The Virality of Hate Speech on Social Media

2025-05-06 0 0 1.83MB 22 页 10玖币

侵权投诉

ABDURAHMAN MAAROUF, LMU Munich & Munich Center for Machine Learning, Germany

NICOLAS PRÖLLOCHS, JLU Giessen, Germany

STEFAN FEUERRIEGEL, LMU Munich & Munich Center for Machine Learning, Germany

Online hate speech is responsible for violent attacks such as, e.g., the Pittsburgh synagogue shooting in 2018,

thereby posing a signicant threat to vulnerable groups and society in general. However, little is known about

what makes hate speech on social media go viral. In this paper, we collect

𝑁=25,219

cascades with

65,946

retweets from X (formerly known as Twitter) and classify them as hateful vs. normal. Using a generalized

linear regression, we then estimate dierences in the spread of hateful vs. normal content based on author and

content variables. We thereby identify important determinants that explain dierences in the spreading of

hateful vs. normal content. For example, hateful content authored by veried users is disproportionally more

likely to go viral than hateful content from non-veried ones: hateful content from a veried user (as opposed

to normal content) has a

3.5

times larger cascade size, a

3.2

times longer cascade lifetime, and a

1.2

times larger

structural virality. Altogether, we oer novel insights into the virality of hate speech on social media.

CCS Concepts: •Human-centered computing

→

Empirical studies in collaborative and social com-

puting;Social media;•Applied computing →Sociology.

Additional Key Words and Phrases: Hate speech, Twitter/X, social media, content spreading, virality, regression

analysis

1 INTRODUCTION

Hate speech on social media poses a widespread problem of societal signicance. In 2021, around

40 % of the U.S. society has personally experienced online hate speech [

]. Hate speech is known

to have a negative impact on the mental as well as physical well-being of online users. In particular,

young adults tend to suer from the psychological consequences [

]. In the context of the Global

South, online hate speech often serves as a tool to seed fear towards ethnic and religious minorities,

especially in the political discourse [

]. Online hate speech further reinforces hateful

attitudes (i.e., radicalization) and motivates violent acts including hate crimes [

]. For example,

online hate speech has played a crucial role in the 2018 Pittsburgh synagogue shooting [

], the

2017 Rohingya genocide in Myanmar [33], and anti-Muslim mob violence in Sri Lanka [56].

While negative consequences of hate speech have been widely documented, our understanding

of how hateful content spreads on social media is still emerging. Prior work has studied targets of

hate speech, which are characterized by higher social activity, older accounts, and more followers

[

]. Moreover, several studies have analyzed characteristics of hateful users [

]. Those

have been found to be associated with a higher social media activity, more complex word usage,

and a higher density in terms of connections among each other, compared to normal users. Social

media cascades by hateful users have typically a larger spread, longer lifetime, and higher virality

[

]. However, all of these analyses were conducted at the user level, and

not

at the content level.

In particular, no prior work analyzed the dierences in (and the determinants for) the spreading

dynamics of hateful vs. normal content.

We hypothesize that dierences in the spreading of hateful vs. normal online content can be

explained by characteristics of both the author and the content. For instance, in terms of author

characteristics, one may expect that a user with a veried status disproportionately promotes the

spreading of hateful content. The reason is that a veried status lends the author a higher credibility

Authors’ addresses: Abdurahman Maarouf, LMU Munich & Munich Center for Machine Learning, Germany, a.maarouf@

lmu.de; Nicolas Pröllochs, JLU Giessen, Germany, nicolas.proellochs@wi.jlug.de; Stefan Feuerriegel, LMU Munich & Munich

Center for Machine Learning, Germany, feuerriegel@lmu.de.

arXiv:2210.13770v3 [cs.SI] 25 Nov 2024

2 Abdurahman Maarouf, Nicolas Pröllochs, and Stefan Feuerriegel

and trustworthiness [

], and, as a result, others may be more willing to share hate speech if it

comes from a gure with large social inuence. Moreover, societal leaders play an important role

in shaping the social norms of other users [

]. Hence, individuals with a special social status (such

as a veried status on social media platforms) may be able to let hateful content appear socially

acceptable, thus adding to the proliferation of hate speech. In terms of content characteristics, one

may expect that, for instance, the use of mentions (i.e., using “@username” in tweets) aects the

spread of hate speech. Mentions make hate speech appear directed against a specic user or entity,

which diers from generalized hate that is against a general group of individuals [

]. Hence, it

is likely that directed vs. generalized hate also have dierent spreading dynamics. Note that our

results later provide empirical evidence in favor of these hypotheses (e.g., we nd that a veried

status adds to the proliferation of hate speech, while mentions and hashtags hamper the spread).

