Not All Asians are the Same A Disaggregated Approach to Identifying Anti-Asian Racism in Social Media

2025-05-02 0 0 1.44MB 15 页 10玖币

侵权投诉

Not All Asians are the Same: A Disaggregated Approach to

Identifying Anti-Asian Racism in Social Media

Fan Wu

Arizona State University

Tempe, Arizona, USA

Sanyam Lakhanpal

Arizona State University

Tempe, Arizona, USA

Qian Li

Arizona State University

Tempe, Arizona, USA

Kookjin Lee∗

Arizona State University

Tempe, Arizona, USA

Kookjin.Lee@asu.edu

Doowon Kim

University of Tennessee, Knoxville

Knoxville, Tennessee, USA

Heewon Chae

Arizona State University

Tempe, Arizona, USA

Hazel K. Kwon

Arizona State University

Tempe, Arizona, USA

ABSTRACT

Recent policy initiatives have acknowledged the importance of dis-

aggregating data pertaining to diverse Asian ethnic communities

to gain a more comprehensive understanding of their current sta-

tus and to improve their overall well-being. However, research on

anti-Asian racism has thus far fallen short of properly incorporat-

ing data disaggregation practices. Our study addresses this gap by

collecting 12-month-long data from X (formerly known as Twit-

ter) that contain diverse sub-ethnic group representations within

Asian communities. In this dataset, we break down anti-Asian toxic

messages based on both temporal and ethnic factors and conduct a

series of comparative analyses of toxic messages, targeting dierent

ethnic groups. Using temporal persistence analysis, 𝑛-gram-based

correspondence analysis, and topic modeling, this study provides

compelling evidence that anti-Asian messages comprise various

distinctive narratives. Certain messages targeting sub-ethnic Asian

groups entail dierent topics that distinguish them from those tar-

geting Asians in a generic manner or those aimed at major ethnic

groups, such as Chinese and Indian. By introducing several tech-

niques that facilitate comparisons of online anti-Asian hate towards

diverse ethnic communities, this study highlights the importance

of taking a nuanced and disaggregated approach for understanding

racial hatred to formulate eective mitigation strategies.

CCS CONCEPTS

•General and reference

→

General conference proceedings;

•Social and professional topics

→

Race and ethnicity;•Net-

works →Social media networks.

∗Corresponding author

Permission to make digital or hard copies of all or part of this work for personal or

classroom use is granted without fee provided that copies are not made or distributed

for prot or commercial advantage and that copies bear this notice and the full citation

on the rst page. Copyrights for components of this work owned by others than ACM

must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,

to post on servers or to redistribute to lists, requires prior specic permission and/or a

fee. Request permissions from permissions@acm.org.

WWW ’24, MAY 13 – 17, 2024, Singapore

ACM ISBN 978-1-4503-XXXX-X/18/06. . . $15.00

https://doi.org/XXXXXXX.XXXXXXX

KEYWORDS

Anti-Asian sentiment, Racism against Asian, Panethnicity, Disaggre-

gated Asian American data, Topic modeling, Social media mining

ACM Reference Format:

Fan Wu, Sanyam Lakhanpal, Qian Li, Kookjin Lee, Doowon Kim, Hee-

won Chae, and Hazel K. Kwon. 2024. Not All Asians are the Same: A

Disaggregated Approach to Identifying Anti-Asian Racism in Social Me-

dia. In Proceedings of (WWW ’24). ACM, New York, NY, USA, 15 pages.

https://doi.org/XXXXXXX.XXXXXXX

1 INTRODUCTION

In 2023, the U.S. government released its inaugural report of the

White House Initiative on Asian Americans, Native Hawaiians, and

Pacic Islanders (WHIAANHPI) [

], which aims to develop strate-

gies to enhance justice, equity, and the overall well-being of this pop-

ulation (collectively referred to as Asians hereafter). One of the key

priorities of this initiative is to “make disaggregated data collection

and reporting the norm” across the federal agencies (WHIAANHI,

[

, p.22]). Given the diverse range of ethnic groups within the

Asian American population, the use of disaggregated data prac-

tices is imperative for attaining a thorough understanding of these

distinct Asian communities and relevant policy-making [

]. For

example, when information is reported in an aggregated manner,

the average cancer rate for Asian women is lower than that for

white women. However, when examining segmented records, it

becomes evident that Laotian women have cancer rates more than

nine times higher than those for white women (WHIAANHI, [

p.22]). This dierence highlights the critical need for disaggregated

data, as it reveals the signicant disparities within the Asian Amer-

ican population, enabling policymakers to develop targeted and

eective interventions for specic communities like Laotian women.

