=Paper=
{{Paper
|id=Vol-2699/paper38
|storemode=property
|title=The Ebb and Flow of the COVID-19 Misinformation Themes
|pdfUrl=https://ceur-ws.org/Vol-2699/paper38.pdf
|volume=Vol-2699
|authors=Thomas Marcoux,Ester Mead,Nitin Agarwal
|dblpUrl=https://dblp.org/rec/conf/cikm/MarcouxMA20
}}
==The Ebb and Flow of the COVID-19 Misinformation Themes==
https://ceur-ws.org/Vol-2699/paper38.pdf
The Ebb and Flow of the COVID-19 Misinformation
Themes
Thomas Marcoux Esther Mead Nitin Agarwal
txmarcoux@ualr.edu elmead@ualr.edu nxagarwal@ualr.edu
University of Arkansas at Little Rock, Little Rock, AR 72204, USA
otal entity in determining how each nation responds
to the crisis. We have seen a variety of contradictory
Abstract statements on the national and international scene in-
fluencing opinions, in some cases politically polarizing
The COVID-19 pandemic has seen the
the issue of how to respond to the pandemic. But we
emergence of unique misinformation
have also seen cases of direct, physical - i.e. direct mail
narratives in various outlets, through
scams - attempts at preying on the uninformed or vul-
social media, blogs, etc. This on-
nerable such as personal protective equipment (PPE)
line misinformation has been proven to
marketing schemes. In both cases, it is obvious that
spread in a viral manner and has a di-
information has a very real impact on the lives and
rect impact on public safety. In an effort
livelihood of many. As such, we propose a study of
to improve public understanding, we cu-
the themes and chronological dynamics of the spread-
rated a corpus of 543 misinformation
ing of misinformation about COVID-19. Our corpus
pieces whittled down to 243 unique mis-
is a collection of unique misinformation stories1 man-
information narratives along with third
ually curated by our team. To highlight and visualize
party proofs debunking these stories.
these misinformation themes, we use topic modeling,
Building upon previous applications of
and introduce a tool to visualize the evolution of these
topic modeling to COVID-19 related
themes chronologically.
material, we developed a tool leveraging
topic modeling to create a chronological
visualization of these stories. From our 2 Literature Review
corpus of misinformation stories, this The information community has been tackling the is-
tool has shown to accurately represent sue of misinformation surrounding the COVID-19 pan-
the ground truth reported by our cu- demic since early in the outbreak. We base the claims
rator team. This highlights some of found in this paper on the findings that misinforma-
the misinformation narratives unique to tion spreads in a viral fashion and that consumers of
the COVID-19 pandemic and provides a misinformation tend to fail at recognizing it as such
quick method to monitor and assess mis- [Pen+20]. In addition to this, we believe this research
information diffusion, enabling policy- is essential as rampant misinformation constitutes a
makers to identify themes to focus on danger to public safety [Kou+20]. We also believe this
for communication campaigns. research is helpful in curbing misinformation since re-
searchers have found that simply recognizing the ex-
1 Introduction istence of misinformation and improving our under-
Following the discovery and subsequent spread of the standing of it can enhance the larger public’s ability to
COVID-19 pandemic, information has become one piv- recognize misinformation as such [Pen+20]. In order
to better understand the misinformation surrounding
Title of the Proceedings: “Proceedings of the CIKM 2020 Work- the pandemic, we look at previous research that has
shops, October 19-20, Galway, Ireland” Editors of the Proceed- leveraged topic models to understand online discus-
ings: Stefan Conrad, Ilaria Tiddi
sions surrounding this crisis. Research has shown the
2020 Copyright c for this paper by its authors: Use permitted
under Creative Commons License Atrribution 4.0 International 1 Stories can be explored at our official website
(CC BY 4.0). CEUR Workshop Proceedings (CEUR-WS.org) https://cosmos.ualr.edu/covid-19
benefits of using this technique to understand fluctu- ated by EUvsDisinfo in March of 2020 [EUv20]. EU-
ating Twitter narratives [Sha+20] over time, and also vsDisinfo’s database, however, was primarily focused
in understanding the significance of media outlets in on “pro-Kremlin disinformation efforts on the novel
health communications [Liu+20]. coronavirus”. Most of these items represented false
To implement topic modeling, we use the Latent narratives that were communicating political, mili-
Dirichlet Allocation model. Within the realm of natu- tary, and healthcare conspiracy theories in an at-
ral language processing (NLP), topic modeling is a sta- tempt to sow confusion, distrust, and public discord.
tistical technique designed to categorize a set of doc- Subsequently, misinformation stories were continually
uments within a number of abstract “topics”[BLS09]. gleaned from publicly available aggregators, such as
A “topic” is defined as a set of words outlining a gen- POLITIFACT2 , Truth or Fiction3 , FactCheck.org4 ,
eral underlying theme. For each document, which POLYGRAPH.info5 , Snopes6 , Full Fact7 , AP Fact
in this case, is an individual item of misinformation Check8 , Poynter9 , and Hoax-Slayer10 . The follow-
in our data set, a probability is assigned that desig- ing data points were collected or each misinforma-
nates its “belongingness” to a certain topic. In this tion item: title, summary, debunking date, debunking
study, we use the popular LDA topic model due to source, misinformation source(s), theme, and dissemi-
its widespread use and proved performances [BNJ03]. nation platform(s). The time period of our data set is
One point of debate within the topic modeling com- from January 22, 2020 to July 22, 2020. The data set
munity is the elimination of stop-words: i.e., should is comprised of 548 unique misinformation items. For
analysts filter common words from their corpus before many of the items, multiple platforms were used to
training a model. Following recent research claiming spread the misinformation. For example, oftentimes
that the use of custom stop-words adds little benefits a misinformation item will be posted on Facebook,
[SMM17], we followed the researchers’ recommenda- Twitter, YouTube, and as an article on a website. For
tion and removed common words after the model had our data set, the top-used platforms used for spread-
been trained. ing misinformation were websites, Facebook, Twitter,
Our model choice has seen use in previous research YouTube, and Instagram, respectively.
using LDA for short texts, specifically for short so-
cial media texts such as tweets [ZML17]. Some other 3.2 Topic Modeling
social media research using homogeneous social me-
In order to derive lexical meaning from this corpus, we
dia sources such as tweets or blog posts use associated
built a pipeline executing the following steps. First, we
hashtags to provide further context to topic models
processed each document in our text corpus. All that
[ARL17]. This is a promising lead to expend this re-
is needed is a text field identified by a date. Because
search towards big data social media corpora.
in most cases of word of mouth or social media it is
In this paper, we propose to leverage topic models
impossible to pinpoint the exact date the idea first
to understand the main underlying themes of misinfor-
emerged, we use the date of publication of the cor-
mation and their evolution over time using a manually
responding third party “debunk piece”. We trained
curated corpus of known fake narratives.
our LDA model using the Python tool Gensim11 us-
ing the methodology and pre-processing best practices
3 Methodology as described by its author [ŘS10] as well as best stop
This study uses a two-step methodology to produce words practices as described earlier [SMM17]. In this
relevant topic streams. First, through a manual cu- study, we found that generating 20 different topics
rating process, we aggregate different misinformation best matched the ground truth as reported by the re-
narratives for later processing. We consider misinfor- searchers curating the misinformation stories.
mation narratives, any narrative pushed through a va- Once the model was trained, we ordered the docu-
riety of outlets (social media, radio, physical mail, etc.) ments by date and created a numpy matrix where each
that has been or is later believably disproved by a third document is given a score for each topic produced by
party. This corpus constitutes our input data. Sec- the model. This score describes the probability that
ondly, we use this corpus to train an LDA topic model 2 https://www.politifact.com/coronavirus/
and to generate subsequent topic streams for analysis. 3 https://www.truthorfiction.com/
4 https://www.factcheck.org/
We describe these two steps in more details in the next
5 https://www.polygraph.info/
sections. 6 https://www.snopes.com/fact-check/
7 https://fullfact.org/health/coronavirus/#coronavirus
3.1 Collection of Misinformation Stories 8 https://apnews.com/APFactCheck
9 https://www.poynter.org/ifcn-covid-19-misinformation
Initially, the misinformation stories in our data set 10 https://www.hoax-slayer.net/category/covid-19/
were obtained from a publicly available database cre- 11 https://radimrehurek.com/gensim/
the given document is categorized as being part of a potential COVID-19 vaccine, and items promoting the
topic, i.e. if a score is high enough (here, a 10% prob- use of hydroxychloroquine. During the month of June,
ability), the document is considered part of the topic. the prominent theme shifted significantly to attempts
This allowed us to leverage the Python Pandas12 li- to convince citizens that face masks are either more
brary to plot a chronological graph for each individual harmful than not wearing one, and how to avoid rules
topic. We averaged topic distribution per day and used that required their use. Phishing scams also remained
a moving average window size of 20. This helped in prominent during June. During the month of July, the
highlighting the overarching patterns of the different dominant themes of the misinformation items shifted
narratives. The tool is publicly available and can be back to attempts to downplay the deadliness of the
found in the footnotes13 . novel coronavirus. Another prominent theme in July
were attempts to convince the public that COVID-19
4 Results testing is inflating the results.
In this section, we discuss the thoughts of our data
collection team and the ground truth as they were ob- 4.2 Topic Streams
served, and compare these with the results obtained
After using the tool described in 3.2, we generated the
through our topic modeling visualization tool.
graphs and tables described and discussed in this sec-
tion. Our data contains 243 unique misinformation
4.1 Prominent Misinformation Themes Over
narratives spanning from January 2020 to June 2020.
Time
The data was curated by our research team through
Although a variety of misinformation themes were the process described in the methodology. Each en-
identified, particularly dominant themes stood out, try contains, among other fields, a “date” used as a
changing over time. These themes were considered chronological identifier, a “title” describing the gen-
as dominant based on a simple sum of their frequency eral idea the misinformation is attempting to convey,
of occurrence in our data set. During the month of and a “theme” field putting the story in a concisely
March, the prominent misinformation theme was the described category. For example, a story given the ti-
promotion of remedies and techniques to supposedly tle “US Department of Defense has a secret biological
prevent, treat, or kill the novel coronavirus. Dur- laboratory in Georgia” is categorized in the following
ing the month of April, the prominent themes still theme: “Western countries are likely to be purpose-
included the promotion of remedies and techniques, ful creators of the new virus.” Each topic was repre-
but additional prominent themes began to stand out. sented by an identification number up to 20 and a set
For example, several misinformation stories attempted of 10 words. We picked the three most relevant words
to downplay the deadliness of the novel coronavirus. that best represented the general idea of each topic.
Others discussed the anti-malaria drug hydroxychloro- Notably, obvious words such as covid or coronavirus
quine. Others promoted the idea that the virus was a were removed from the topic descriptions since they
hoax meant to defeat President Donald Trump. Oth- are common for every topic.
ers consisted of various attempts to attribute false In Tables 1 and 2, we described some of the twenty
claims to high-profile people, such as politicians and topics found by each of our LDA models. These topics
representatives of health organizations. Also in April, were chosen because they each described a precise nar-
although first signs of these were seen in March, the rative and have a low topic distribution (or proportion
idea that 5G caused the novel coronavirus began to within the corpus). A low proportion is desirable be-
become more prevalent. During the month of May, cause this indicates the detection of a unique narrative
the prominent themes shifted to predominantly false within the corpus; as opposed to an overarching topic
claims made by high-profile people, followed by at- including general words such as “world”, “outbreak”,
tempts to convince citizens that face masks are ei- or “pandemic”. Do note that topic inclusiveness is
ther more harmful than not wearing one, or are in- not exclusive and documents can be part of multiple
effective at preventing COVID-19, and how to avoid topics.
rules that required their use. The number and variety This becomes apparent in the tables below: from
of identity theft phishing scams also increased during our topic model, we found a dominant topic encom-
May. Misinformation items attempting to attribute passing 68% of narratives. It includes words such as
false claims to high-profile people continued through- “Trump”, “outbreak”, “president”, etc. Some other
out May. Also becoming prominent in May were mis- narratives also included words such as “flu”, “news”,
information items attempting to spread fear about a or “fake”. Because the evolution of these narratives are
12 https://pandas.pydata.org/ consistent across the corpus and show little temporal
13 https://github.com/thomas-marcoux/TopicStreamsTools fluctuation, we chose not to report on them further.