In this work, we analyze dierences in the spreading of hateful vs. normal content on social

media. Specically, we aim to generate an understanding of what makes online hate speech go

viral:

RQ: What are the determinants for the dierences in the spreading dynamics of hateful vs. normal

content on online social media?

Data: We collected a comprehensive dataset of

𝑁=25,219

retweet cascades from X (formerly

known as Twitter). Each cascade was human-labeled into two categories: hateful and normal. Our

dataset comprises both the root tweet and all retweets, i.e., the complete retweet cascades. The root

tweets in our dataset have received 65,946 retweets by 62,051 dierent users. For each root tweet,

we collected an extensive set of both author characteristics (i.e., number of followers, number of

followees, veried status, tweet volume, and account age) and content characteristics (i.e., media

items, hashtags, mentions, and tweet length).

Methods:

We perform an explanatory regression analysis using generalized linear models (GLMs).

Thereby, we aim to identify author and content characteristics that explain the dierences in

the spread of hateful vs. normal content. To quantify the spread on social media, we analyze the

following structural properties: (i) cascade size, i.e., how many times the root tweet was retweeted

and which thus measures the overall exposure; (ii) cascade lifetime, i.e., how long the cascade lived;

and (iii) structural virality, i.e., a measure for the eectiveness of the spreading process [19].

Findings: To the best of our knowledge, this is the rst work identifying what makes hate speech

go viral. Using explanatory regression modeling, we yield the following novel ndings:

(1)

Cascades with hateful content (as compared to cascades with normal content) grow larger

in size, live longer, and are of larger structural virality.

(2)

Author characteristics explain dierences in the spread of hateful vs. normal content. In

particular, cascades with hateful content from veried authors are disproportionally larger

in size, longer in lifetime, and higher in structural virality.

(3)

Content characteristics explain dierences in the spread of hateful vs. normal content. In

particular, cascades with hateful content including mentions or hashtags are associated

with a smaller size, lifetime, and structural virality.

2 RELATED WORK

In the following, we review dierent literature streams that are particularly relevant to our work,

namely our subject (online hate speech) and our methods (modeling spreading dynamics).

1Code and data for our analysis are available via https://github.com/abdumaa/hatespeech_virality.

The Virality of Hate Speech on Social Media 3

2.1 Hate Speech on Social Media

Hate speech has been dened by the United Nations Strategy and Plan of Action as “any kind of

communication in speech, writing or behaviour, that attacks or uses pejorative or discriminatory

language with reference to a person or a group on the basis of who they are, in other words, based

on their religion, ethnicity, nationality, race, colour, descent, gender, or other identity factor.” While

the legal denitions vary across countries, the common principle is that hate speech expresses

animosity against specic groups, often even calling for violence.

Authors of hateful content on X have several characteristics that distinguish them from normal

authors. For example, hateful content is typically created by authors with a higher tweet activity, a

more specic, non-trivial word usage, and a more dense connection among each other, as compared

to normal users [

]. Online hate is also a common concern in the Global South: Prior research

on social media in the Global South characterizes hateful users by creating and spreading an illusion

of fear, thereby eliciting hate towards minority groups [

]. For example, in India, hateful users are

particularly active in the political discourse by using symbols and past events (e.g., the COVID-19

pandemic) to spread hate against contesting parties and minorities [

]. Moreover, hateful users

in the Global South are characterized as having a larger following as well as being more veried,

leading to a more polarized and inuenced audience [

]. Some works focus also on specic types

of online hate speech such as hate against politicians [23] or religious hate [2, 26].

Other works examine the users targeted by hate speech. For example, gures from American

politics are more likely to receive hateful replies on X when they are persons of color from the

Democratic party, white Republicans, women, and/or authors of content with negative sentiment

[

]. Moreover, users employing moralized language are more likely to receive hate speech [

In general, hate speech on X is predominantly addressed against users with an older account, a

higher tweet activity, and more followers, as compared to normal users [16].

Note that the above works perform analyses at the user level, but not at the content level. These

works can answer who composes/receives hate speech but not how hate travels. Hence, the spreading

dynamics of hate speech on social media remain unknown.