Indeed, the importance of collecting and reporting disaggregated

data extends beyond Asian Americans and should be applied to all

“panethnic” communities worldwide [28].

Addressing anti-Asian hate can also benet from disaggregated

data practices. Research on anti-Asian hate has attracted signicant

attention, especially in response to the surge in Sinophobia, a fear

or dislike of China or its people, and hate crimes targeting Asians

arXiv:2210.11640v2 [cs.SI] 12 Feb 2024

WWW ’24, MAY 13 – 17, 2024, Singapore Wu, et al.

in the midst of the COVID-19 pandemic. Negative sentiments to-

wards China and Chinese, as evidenced by derogatory labels such

as “Chinese virus,” along with implicit biases against Asians, have

increased during the pandemic [

]. Federal law enforcement

agencies in the U.S. have alerted the surge in anti-Asian hate crimes

during this period [

]. Various advocacy eorts, including hashtag

campaigns such as “#racismisvirus" and “#stopAsianhate" have also

emerged to counter such anti-Asian sentiments and hate crimes.

As a result, the majority of recent studies on anti-Asian hate

have utilized datasets pertaining to the inuence of the COVID-19

pandemic, focusing on the evidence and consequences of Sinopho-

bia [

]. While the pandemic has undoubtedly served as an

important backdrop for recent Asian hate research, existing liter-

ature has failed to fully acknowledge the problem of anti-Asian

sentiments as an enduring social issue that transcends being merely

a byproduct of the pandemic. Furthermore, it does not adequately

acknowledge that the problem of anti-Asian hate aects a wide

range of ethnic groups within Asian populations, extending beyond

the Chinese community.

The purpose of this study is to ll this void by examining online

anti-Asian hate using a disaggregated-data approach. In particular,

this study broadens the observation period to cover an extended

time frame that encompasses the pre-pandemic, peak pandemic, and

post-peak pandemic phases, and conducts comparative analyses

using disaggregated data based on both temporal and sub-ethnic

breakdowns. This disaggregated approach enables the identication

of nuanced distinctions in the animosity directed toward dierent

ethnic groups within Asian populations. Moreover, it facilitates a

deeper understanding of the intricate inter-ethnic dynamics within

pan-Asian communities.1

The study aims to contribute to the literature by (1) creating

a longitudinal multi-ethnic Asian hate dataset, (2) investigating

temporal trends of anti-Asian messages on X (formerly known as

Twitter), and (3) introducing techniques that enable comparisons of

anti-Asian topics across multiple ethnic communities within pan-

Asian populations. The empirical results presented in this paper

address the following research questions.

(1)

RQ1: (a) Are there changes in the magnitude of anti-Asian

messages over time? (b) How do the trends over time vary

across dierent ethnic groups?

(2)

RQ2: (a) How semantically distant are anti-Asian messages

when comparing those aimed at Asians in a general sense to

those directed at specic sub-ethnic groups? (b) How do the

semantic distances change over time?

(3)

RQ3: (a) How are these topics distributed among messages

targeting Asians in a general sense, those targeting major

ethnic groups like Chinese and Indian, and those directed at

smaller ethnic groups? (b) What are the prevalent topics of

anti-Asian messages?

We collect a 12-month-long social conversations on X (formerly

known as Twitter) that contain diverse sub-ethnic group represen-

tations within Asian communities. Using this dataset, we disag-

gregate anti-Asian toxic messages based on temporal and ethnic

https://www.pewresearch.org/race-ethnicity/2022/08/02/what-it-means-to-be-

asian-in-america/

breakdowns and conduct a series of comparative analyses of toxic

messages targeting various ethnic groups.

Findings from temporal persistence analysis,

𝑛

-gram-based corre-

spondence analysis, and topic modeling reveal several key insights.

First, there is a substantial increase in the number of anti-Asian mes-

sages (especially anti-Chinese) in response to the declaration of the

pandemic, but the average toxicity score has not much aected by

the pandemic. Second, results align with previous research focused

on online hatred towards the Chinese ethnicity, highlighting that

toxic messages, broadly referring to ‘Asians’, had more semantic

similarities with those targeting the Chinese ethnicity than mes-

sages aimed at other specic groups within the Asian community

and that the volume of messages targeting other sub-Asian ethnic

groups was relatively low. Third,

𝑛

-gram-based analysis shows that

toxic messages that attack minority ethnic groups display orthog-

onal semantic features compared to majority-ethnicity-attacking

(e.g., Chinese, Indian) or generic-Asian-attacking messages. In con-

trast, when analyzing minority ethnic groups collectively using

topic modeling, generic-Asian-attacking messages demonstrate

more similar narrative patterns to the collective set of minority

Asian ethnic groups than to a single large group such as Chinese

or Indian.

