2019 | 2020 | 2021 | |
|---|---|---|---|---|---|---|
| Number of messages | 410 | 2601 | 1456 | 5865 | 4059 | 15,603 |
| Max. daily messages | 75 | 49 | 22 | 449 | 663 | 3,696 |
| Most active day | Dec. 14 | Feb. 8 | Sep. 16 | Jun. 27 | Dec. 23 | Jan. 9 |
Figure 2
Messages per day in 2020.
Figure 3
Messages per day in 2021.
Table 3
Statistics on 5 main categories: incendiary speech, pejorative words and expressions, insulting/offensive/abusive uses, in/out-group, code words.
| TAG | I. INCENDIARY | II. PEJORATIVE | III. OFFEN- SIVE USES | IV. IN/OUT- GROUP | V. CODES | ALL |
|---|---|---|---|---|---|---|
| Number of messages | 98 | 273 | 115 | 40 | 261 | 787 |
| Fraction | 12% | 35% | 15% | 5% | 33% | 100% |
Table 4
Confusion matrix of multi-class annotation.
| MANUAL\AUTO | WORD LISTS | HATE SPEECH | OFFENSIVE | NEITHER | RECALL |
|---|---|---|---|---|---|
| I incendiary | 29 | 0 | 8 | 61 | 0.378 |
| II pejorative | 85 | 13 | 30 | 89 | 0.590 |
| III offensive uses | 44 | 3 | 10 | 54 | 0.514 |
| IV in-/out-group | 12 | 5 | 4 | 18 | 0.538 |
| V codes | 24 | 0 | 8 | 181 | 0.150 |
| neither | 317 | 11 | 104 | 3395 | 0.887 |
| precision | 0.380 | 0.656 | 0.366 | 0.894 |
Figure 4
Automatically predicted offensive language labels over time. White numbers show the fraction of the ‘neither’ tag.
Figure 5
Sub-category statistics for pejorative words and expressions and insulting/offensive/abusive uses categories.
Figure 6
Binary confusion matrix for manual and automated annotation.