2.2 Spreading Dynamics on Social Media

Previous research has aimed at a better understanding of the spreading dynamics of social media

content and thus why some tweets go viral and others do not. For this, dierent structural properties

of retweet cascades have been analyzed: (1) cascade size, (2) cascade lifetime, and (3) structural

virality. Here, cascade size refers to how many times a tweet was retweeted and thus measures

the overall exposure [

]. The cascade lifetime refers to the overall duration the cascade

was active [

]. The so-called structural virality is a statistical metric to quantify of the trade-

o between depth and breadth of the cascade and should thus capture how eectively a tweet

propagates [

]. The denition of structural virality is related to the Wiener index, and thus

measures the average distance between all pairs of nodes in the cascade.

Dierences in the structural properties of cascades have been found across dierent dimensions.

These can be grouped into author characteristics and content characteristics. (1) Author charac-

teristics include, e.g., number of followers, number of followees, tweet volume (i.e., the previous

activity on X), account age, and veried status [

]. For example, authors with more followers are

theorized to have larger “social inuence,” and their tweets should thus reach a larger audience

[

]. (2) Content characteristics include, e.g., the use of media items (e.g., images, videos,

URLs), mentions, and hashtags [

]. For example, URLs and hashtags have been found to be

positively associated with spreading dynamics [55].

4 Abdurahman Maarouf, Nicolas Pröllochs, and Stefan Feuerriegel

2.3 The Spread of Online Hate Speech

There are only a few works that have modeled the spread of online hate speech, yet with a dierent

objective from ours. On the one hand, there are predictive models that aim to forecast certain

behavior. For example, some works aim to predict the probability with which hate speech is

retweeted [

]. However, predictive models are only trained on hateful content and, therefore,

do not oer explanations on why the spread of hateful vs. normal content diers.

On the other hand, there is one work making quantitative comparisons between hateful and

non-hateful authors on Gab [

]. Therein, the authors create a labeled dataset of hateful vs. normal

posts and reposts on the platform Gab. The authors report summary statistics and nd that cascades

by hateful users live longer, are larger in size, and exhibit more structural virality. Moreover, they

nd that authors of hateful content are more inuential, cohesive, and proactive than authors of

normal content. However, this work is dierent from ours in that it compares the spread by hateful

vs. normal authors, but

not

by hateful vs. normal content. Hence, the determinants for why hateful

content goes viral remain unknown.

Research gap: To the best of our knowledge, there is no prior work studying what makes hate

speech go viral on online social media. For this reason, we provide the rst analysis identifying

determinants that explain dierences in the spread of hateful vs. normal content on online social

media.

3 DATA

3.1 Data Collection

We analyze a comprehensive dataset of hateful and normal tweets on the social media platform X

[

]. We select X in our study as it represents a platform with high popularity. In 2022, it counted

around 436 million monthly active users.

Moreover, hate speech is still widespread on X. Estimates

suggest that there were around

1.13

million accounts on X violating against their policy banning

hateful content in the second half of the year 2020, which marks an increase of 77% compared to

the rst half of the year.3

We use the dataset from Founta et al

. [17]

in which the tweets were human-labeled via crowd-

sourcing into two categories: hateful and normal. The dataset was created using a boosted sampling

procedure, ensuring an unbiased dataset as well as a large number of annotations for the minority

class of hateful tweets. The nal labels were generated through an iterative process of multiple

annotation rounds. In the rst rounds, annotators were given brief conceptual denitions, allowing

for exploratory annotations. The nal denitions and design choices for the annotation process

were determined based on the evaluations of these exploratory rounds. As a result, the subsequent

and nal round benets from an optimized design choice and more precise denitions, resulting in

more accurate annotations. The nal label for each tweet is determined based on a majority vote of

ve crowdsourced annotations. For more details on the dataset, we refer to Founta et al. [17].

We additionally process the dataset as follows. Using the Twitter Historical API, we rst retrieve

the type of each tweet, that is, whether it is a root tweet, a reply, or a retweet. To achieve this, we

extract the eld “referenced_tweet” (using the “tweet_id” provided by Founta et al

. [17]

) which

contains the type of the tweet as well as the id of the root tweet. We replace retweets with their

corresponding root tweet. As the text of the root tweet and its retweets are identical, we kept the

same annotation in case of replacement. In addition, we lter out replies, such that the dataset

only consists of root tweets. In the next step, we collect all retweets for each root tweet in order to

2https://www.statista.com/

3https://blog.twitter.com/en_us/topics/company/2021/an-update-to-the-twitter-transparency-center

The Virality of Hate Speech on Social Media 5

construct the retweet cascades. Here, we follow the methodological approach in [

] to retrieve the

underlying retweet paths. Specically, we rst query the Twitter Historical API for all retweets that

refer to the ids of the root tweets using the information provided in the eld “referenced_tweet”.

While tweets and retweets have timestamps, the Twitter Historical API does not provide the true

retweet path. Instead, all retweets point to the root tweet, which does not reect the true retweet

cascade (i.e., users can also retweet the retweets of other users, not the root tweet). To overcome

this, we use the method of time-inferred diusion [

]. This method leverages X’s follower graph as

well as the timestamps of the retweets to reconstruct the true retweet path. By considering the

reverse chronological order of follower-followee information along with users’ join dates, we infer

the followership network. Follower-followee information is inferred at the time of the retweet.

The resulting dataset contains

𝑁=25,219

cascades, out of which 991 are classied as hateful and

24,228

as normal. Overall, the cascades comprise

91,165

tweets (root tweets and retweets), out of

which

22,313

are in cascades with hateful content and

68,852

are in cascades with normal content.

3.2 Computed Variables

Explanatory variables (EVs): We compute a comprehensive set of variables at the cascade level,

which will later serve as EVs in our regression model. The key EV in our analysis is given by Hateful,

which is a binary variable denoting whether the cascade resulted from hateful content (=1 if the

content is hateful, and =0 otherwise). We group all remaining variables into author and content

variables as follows.

Author variables: To characterize “who” has authored the root tweets, we use the Twitter API to

retrieve the number of followers and followees for each author. These variables have been shown

to be important determinants for why tweets go viral [

]. Analogous to prior work [

], we

also compute variables that capture an author’s activity on X, namely the tweet volume (i.e., the

number of tweets divided by the account age) and the account age. Furthermore, users of public

interest on X are assigned a veried status

, for which we encode a corresponding binary variable

(=1 if veried, and =0 otherwise).

Content variables: We compute variables to capture “what” is included in the tweets. Specically,

we encode a binary variable indicating whether the tweet includes a media item (=1 if it has a

media item, and =0 otherwise). The media item can refer to an attached image, video, poll and/or

URL. Tweets can further mention other users using “@username,” which we again encode as a

binary variable (=1 if the tweet has a mention, and =0 otherwise). We also encode a binary variable

whether the tweet contains hashtags (=1 if the tweet has a hashtag, and =0 otherwise). Finally, we

compute the length of the raw tweet (in characters), excluding URLs, mentions, and hashtags.

Dependent variables (DVs): Recall that we aim to explain dierences in the cascade structure of

hateful vs. normal content. In order to quantify the structural dierences in cascades, we compute

several DVs. For this, let a cascade

𝑖=

, . . . , 𝑁

be given by a tree structure

𝑇𝑖=(𝑟𝑖, 𝑡𝑖0, 𝑅𝑖)

with

root tweet

𝑟𝑖

, a timestamp

𝑡𝑖0

of the root tweet, and a set of retweets

𝑅𝑖={(𝑝𝑖𝑘 , 𝑙𝑖𝑘 , 𝑡𝑖𝑘 )}𝑘

, where

each retweet is a 3-tuple comprising a parent 𝑝𝑖𝑘 , a level of depth 𝑙𝑖𝑘 , and a timestamp 𝑡𝑖𝑘 .

We then compute the following DVs:

•

Cascade size

𝑦CS

𝑖

:The overall number of tweets and retweets in the cascade, that is,

|𝑅𝑖| +

•

Cascade lifetime

𝑦CL

𝑖

:The overall size of the time frame during which the tweet travels

through the network, dened as max {𝑡𝑖𝑘 }𝑘−𝑡𝑖0.

•

Structural virality

𝑦SV

𝑖

:The so-called structural virality [

] is a measure for the trade-o

between cascade depth and breadth. Formally, it is dened as the average “distance” between

4The data is in the supplements and available via https://anonymous.4open.science/r/hatespeech_virality-2218/

5https://help.twitter.com/en/managing-your-account/about-twitter-veried-accounts

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

TheViralityofHateSpeechonSocialMediaABDURAHMANMAAROUF,LMUMunich&MunichCenterforMachineLearning,GermanyNICOLASPRÖLLOCHS,JLUGiessen,GermanySTEFANFEUERRIEGEL,LMUMunich&MunichCenterforMachineLearning,GermanyOnlinehatespeechisresponsibleforviolentattackssuchas,e.g.,thePittsburghsynagogueshootingin2018,th...

展开>> 收起<<

The Virality of Hate Speech on Social Media.pdf

共22页,预览5页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

The Virality of Hate Speech on Social Media

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: