Understanding COVID-19 Vaccine Campaign on Facebook using Minimal Supervision Tunazzina Islam

2025-05-06 0 0 8.56MB 11 页 10玖币
侵权投诉
Understanding COVID-19 Vaccine Campaign on
Facebook using Minimal Supervision
Tunazzina Islam
Department of Computer Science
Purdue University, West Lafayette, IN 47907, USA
islam32@purdue.edu
Dan Goldwasser
Department of Computer Science
Purdue University, West Lafayette, IN 47907, USA
dgoldwas@purdue.edu
Abstract—In the age of social media, where billions of internet
users share information and opinions, the negative impact of
pandemics is not limited to the physical world. It provokes a surge
of incomplete, biased, and incorrect information, also known
as an infodemic. This global infodemic jeopardizes measures to
control the pandemic by creating panic, vaccine hesitancy, and
fragmented social response. Platforms like Facebook allow adver-
tisers to adapt their messaging to target different demographics
and help alleviate or exacerbate the infodemic problem depending
on their content. In this paper, we propose a minimally supervised
multi-task learning framework for understanding messaging
on Facebook related to the COVID vaccine by identifying ad
themes and moral foundations. Furthermore, we perform a
more nuanced thematic analysis of messaging tactics of vaccine
campaigns on social media so that policymakers can make better
decisions on pandemic control.
Index Terms—COVID-19 vaccine, social media, facebook ads,
minimal supervision, weak labeling.
I. INTRODUCTION
Since January 2020, worldwide public health has been
threatened by the novel coronavirus – the outbreak was declared
a global pandemic by the World Health Organization (WHO)
[1]. COVID-19 is the first pandemic in the history in which
technology and social media are being used on a massive scale
to keep people safe, informed, productive, and connected. Yet,
at the same time, the growing proliferation of social media can
be used for spreading hoaxes and false information leading to
what is commonly referred to as an infodemic [2]. Social
Fig. 1: Example of an ad highlighting analysis dimensions (1)
theme, (2) moral foundation.
media discourse help increase polarization around topics related
to COVID-19 vaccines, such as the vaccine mandate, natural
immunity, vaccine efficacy, religious sentiment, and vaccine
equity. Moral Foundation Theory (MFT) [4], [5] suggests a
theoretical framework for analyzing the morality, containing
six basic moral foundations (Table. I). Past work has shown
that the theory can help explain ideological differences and
*networkforphl.org
CARE/HARM:
Saying that someone other than the speaker deserves care or gets harmed.
Reflects the base of Maslow’s Hierarchy of Needs [3]. Security, Shelter, Food, Water,
Warmth.
FAIRNESS/CHEATING:
Justice, rights, and autonomy, comparison to other groups.
Equality of Opportunities. Social Intolerance to “Free-Rider".
LOYALTY/BETRAYAL:
Patriotism and self-sacrifice for the group (Or failure to provide
it, in the case of betrayal). It is active anytime people feel that it’s “one for all, and all
for one.”
AUTHORITY/SUBVERSION:
Deference (or opposition) to legitimate authority and
respect for traditions. Social order and the obligations of hierarchical relationships,
such as obedience, respect, and the fulfillment of role-based duties.
SANCTITY/DEGRADATION:
Not simply a religious value. Respect for the human
spirit. Social aversion of personal degradation. Degradation causes disgust. Physical and
spiritual contagion, including virtues of chastity, wholesomeness, and control of desires.
It underlies the widespread idea that the body is a temple which can be desecrated by
immoral activities and contaminants (an idea not unique to religious traditions).
LIBERTY/OPPRESSION:
Feelings of people towards those who dominate them and
restrict their liberty. The hatred of bullies and dominators motivates people to come
together, in solidarity, to oppose or take down the oppressor.
NONE:Does not fall under any other foundation.
TABLE I: Six basic moral foundations.*
social group membership [6]–[12]. Often there is a significant
correlation between the vaccine debate and its moral foundation
(MF).
However, sponsored contents have been used to reach
more people on social media to disseminate their agendas.
For example, Fig. 1 presents a sponsored ad on Facebook
representing ‘vaccine mandate’ as the ad theme and over-
reach of power and takes away the right, falling under the
‘liberty/oppression’ moral foundation. Therefore, detecting
moral foundation and theme from text are vital components
in understanding advertisers’ intention, key talking points,
policies.
Our goal in this paper is to take a first step toward analyzing
the landscape of vaccination campaigns on social media.
We focus our experiments on a timely topic, COVID-19
vaccination campaign. Our main contributions are twofold: (1)
to identify the ad theme & moral foundation; (2) to build on
this characterization to analyze the messaging across different
demographics, geographic, and timelines. We analyze over
28K
COVID vaccine related ads on Facebook, associating ads
with
6
moral foundations and including ‘none’, it’s a
7
-class
classification problem. We also identify the theme of the ads,
a15-class classification problem.
Our theme analysis is motivated by previous studies [13]–
[15] that created code-book of COVID-19 vaccine arguments.
Besides, our moral foundation analysis is inspired by social
science studies [16]–[18] that demonstrated relation between
arXiv:2210.10031v2 [cs.CL] 16 Nov 2022
Moral Foundations Example of messages
Care/Harm Protect yourself and others.
Help those most vulnerable.
Public health can assist you.
Stay healthy and safe.
Fairness/Cheating Everyone has an interest in beating this outbreak.
Infection does not discriminate.
We have an interest in everyone getting appropriate care.
Vaccine should be free for everyone.
Loyalty/Betrayal Do your part, take the shot for your family, friends, country.
We need to protect our community.
I’m loyal to you and want to keep you safe.
Limited resources should go first to healthcare workers and those caring for us.
Authority/Subversion Scientific evidence and common sense show that protective measures really work.
Listen to your local public health official.
Respect healthcare workers and the risks they are taking.
Trust science.
Be a good role model for others.
Sanctity/Degradation Be willing to sacrifice your wants for community needs.
Help nurture the spirits of those needing comfort.
Look for ways to serve others.
Liberty/Oppression COVID can threaten our safety and freedom.
We want our community to be free from fear of contagion.
The quicker we beat this, the quicker we recover and return to normal.
TABLE II: Example messages corresponding to each moral
foundation provided to annotators.
moral foundation and COVID related health decisions.
In this paper, we suggest a minimally supervised multi-
task learning approach to understand COVID-19 vaccine
campaign in Facebook. The purpose of minimal supervision
is to compensate for the lack of annotated data by exploiting
the maximum potential of the available data. For MF, we
generate weak labels from dedicated lexicons developed for
identifying moral foundation. For theme, we use a pre-trained
textual inference model to identify paraphrases in a large
collection of COVID-19 vaccination ads from Facebook and
assign theme based on cluster assignment (Details in
IV-A
).
We focus on the following research questions (RQ) to analyze
vaccine campaigns on social media:
•RQ1.
What are the narratives of the messaging? (section
V-C)
•RQ2.
How does entity type fulfill messaging roles?
(section V-D)
•RQ3.
Which demographics and geographic are reached
by the advertisers and their messages? (section V-E)
•RQ4.
Do ads follow current COVID status? (section
V-F
)
We summarize the main contributions of this paper as the
following:
1)
We formulate a novel problem of using minimal supervi-
sion to analyze the landscape of vaccine campaigns on
Facebook. Our dataset is publicly available here†.
2)
We suggest a minimally supervised multi-task learning
framework with three different learning strategies to
identify ad theme and moral foundation.
3)
We investigate the COVID vaccine ads on Facebook from
four angles: narratives (thematic and moral foundation
analysis), entity types (who is funding the ad), reach
(who saw the ads), and whether the ads reflect current
COVID situations.
II. RELATED WORK
Recent studies have shown narrative analysis and opinion
mining of COVID-19 pandemic discourse in social media
†https://github.com/tunazislam/Covid_FB_AD_MinimalSup
Themes Definition
EncourageVaccination Promoting vaccination to control pandemic.
VaccineMandate Arguments about vaccine mandate, vaccine passport/card.
VaccineEquity Acknowledging no nation, state, or individual’s life
is more important or more deserving than another’s.
VaccineEfficacy Arguments saying that the vaccine is safe, lessens the symptoms.
GovDistrust Arguments saying people do not have trust on Governmental
institutions or authority figures.
GovTrust Arguments saying people have trust on Governmental
institutions or authority figures.
VaccineRollout Information about vaccination sites and availability of appointments.
VaccineSymptom Symptoms associated with the vaccine, e.g., fever, sore arm etc.
VaccineStatus Information regarding rate of vaccination, hospitalization, death etc.
VaccineReligion Arguments about religion and vaccine.
VaccineDevelopment Broadcasting information about the vaccine development and approval.
CovidPlan Good policies to deal with COVID-19.
VaccineMisinformation Conspiracy theories, fake news related to vaccine.
NaturalImmunity Natural methods of protection against COVID.
Vote Encourage residents to vote by iterating messages related to COVID vaccine.
TABLE III: Theme definition provided to the annotators.
and news media [19]–[25]. Also, there are recent studies
on online perceptions about COVID-19 vaccination related
to public health measures [26]–[28] and moral foundations
[29]–[33]. Nowadays, targeted online advertising is one of the
main communication channels, allowing hyper-local sponsors
to campaign during the pandemic. Sponsored content on social
media can be shared with various narratives, including infor-
mation and misinformation, to disseminate agendas targeting
specific demographics and geographic. Mejova and Kalimeri
[34] analyzed a smaller set of COVID-19 related Facebook ads
messaging by identifying advertisers and their targets. Silva
and Benevenuto [35] monitored COVID related Facebook Ads
in Brazil to identify misinformation. Our work takes a different
approach to analyze COVID vaccine related Facebook ads by
identifying themes and moral foundation that motivate sponsors.
Our work falls under the broad scope of weak supervision [36]–
[39] and multi-task learning [40]–[44].
III. DATASET DETAILS
We collected approximately
28,000
COVID vaccine related
ads focusing on United States from December 2020 - January
2022 using Facebook Ad Library API
‡
with the search
term ‘COVID-19 vaccine’, ‘COVID vaccine’, ‘vaccination’,
‘vaccine’, ‘coronavirus vaccine’, ‘corona vaccine’. Our collected
ads were written in English. For each ad, the API provides
ad ID, title, ad body, funding entity, spend, impressions,
distribution over impressions broken down by gender (male,
female, unknown), age (
7
groups), location down to states in
the USA. We have duplicate content among those collected
ads because the same ad has been targeted to different regions
and demographics with unique ad id. We have
9,920
ads with
different contents.
A. Data Annotation
We manually annotated
557
ads for themes and moral
foundation. To ensure quality work, we provided annotators
with
23
examples covering all six moral foundations (Table. II)
and theme definition of
15
themes (Table III). Two annotators
from the Computer Science department manually annotated
a subset of ads (
20%
) to calculate inter-annotator agreement
‡https://www.facebook.com/ads/library/api
using Cohen’s Kappa coefficient [45]. This subset has inter-
annotator agreements of
73.80%
for MF and
65.60%
for theme
which are substantial agreements. In case of a disagreement, we
resolved it by discussion. Rest
80%
of the data was annotated
by one of the graduate students between the two. We had one
female and one male annotator, and the age range was
30 −40
.
IV. METHODOLOGY
In this section, we start by describing the labeling technique
that produces numerous but imprecise (weak) labels. Then, we
put forward two learning strategies to better exploit the available
labels. Finally, we show the main components of multi-task
learning model. An overview of the model is illustrated in Fig.
2.
Fig. 2: An overview of our proposed framework. Best viewed
