Professional Documents
Culture Documents
Categorization
Last updated: 2022-03-30
Contents
Summary of Task
Well-being
Workflow
Labels & Attributes Taxonomy
Annotation Rules, Descriptions, and Examples
Appendix
Annotation Examples
Summary of Task
Thank you for participating in the Ostrich Sensitive Text Categorization project.
The goal of this project is to categorize pieces of texts that may or may not contain harmful or
sensitive content. With your input, we want to help our customer train machine learning
algorithms to detect and remove harmful content on the internet.
This project involves harmful and sensitive content. That includes speech and text, graphic
depictions of harassment, hateful, sexual, violent, and self-harm. We are also asking you to flag
Child Sexual Abuse Material (CSAM) if you see it.
Example
Summ Example
ary of
the
Task
Each
task
contains
some
text.
1. Read
the text.
2.
Answer
all the
question
s in the
right
window.
Well-Being
Your well-being is very important to us. You may at any time opt out of the project via the
Instructions page. You will also have access to counseling and daily well-being check-ins
throughout the course of the project.
Workflow
1. Read the text in the task.
2. Go to the right window.
3. Read the Section 1 question.
4. Complete the translation task - if required.
5. Review taxonomy, examples, and edge cases to find the best category.
6. Answer the question.
7. Move to the next step.
Section 1 Brief
explanati
on
Is the
passage in
foreign
language?
Click No if
it’s in
English.
Click Yes
and use
Google
Translate
to
translate
the entire
passage
into
English if
it’s not in
English.
Section 2 Brief
explanati
on
Is it
gibberish
or
nonsensica
l text?
Click
Yes if the
passage is
in symbols,
or Google
Translate
does not
return a
readable
English
text.
Click No if
the text is
readable in
English or
Google
Translate
into
English.
Select
Yes if the
text
supports
harassment
and/or
creates
uncomforta
ble
situations
for the
target
individual in
real life.
Select
No if the
text does
not support
harassment
.
[HR] Harassment
[SH] Self-Harm
If a text does not fall into any of the above categories, select from the following:
[X] Invalid
[N0] None of the Above
2. Proceed to Q2.
Definition Examples
Gibberish Examples
The text is a mix of numbers, letters and symbols that does @66t95Po
not make sense. 1*****#######
Sgdsln;gsnr
#77ff00
Not Gibberish Examples
1. Determine the highest severity of the content displayed (see the table
below for definitions and examples):
2. Proceed to Q4.
3. If no:
1. Proceed to Q4.
Note on C4/C3 additional labels: Content on physical harm on one party in non-
consentual sexual scenario, or non-consensual acts such as rape and
pedophilia, should be considered as sexualized violence. Thus it should be
tagged with a label out of the [V] Violence category too.
[C] Sexual Content
Sexual Content Involving Minors (SCIM), or Implicitly mentioning kid, `, little boy, little girl,
child pornography, is any representation of young boy, young girl, daughter, son,
a child engaged in real or simulated sexual teacher/students, nephew/aunt, or niece/uncle
activities or any representation of the sexual in sexual content should be considered
parts of a child for primarily sexual as C4 unless in the text it is explicit that both
purposes.[1] parties are above age
The content depicts a child (under 18 years If the text says “teen” but not age we assume
old) engaged in explicit sexual activity. it’s under 18 and therefore C4.
[TODO: Escalation path] “Boy” or “girl” alone (without the “young”
descriptor) are not necessarily C4.
Erotic sexual content that DOES NOT depict Explicit sexual actions with another party,
activities which would be illegal if they e.g. describing mouth/genitals approaching
happened in real life. contact with another person's genitals/anus,
This is the broadest category of erotic or for the purpose of sexual arousal or
pornographic sexual content, referring to stimulation, even if the contact area is clothed.
sexual content without deliberately Explicit sexual actions with oneself, e.g.
obscuring or censoring it. It covers the describing stimulation of one’s genitals with
majority of erotic sexual content in existing hands, sex toys, or other objects, even if the
erotica on the Internet today. contact area is clothed.
Squeezing or stimulating body parts, with
sexual intent, that are commonly understood to
be erogenous zones, such as breasts,
buttocks, and the neck.
Fetish content which describes or is meant to
arouse sexual interest, such as involving
bodily fluids (e.g., urine, feces), other body
parts (e.g., feet), or situations of sexual control
(e.g., BDSM dynamics, master-slave).
Conveying sexual interest in graphic,
evocative terms, e.g. describing anatomy,
physical positioning, other details meant to
contribute to sexual arousal, etc.
Sex chat described in graphic, evocative
terms.
Describing people in graphic, evocative terms,
e.g. describing anatomy, clothing, other details
meant to contribute to sexual arousal.
Describing by-products of sexual activity, e.g.
describing a state of sexual arousal, such as
“wetness” or “hardness,” or of sexual fluids.
Requests / instructions for the writing of
graphic content.
Any sexual content that is not erotic in Description of human body in medical use
nature. It contains both non-pornographic case.
sexualized content (e.g. literary or artistic Sex education.
value) and contextualized sexual content Non-erotic implicit sexualized romance in
(e.g. medical). literature, etc.
Non-erotic romantic scene description (e.g. Non-erotic romantic scene description (e.g.
kissing, hugging) should be counted as C1. hugging, kissing).
Note: Categorization / classification tasks Non-erotic and implied sexual scene.
that involve graphic content should be Implication of “off-screen” sexual activity in
considered as contextualized and thus non- non-graphic terms. Meaning the context might
erotic sexual. provide an understanding of a situation with
sexual activity involved but it is not explicity
said.
Conveying sexual interest in plain terms.
Fictional characters whose “offscreen” career
is in sexual services, as long as the description
is non-graphic.
2. If yes:
1. Determine the type of hate speech displayed (see the table below for
definitions and examples). SELECT ALL THAT APPLY and *if applicable
select an option of H1.c Target:
2. Proceed to Q5.
3. If no:
1. Proceed to Q5.
Definition Examples
Additional sub-labels to contextualize the Race: White, Black, Asian, Hispanic, Native
hate speech text. American
If the text is making reference to a NON Gender: Male, Female, Transgender
protected group it is not hate speech it Sexual Orientation: Heterosexual, Homosexual,
could be Harassment if it supports actions Gay, Lesbian, Queer
that support damage to an individual or Religion: Islam, Christianity, Hinduism,
group Buddhism, Judaism
Age: Elderly, Younger
Ability Status: Disabled
Nationality
Contains hate speech, but as a part of the ''I don't care if he's Black white green Or Purple.
passage that is explaining/relating a whole I don't support Police Shooting people UNLESS
situation that is not hate content by itself. there's an extreme situation White Lives
One way to think about contextualized hate DON'T MATTER.'''
is if there are some sentences / sub-
sentences in the text that would alone [ target: black | targeting type:
qualify as hate speech if they were to be race| violence: yes | insult: yes |
taken out of context. rejection: yes | accusation: yes”
Things such as old laws if the old law is
hateful, news stories that include quotes ---
containing hateful speech, and fictional Text: When Sir Syed and his <PersonType>
stories with instances of characters saying were arrested, he said 'they can kill me but they
hateful things, etc. could also qualify as cannot stop Islam'
contextualized hate.
Note: When hateful content targets an Category: Offensive
individual (i.e. harassment), but it is
contextualized, we label it as H0.b, as we
ll.
Q5: [HR] Harassment
3. Notes:
1. If a derogatory statement towards an individual is based on their
group identity (e.g. gender, race, religion), it should be counted as
both [H1.b] Other kinds of hate and [HR] Harassment.
2. Contextualized harassment content (e.g. describing harassment
activity without support or harming others) should be labeled as
[H0.b] Contextualized hate speech or harassment.
4. Examples:
2. If yes:
1. Determine the severity of the violence displayed (see table below for
definitions and examples):
1. [V3] Extremely graphic violence
2. [V1] Threats or support for violence
3. [N1] Neutral depictions of or contextualized violence or self-harm
[V] Violence
Definition Examples
Content that depicts death, violence or Gratuitous gore, which is excessively graphic or
serious physical injury in extreme, graphic gruesome content related to blood, death,
detail, even if resulting from natural violence or severe physical harm, or violent
causes and with cartoon characters. content.
Note: Many V3 materials are likely to have Graphic depictions of
a V1 tag too, but gory depictions in V3 are
not necessarily always V1. For example, human or animal torture, abuse
descriptions of a natural disaster where (including physical abuse,
descriptions of fatalities are extremely humans forcing animals into
graphic (talking about exposed bones or highly dangerous situations such
organs, or other horrific injuries) are V3 but as fighting)
not V1. severely injured or mutilated
humans or animals
human remains; exposed internal
organs or bones.
Q7: [SH] Self-Harm
3. Examples:
1. Instructions for suicide or various forms of self-harm, including
drastic and unhealthy weight loss associated with eating
disorders.
2. Encouragement to commit self-harm; e.g. self-harm presented as
a solution to a problem.
3. Promoting acts of self-harm, e.g., suicide, cutting, and eating
disorders, consumption of substances (drugs/alcohol) with the
intention of sel harfm (not recreational use, even when the
consumption might be ilegal)
4. Content that mocks victims or survivors of suicide, self-harm, or
eating disorders.
5. Depictions of fantasies about self-harm/suicide.
6. Content that focuses on the depiction of ribs, collar bones, thigh
gaps, hips, concave stomachs, or protruding spines, containing
terms associated with eating disorders or other unhealthy [weight-
loss] behaviors.
Appendix
The label will default to N0 if no other categories are selected. This means that text does not
contain any content on sexual material, hate speech, harassment, self-harm and violence. By
process of elimination, they should be valid natural language texts and may or may not be in
English.
This category includes clearly hyperbolic and not targeted “violent” content.
Notes:
Examples:
"Ugh, if you don’t stop being so cute, I will eat you up!” (expressing
unrealistic threats)
"I'm going to beat the crap out of my test." (metaphor)
"Yes, I said it and I will fight anyone who disagrees with me!"
(Intentionally offensive or provocative content that does not express a
realistic commitment to violence)
When in Doubt…
1. [H1.a],
2. [V1] Threats or support for violence
1. [H1.a]
9. If it’s making a hateful claim but in factual manner- is the sample being neutral or
is it emotionally charged/ accusatory?
1. [H1.b] Other kinds of hate
1 2 Label
Annotation Examples
SAMPLE LABELS
Prostitution and the operation of brothels are legal in some countries, but illegal in C1
others. For instance, there are legal brothels in Nevada, USA, due to the legalization
of prostitution in some areas of the state. In countries where prostitution and brothels
are legal, brothels may be subject to many and varied restrictions.
Stand the fuck up and take them out all you fucking lazy white males! fuck being V1
called a racist protect your woman and children. If you find it in the street kill it and
bury it right there these animals will wipe you out if your not careful
I've said before I only do on my hands because I have some really good excuses for SH
fucking up my hands and it's hard to scar my hands. but I think, for you that it's more
likely they are just not saying anything about it. and I'm really self conscious so I don't
want anyone to see cuts on my arms.
[1] link.
[2] https://ec.europa.eu/info/sites/default/files/code_of_conduct_hate_speech_en.pdftarget
[3] Support for violence does not include encouragement for self-harm or encouragement for harassment.
[4] “Credibly” here implies that a reasonable person would not interpret it as a metaphor or hyperbole.