In essence, this study underscores the importance of recognizing

and addressing the diversity of anti-Asian hate speech. Online anti-

Asian hate speech is complex and nuanced, encompassing various

ethnic backgrounds and the intricate web of biases that exist both

within and beyond the Asian community. In this sense, a multifac-

eted and disaggregated data approach is necessary to understand

and combat the hateful discourse. The methodological approaches

we develop in this paper may be useful to researchers and policy-

makers striving to better comprehend and confront these pressing

challenges, fostering a more inclusive and equitable digital land-

scape for all. Importantly, while the primary focus of this study is

on Asians, “panethnicity” is a form of identication observed glob-

ally, encompassing communities like Latino, Yoruba, or Roma [

Therefore, disaggregated data practices have universal applicability

in addressing social issues relevant to panethnic communities.

2 RELATED WORK AND PROBLEM

STATEMENT

2.1 Online Hate/toxic Speech Research

Hate and toxic speech involves abusive and aggressive language

that attacks a person or group based on attributes such as race,

religion, ethnic origin, national origin, sex, disability, sexual orien-

tation, or gender identity [

]. Much eort in this research

domain has been put on message discovery solutions based on nat-

ural language techniques and models to detect and classify hate

speeches more eciently [

]. Especially, deep learn-

ing has emerged as a powerful technique that learns hidden data

representations and achieves better performance in detecting on-

line hate speech [

]. As a computational aide, state-of-the-art

deep learning models such as BERT

, a BERT ne-tuning model,

RoBERTa [22] have been extensively employed [10, 30].

2Bidirectional Encoder Representations from Transformers [7]

Not All Asians are the Same: A Disaggregated Approach to Identifying Anti-Asian Racism in Social Media WWW ’24, MAY 13 – 17, 2024, Singapore

2.2 Online Anti-Asian Hate Speech Research

Anti-Asian hate speech has recently received attention in response

to the outbreak of COVID-19, during which racism and hateful

messages against Asians have become rampant [

]. On-

line anti-Asian hate speech research has evolved into four types—

COVID-specic hate speech, general anti-Asian sentiments, anti-

Chinese political sentiments, and counter-hate movements such

as “#racismisvirus” and “#stopAsianhate” [

]. Like previous stud-

ies on racist hate speech, anti-Asian speech research has focused

on detecting and classifying anti-Asian toxic contents [

Most of these studies have centered specically on the COVID-19

pandemic. For example, a study introduced a new classier that

identies and categorizes online anti-Asian tweets during COVID-

19 into four classes: hostility against East Asia, criticism of East

Asia, meta-discussions of East Asian prejudice, and a neutral class

[

]. Several studies have focused on the trends and features of

anti-Asian sentiment during COVID-19 [

] and found that

antipathy against Chinese had spillover eects on Asians in general.

One study uses a large-scale web-based media database to compare

global sentiments toward Asians across 20 countries before and

after the pandemic, nding that even though anti-Asian sentiments

are deep-seated and predicated on structural undercurrents of cul-

ture, the pandemic has indirectly and inadvertently exacerbated

those anti-Asian sentiments [27].

2.3 Filling the Void: Considering Temporal and

Ethnic Heterogeneity in Asian Hate Speech

While existing research has developed various statistical/machine

learning (ML) techniques (e.g., hate speech detection) to identify

patterns in anti-Asian sentiments of online speech, the vast amount

of research has been situated in a specic empirical context, that

is, the COVID-19 pandemic, resulting in a rather skewed research

trend. Although COVID-19 has resurfaced the concerns about anti-

Asian hate, anti-Asian racism has been an enduring problem of

inter-ethnic relations. Furthermore, empirical datasets related to

COVID-19 often feature a disproportionately large number of mes-

sages concerning China and Chinese, leading to an assessment of

anti-Asian sentiments that is centered around Chinese-related con-

tents [

]. Even many studies, which examine a generically-

dened ‘Asians’, have (misleadingly) alluded to Asians as being a

homogeneous unity, dismissing the essence of “panethnicity” [

]

that Asian is a concept that bridges very diverse sub-ethnic groups.