in electronic format (zoomed in).
A. Weak Label Generation
In this section, we describe how to generate weak labels
from ad content that can be incorporated as weak sources in
our model.
1) Themes: First, we go through relevant research conducted
by the health informatics, computational social science, public
health, psychology communities and ground our analysis
through constructing a list of potential themes for COVID
related ads [15], [46]–[49]. Then, we consult with two
researchers in Computational Social Science and finalize the
relevant themes with corresponding phrases. The full list of
phrases for each theme can be observed in Table IV. To
generate the weak labels for themes, we ground the phrases
(from corresponding theme) in a set of
28k
unlabelled COVID-
19 vaccine related ads and match similarity between their
Sentence BERT embeddings [50]. We measure the cluster
purity using silhoutte score [51]. We use threshold based on
closest distance to limit assignments. Bar plot with the number
of assigned ads to each cluster with and without threshold and
2D
visualizations of clusters are shown using t-SNE [52] in
Fig. 3. We use threshold ≤0.5resulting 21,851 ads.
2) Moral Foundation: Weak label for moral foundation is
generated by analyzing text based on the MFT relying on the
use of a lexical resource, the Moral Foundations Dictionary
(MFD) [6]. Similar to Linguistic Inquiry and Word Count
(LIWC) [53], [54], MFD associates a list of related words
with each one of the moral foundations. We analyze ad’s text
by counting the number of occurrences of words in the text
(a) Without threshold. (b) threshold ≤0.5
(c) threshold ≤0.4(d) t-SNE without threshold.
(e) t-SNE (threshold ≤0.5) (f) t-SNE (threshold ≤0.4)
Fig. 3: Bar plot and 2D visualization of cluster assignment for
themes without and with threshold. Best viewed in electronic
format (zoomed in).
which also match the words in the MFD. In this process, same
ad might get multiple moral foundations based on lexicon
matching. To assign one MF for each ad, we pick the MF
having the highest number of keyword matching with our text.
Given that MFD does not have lexicon for liberty/oppression
moral foundation, we use the same lexicon curated by Pacheco
et al. [15]. We annotate an ad as liberty/oppression MF if it
contains at least two keywords.
3) Quality of Weak Label: To assess the weak label quality,
we compare the weak labels with the ground truth labels (
557
ads). The accuracy and macro-avg F1 score of the weak label for
theme are
0.513
and
0.337
respectively. For moral foundation,
the accuracy and macro-avg F1 score are
0.417
and
0.248
correspondingly. We observe that the accuracy and macro-avg
F1 score of the weak label are significantly better than random
(
0.067
) for theme and comparatively better than random for
moral foundation (
0.143
), indicating that our noisy labeling
approach has acceptable quality.
B. Learning Strategies
We devise two learning strategies so that model can have
access to larger datasets, which may benefit its generalization
capabilities at both tasks.
1) Hybrid Learning: To avoid the risk of highly biased
model, instead of using fully weakly supervised batches to
train the multi-task model, we create mixed batches with part
gold, part noisy labels.
2) Two-stage Learning: We separate the learning process
into two stages, i.e., (1) pre-training stage using a large but
noisy dataset, (2) fine-tuning stage using gold labels. We use
a transfer learning technique by freezing the hidden layers
except for the task-specific layers of our model. Therefore, we
摘要：

UnderstandingCOVID-19VaccineCampaignonFacebookusingMinimalSupervisionTunazzinaIslamDepartmentofComputerSciencePurdueUniversity,WestLafayette,IN47907,USAislam32@purdue.eduDanGoldwasserDepartmentofComputerSciencePurdueUniversity,WestLafayette,IN47907,USAdgoldwas@purdue.eduAbstract—Intheageofsocialmedi...

展开>> 收起<<
Understanding COVID-19 Vaccine Campaign on Facebook using Minimal Supervision Tunazzina Islam.pdf

共11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！

相关推荐

更多
立即下载
分类：图书资源 价格：10玖币 属性：11 页 大小：8.56MB 格式：PDF 时间：2025-05-06

开通VIP享超值会员特权

  • 多端同步记录
  • 高速下载文档
  • 免费文档工具
  • 分享文档赚钱
  • 每日登录抽奖
  • 优质衍生服务

作者详情

相关内容

更多

热门标签

人际关系 配电装置 动力学连接体 力的合成 高考理综 全宋诗作者索引 公务员考试
/ 11
评分 收藏
分享
立即下载
客服
关注