For these reasons, the narratives we focused on below
show a low percentage of distribution.
Table 1: Most frequent dominant topics from titles.
ID Word 1 Word 2 Word 3 Proportion
10 china chinese spread 2%
12 scam hydroxy... health 2%
17 state donald trump 2%
18 vaccine gates bill 5%
Table 2: Most frequent dominant topics from themes.
Figure 1: Topic distribution of titles for topic 10 (key-
ID Word 1 Word 2 Word 3 Proportion words: china, chinese, spread)
3 fear spread western 2%
9 predicted pandemic vaccine 2% online narratives that focused on the provenance of the
16 phishing hydroxy... vaccine 2% virus during the early stages.
Figure 2 shows the evolution of Topic 12, the topic
describing narratives related to health, home reme-
4.2.1 Using narrative titles as a corpus dies, and general hoaxes and scams stemming from
the panic. We can see it was consistent with the rise
The general narratives described by the topics were
of cases in the United States and panic increased as
thus:
with the spread of the virus. It is interesting to note
that this figure roughly coincides with the daily num-
• Topic 10 described the narratives related to the
ber of confirmed cases for this time period [Rit+20].
Chinese government and its responsibility in the
spread of the virus. These stories represented an
estimated 2% of the 243 stories collected.
• Topic 12 described the narratives related to per-
sonal health and scams or misinformation such as
the benefits of hydroxychloroquine. These stories
represented an estimated 2% of the 243 stories
collected.
• Topic 17 described the narratives related to the re-
sponse of Donald Trump and his administration.
These stories represented an estimated 2% of the
243 stories collected.
• Topic 18 described the narratives related to the
involvement of Bill Gates in various conspiracies,
mostly linked to vaccines. These stories repre- Figure 2: Topic distribution of titles for topic 12 (key-
sented an estimated 4% of the 243 stories col- words: hydroxychloroquine, health, scam)
lected.
Figure 1 shows the evolution of Topic 10, the topic Figure 3 shows the evolution of Topic 17. This topic
describing China-related narratives. It shows that described stories related to Donald Trump and his ad-
these narratives were already in full force from the be- ministration. These stories generally referred to claims
ginning of our corpus and slowly came to a near halt that the virus was manufactured as a political strat-
during the month of April. We notice a short spike egy, or claims that various public figures were speaking
again towards the end of the corpus during the month out against the response of the Trump administration.
of June. This is consistent with the ground truth of Figure 4 shows the evolution of Topic 18. This
68% of narratives as well. This time including words
such as “attempt”, “countries”, and “purposeful”. As
for section 4.2.1, we chose not to report on that topic
as well as other smaller but general topics showing
little fluctuation. Therefore, the narratives we focused
on below show a low percentage of distribution. The
general narratives described by the topics are thus:
• Topic 3 described the narratives related to the
speculations on the spread of the virus, especially
in an international relations context. These sto-
ries represented an estimated 2% of the 243 stories
collected.
• Topic 9 described the narratives related to sto-
ries claiming the creation and propagation of the
Figure 3: Topic distribution of titles for topic 17 (key- virus were either designed or predicted, along with
words: donald, trump, state) voices claiming a vaccine already exists. These
stories represented an estimated 3% of the 243
topic described stories such as Bill Gates and his stories collected.
perceived involvement with an hypothetical vaccine,
• Topic 16 described the narratives related to per-
and other theories describing the virus’ appearance
sonal health and scams or misinformation such as
and spread as an orchestrated effort. As with Figure
the benefits of hydroxychloroquine. These stories
1, these narratives were especially strong early on
represented an estimated 2% of the 243 stories
(albeit this narrative remained active for a slightly
collected.
longer time), before coming to a near halt.
Figure 5 shows the evolution of Topic 3. It is linked
We notice that as theories about the origins of the to early fear of the virus and presented narratives as
virus slowed down, hoaxes and scams on personal pro- opposing the western block with the East, notably
tection increased as shown on Figure 2. China. It matched closely with Figure 1 and its China-
related narratives. In both cases, we see an early dom-
inance of the topic followed by a near halt as the virus
touched the United States.
Figure 4: Topic distribution of titles for topic 18 (key-
words: bill, gates, vaccine)
Figure 5: Topic distribution of themes for topic 3 (key-
4.2.2 Using narrative themes as a corpus words: fear, spread, western)
For this section, we inputted narrative themes as the
corpus. Note that the topic IDs are independent from Figure 6 describes the evolution of narratives claim-
the previous set of topics using titles. Similarly to ing the virus was predicted or even designed. This
section 4.2.1, we found a dominant topic encompassing figure is consistent with the results shown by Figure
4 which shows claims regarding Bill Gates, early vac- We have shown the potential of using topic modeling
cines, etc. They both showed stories of early knowl- visualization to get a bird’s eye view of the fluctuating
edge of the virus and peaked early, appearing more or narratives and an ability to quickly gain a better un-
less sporadically as time goes on and as cases increased. derstanding of the evolution of individual stories. We
have seen that the tool is efficient to chronologically
represent actual narratives pushed to various outlets,
as confirmed by the ground truth observed by our mis-
information curating team. This work illustrates a rel-
atively quick technique for allowing policy makers to
monitor and assess the diffusion of misinformation on
online social networks in real-time, which will enable
them to take a proactive approach in crafting impor-
tant theme-based communication campaigns to their
respective citizen constituents.
We have also seen in this study that using carefully
curated “themes” - which offer a lexical value close to
the abstract topics provided by the LDA model - yields
similar results to using misinformation narratives “ti-
tle”. This paves the way for scaling this method with
much larger corpora such as a set of news headlines,
Figure 6: Topic distribution of themes for topic 9 (key- blog titles, or social media posts.
words: predicted, pandemic, vaccine) LDA is generally viewed as more reliable due to the
control one can have over the number of topics. Find-
Figure 7 is parallel to Figure 2. Both showed hoax ing an optimal level of granularity through trial and er-
stories promoting scams and health-related misinfor- ror tends to perform well when tailored to the use-case.
mation. We noticed an early rise on Figure 7, most Because the LDA topic model may become difficult to
likely due to the inclusion of the keyword “vaccines” scale, however, we consider using the HDP (Hierarchi-
in the topic, which caused some overlap with Topic 9 cal Dirichlet Process) model for future works involving
as shown in Figure 6. multiple larger corpora. This model attempts to infer
the number of topics computationally, which may be-
come more scalable on large sets of documents with an
unknown number of topics.
Acknowledgements
This research is funded in part by the U.S. National
Science Foundation (OIA-1946391, OIA-1920920,
IIS-1636933, ACI-1429160, and IIS-1110868), U.S.
Office of Naval Research (N00014-10-1-0091, N00014-
14-1-0489, N00014-15-P-1187, N00014-16-1-2016,
N00014-16-1-2412, N00014-17-1-2675, N00014-17-1-
2605, N68335-19-C-0359, N00014-19-1-2336, N68335-
20-C-0540), U.S. Air Force Research Lab, U.S. Army
Research Office (W911NF-17-S-0002, W911NF-16-
1-0189), U.S. Defense Advanced Research Projects
Agency (W31P4Q-17-C-0059), Arkansas Research
Figure 7: Topic distribution of themes for topic 16 Alliance, the Jerry L. Maulden/Entergy Endowment
(keywords: hydroxychloroquine, vaccine, phishing) at the University of Arkansas at Little Rock, and the
Australian Department of Defense Strategic Policy
Grants Program (SPGP) (award number: 2020-106-
5 Conclusion 094). Any opinions, findings, and conclusions or
recommendations expressed in this material are those
This study has highlighted some of the narratives that of the authors and do not necessarily reflect the
surfaced during the COVID-19 pandemic. We col- views of the funding organizations. The researchers
lected 243 unique misinformation narratives over six gratefully acknowledge the support.
months and proposed a tool to observe their evolution.
References 19118. url: http://www.ncbi.nlm.nih.
gov/pubmed/32302966.
[BNJ03] David M. Blei, Andrew Y. Ng, and Michael
I. Jordan. “Latent dirichlet allocation”. In: [Pen+20] Gordon Pennycook et al. “Fighting
Journal of Machine Learning Research 3 COVID-19 Misinformation on Social
(2003), pp. 993–1022. Media: Experimental Evidence for a Scal-
able Accuracy-Nudge Intervention”. In:
[BLS09] David M. Blei, John D. Lafferty, and
Psychological Science 31.7 (2020). eprint:
Ashok N. Srivastava. Text Mining: Clas-
https://doi.org/10.1177/0956797620939054,
sification, Clustering, and Applications.
pp. 770–780. doi: 10 . 1177 /
CRC Press, 2009, pp. 71–88.
0956797620939054. url: https :
[ŘS10] Radim Řehůřek and Petr Sojka. “Software //doi.org/10.1177/0956797620939054.
Framework for Topic Modelling with Large
[Rit+20] Hannah Ritchie et al. United States: Coro-
Corpora”. In: May 2010, pp. 45–50. doi:
navirus Pandemic - Our World in Data.
10.13140/2.1.2393.1847.
2020. url: https : / / ourworldindata .
[ARL17] Md. Hijbul Alam, Woo-Jong Ryu, and org / coronavirus / country / united -
SangKeun Lee. “Hashtag-based topic evo- states ? country = ~USA. (accessed:
lution in social media”. In: World Wide 07.29.2020).
Web 20.6 (Nov. 2017), pp. 1527–1549.
[Sha+20] Hao Sha et al. Dynamic topic model-
issn: 1573-1413. doi: 10 . 1007 / s11280 -
ing of the COVID-19 Twitter narrative
017-0451-3. url: https://doi.org/10.
among U.S. governors and cabinet execu-
1007/s11280-017-0451-3.
tives. 2020. arXiv: 2004.11692 [cs.SI].
[SMM17] A. Schofield, M. Magnusson, and D.
Mimno. “Pulling Out the Stops: Rethink-
ing Stopword Removal for Topic Models”.
In: 15th Conference of the European Chap-
ter of the Association for Computational
Linguistics. Vol. 2. Association for Com-
putational Linguistics. 2017, pp. 432–436.
[ZML17] Y. Zhang, W. Mao, and J. Lin. “Model-
ing Topic Evolution in Social Media Short
Texts”. In: 2017 IEEE International Con-
ference on Big Knowledge (ICBK). 2017,
pp. 315–319.
[EUv20] EUvsDisinfo. EUvsDisinfo. March 16,
2020. The Kremlin and Disinformation
About Coronavirus. 2020. url: https://
euvsdisinfo . eu / the - kremlin - and -
disinformation - about - coronavirus/.
(accessed: 07.16.2020).
[Kou+20] Ramez Kouzy et al. “Coronavirus Goes
Viral: Quantifying the COVID-19 Mis-
information Epidemic on Twitter”. eng.
In: Cureus 12.3 (Mar. 2020). Publisher:
Cureus, e7255–e7255. issn: 2168-8184.
doi: 10.7759/cureus.7255. url: https:
//pubmed.ncbi.nlm.nih.gov/32292669.
[Liu+20] Qian Liu et al. “Health Communica-
tion Through News Media During the
Early Stage of the COVID-19 Outbreak in
China: Digital Topic Modeling Approach”.
In: J Med Internet Res 22.4 (Apr. 2020),
e19118. issn: 1438-8871. doi: 10 . 2196 /