While those statistical/ML methods have gained traction as a

pragmatic solution to mitigate the discursive “pollution” in digital

information commons [

], critics point out that such models often

miss contextual nuances, such as bias in dierent demographic and

psycho-graphic subgroups [

]. Some researchers have call for a

more proactive mitigation strategy beyond automated detection. For

example, one study suggested that the polarized opinions sentiment

analyzer system can be used as a plug-in by Twitter to detect and

stop hate speech on its platform [

]. This study recognizes this void

in the existing literature: the predominant focus on the context of

COVID-19 and the negligence of the importance of disaggregating

online hatred messages directed at Asians.

3 DATASETS

3.1 Data Collection

We collect 2.6 million messages from X (Twitter at the time of the

data collection) using its APIs for academic access. The search

period is set from August 2019 to July 2020 to include tweets

from pre-COVID-19 and post-COVID-19 peak periods. We use

search keywords that are related to Asia and 21 sub-ethnic cat-

egories based on the U.S. Census Bureau breakdown.

We pur-

posely choose generic keywords to avoid collecting tweets that are

only specic to an event (e.g., COVID-19). A complete list of the

chosen search keywords is shown in Appendix B.1. With the speci-

ed period and keywords, the initial data set includes 10 million

tweets, out of which 96.3% of tweets contain with eight major key-

words, ‘China’(+‘Chinese’) (31.5%), ‘India’(+‘Indian’) (19%), ‘Japan’

(+‘Japanese’) (16.7%), ‘Korea’(+‘Korean’) (11.‘%), ‘Asia’+(’Asian’)

(10.8%), ‘Pakistan’+(‘Pakistanis’) (3.1%), ‘Vietnam’+(‘Vietnamese’)

(2.3%), and ‘Indonesia’+(‘Indonesian’) (1.7%). Other search key-

words result in less than 1∼2% of the collected tweets.

3.2 Preprocessing

3.2.1 Perspective API. Among the Perspective’s emotional at-

tributes, we refer to the ‘toxicity’ score for initial examination of

our data. Here, the score lies in between [0, 1], with the highest

score 1 being the most toxic. Toxicity is dened as “a rude, disre-

spectful, or unreasonable comment that is likely to make you leave

a discussion”. Toxicity is known to result in the most reliable score

and has been widely used in previous studies [

]. However,

solely relying on toxicity score could both include false positive and

omit false negative anti-Asian tweets because anti-Asian sentiment

is not always expressed in a toxic manner (see Table 4 in Appen-

dix for example). Accordingly, in addition to the toxicity score, we

introduce a manually annotated label, which indicates whether a

tweet contains anti-Asian sentiment. We elaborate it in detail in

the following.

3.2.2 Manual coding. Although the Perspective API provides the

scores that reect the likelihood of assessed tweets being toxic in a

reliable manner, it is challenging to see whether the toxic expression

was being made towards Asian or specic ethnic groups we are in-

terested in. Likewise, it is possible to dismiss anti-Asian tweets that

have low toxicity score. To address this issue, we manually annotate

subsampled tweets to obtain more target-indicative information.

For subsampling, we rst divide the collected tweets into weekly

batches and sort them based on the corresponding toxicity scores.

From each weekly batch, we randomly sample 20 tweets from ten

groups which are broken down based on the toxicity scores (=200

tweets per week), resulting in 10400 tweets in total:

•Group 1: 20 tweets with the scores lie in [0, 0.1],

•Group 10: 20 tweets with the scores lie in [0.9, 1.0].

Then human annotators manually label the tweets on:

[Anti-Asian] Does this tweet contain “anti-

Asian” sentiment? (True/False).

https://www.census.gov/library/stories/2022/05/aanhpi-population-diverse-

geographically-dispersed.html

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：

NotAllAsiansaretheSame:ADisaggregatedApproachtoIdentifyingAnti-AsianRacisminSocialMediaFanWuArizonaStateUniversityTempe,Arizona,USASanyamLakhanpalArizonaStateUniversityTempe,Arizona,USAQianLiArizonaStateUniversityTempe,Arizona,USAKookjinLee∗ArizonaStateUniversityTempe,Arizona,USAKookjin.Lee@asu.eduD...

展开>> 收起<<

Not All Asians are the Same A Disaggregated Approach to Identifying Anti-Asian Racism in Social Media.pdf

共15页,预览3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！

Not All Asians are the Same A Disaggregated Approach to Identifying Anti-Asian Racism in 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: