PaperlayerView exact revision on arXiv

Source-linked AI summary

Fake news agenda in the era of COVID-19: Identifying trends through fact-checking content

Wilson Ceron, Mathias-Felipe de-Lima-Santos, Marcos G. Quiles

arXiv:2012.11004v1cs.SIcs.CY

TL;DR

Research on fake-news topic trends during crises and computational ways to identify them remains limited. The paper applies a time-bounded Markov-inspired topic method to tweets from two Brazilian fact-checking outlets. Their debunks showed similarities and differences, while political and pandemic agendas were intertwined.

  • Problem

    Evidence on fake-news topic trends during crises and computational methods to identify them remains limited.

    Source: §3 · p. 9

  • Method

    The authors use a Markov-inspired method to cluster topics over defined time spans in tweets from two major Brazilian fact-checking initiatives.

    Source: §1 · p. 3, §3.3 · p. 13

  • Results

    The two organizations’ debunks showed similarities and differences, few topics lasted beyond one fortnight, and politics and the pandemic were intertwined.

    Source: §5 · p. 24

  • Takeaways & Limitations

    The method may help news organizations identify topics rising or falling in social-media popularity and inform agenda-setting and cooperation among fact-checkers.

    Source: §5 · p. 24

    The analysis is limited to two major Brazilian fact-checking organizations because other initiatives did not produce enough tweets for inclusion.

    Source: §5 · p. 24

Abstract

from arXiv · show

The rise of social media has ignited an unprecedented circulation of false information in our society. It is even more evident in times of crises, such as the COVID-19 pandemic. Fact-checking efforts have expanded greatly and have been touted as among the most promising solutions to fake news, especially in times like these. Several studies have reported the development of fact-checking organizations in Western societies, albeit little attention has been given to the Global South. Here, to fill this gap, we introduce a novel Markov-inspired computational method for identifying topics in tweets. In contrast to other topic modeling approaches, our method clusters topics and their current evolution in a predefined time window. Through these, we collected data from Twitter accounts of two Brazilian fact-checking outlets and presented the topics debunked by these initiatives in fortnights throughout the pandemic. By comparing these organizations, we could identify similarities and differences in what was shared by them. Our method resulted in an important technique to cluster topics in a wide range of scenarios, including an infodemic -- a period overabundance of the same information. In particular, the data clearly revealed a complex intertwining between politics and the health crisis during this period. We conclude by proposing a generic model which, in our opinion, is suitable for topic modeling and an agenda for future research.

1. Introduction

The introduction motivates studying fact-checking during the COVID-19 infodemic, especially in the underexamined Global South, and proposes a time-bounded topic model to compare Brazilian outlets’ debunks.

1. Introduction

  • The introduction distinguishes misinformation from intentional disinformation and describes the COVID-19 crisis as an infodemic in which false information can confuse people and hamper public-health responses.It illustrates the consequences of online rumors with the death of Fabiane Maria de Jesus, whose killing followed a false claim shared on Facebook [1].Source: §1 · p. 2
  • Fact-checking is presented as a response to harmful falsehoods, but accusations of political bias and uncertainty about what initiatives assess motivate examining their coverage.The introduction also notes that fact-checks can reveal the kinds of claims circulating on social networks.Source: §1 · p. 2
  • The authors identify limited research on fact-checking organizations in the Global South and focus on Brazil amid political instability and the pandemic.They situate the study in a period when Brazil’s political context made the country particularly susceptible to fake-news campaigns.Source: §1 · pp. 2–3
  • The study proposes a time-bounded topic-clustering method and uses it to examine tweets from two major Brazilian fact-checking initiatives during the first half of 2020.Its stated contributions are a method for clustering topics over an established time span and an analysis of topics that emerged during the infodemic.Source: §1 · p. 3
  • The analysis asks which topics the initiatives covered, how their topics related, how politics intertwined with the pandemic, and what makes topic mapping difficult during information overload.The authors frame the work as examining debunked topics, political and pandemic agendas, and repeated themes across the two initiatives.Source: §1 · p. 3

2. Related Work

Related work reviews debates about fact-checking, computational approaches to Twitter, and Brazil’s COVID-19 information environment, motivating analysis of topic trends using two Brazilian outlets.

2.1. Fact-checking: a task force to fight fake news

  • Fact-checking’s efficacy is contested: political interests and prior beliefs may limit receptiveness, while corrections can fail to reduce or may strengthen misperceptions [19, 20, 21].Different initiatives also cover different topics, reinforcing perceptions of bias [22, 23].Source: §2.1 · pp. 4–5
  • Limited resources leave fact-checkers behind misinformation in reach and response time, with an approximately 13-hour delay between consumption and verification [24].Fact-checking also peaks around political events, while false information on Twitter is shared faster than on other platforms, especially political topics [25] [26, 27].Source: §2.1 · p. 5

2.2. Computational Social Science and Twitter

  • Computational social-science research uses Twitter and methods such as sentiment classification, trend analysis, and topic modeling, but LDA may perform poorly on short tweets [42, 43].Other work has aggregated tweets or combined LDA with the Author-Topic Model [45], while PCA and statistical relational learning have also been used to detect false information [50] [51].Source: §2.2 · pp. 6–7
  • Despite prior methods, the evolution of fake-news agendas over time remains underexamined, and combining computational methods with human analysis has been proposed for social-media research [52] [54].The literature also raises concerns about the accuracy and transparency of automated content moderation [53].Source: §2.2 · p. 7

2.3. From a health crisis to a political firestorm: How Brazil is handling the outbreak

  • COVID-19 uncertainty, rumors, and political claims created a challenging information environment, while Brazil’s political context heightened the importance of fact-checking [55, 56] [57, 58].The Brazilian initiatives Agência Lupa and Aos Fatos are described as major producers of fact-checking stories committed to IFCN transparency standards [63, 64] [65].Source: §2.3 · pp. 7–8
  • The study analyzes Lupa and Aos Fatos because they provide public, transparent data, using a Markov-inspired model to examine Twitter topic trends during the pandemic.The authors note that smaller initiatives lacked enough tweets for inclusion in this analysis.Source: §2.3 · p. 9

3. Data and methods

The study addresses limited evidence on crisis-period fake-news trends and topic-modeling methods by analyzing tweets from two prominent Brazilian fact-checking organizations. It proposes a Markov-inspired approach that clusters topics within fortnightly windows to track their evolution.

3.1. Data description

  • The analysis covers 5,115 tweets from @aosfatos and @agencialupa from January to July 2020; smaller initiatives were excluded because they produced too few tweets.The authors describe these as the most prominent Brazilian fact-checking organizations in the study.Source: §3.1 · p. 9

3.2. Data preparation

  • The method clusters topics fortnight by fortnight, addressing the time-independence of conventional models and limitations of temporal alternatives such as variational-inference complexity or requiring a preset topic count.The authors chose fortnightly intervals as a unit they considered suitable for forming communities.Source: §3.2 · p. 10
  • Preprocessing aggregates near-simultaneous thread tweets, removes stop words and punctuation, and excludes retweets, mentions, hashtags, and common terms; Portuguese stemming and lemmatization are omitted.The authors cite Portuguese-language complexity and the risk of losing tweet meaning as reasons for omitting stemming and lemmatization.Source: §3.2 · pp. 10–11

3.3. Markov chains and topic modeling

  • Tweets were represented as directed, weighted graphs of words and their connections, preserving word order; repeated connections increase edge weights.The graph is built from tweets within a time window, and the authors use edge clustering and community detection to identify topics.Source: §3.3 · pp. 11–13
  • Walktrap community detection identifies closely linked word clusters, while the analysis annotates the five largest communities per fortnight and visualizes prominent vertices in word clouds.The word clouds emphasize more frequently used vertices and filter for vertices with more than 10 degrees.Source: §3.3 · pp. 13–14
  • The method identified 23 codes, including Federal Government, Politics, Health, and COVID-19 subtopics; some clusters also formed around English or Spanish retweets.The authors note that words in the same language connect more readily, while rigid clustering can produce a thread or single-tweet cluster.Source: §3.3 · pp. 14–15

4. Results

The results trace changing topics debunked by Brazilian fact-checkers during the pandemic, including overlap between outlets and a growing intertwining of health and political agendas. They also discuss challenges of modeling topics during an infodemic.

4.1. Fact-checking shows the evolution of false information

  • Fact-checkers debunked claims about virus transmission, prevention remedies, death registration, and political and institutional issues.The topics included claims that pets could transmit COVID-19, purported preventive foods, and death over-notification.Source: §4.1 · pp. 15, 18
  • From fortnight 4, when Brazil reported its first case, pandemic discussion dominated; prevention and dissemination claims recurred, alongside economy, politics, and Federal Government topics.Fortnight 4 was an exception for leisure claims during Carnival; earlier, both outlets shared topics in fortnights 1 and 2 but not 3.Source: §4.1 · p. 19
  • Only Leisure showed a strong correlation with the time series; Economy had the next strongest correlations, and the outlets’ coverage overlapped while otherwise selecting different topics.The authors associate Leisure with Carnival and Economy with the financial crisis and controversies involving President Jair Bolsonaro.Source: §4.1 · pp. 19–20

4.2. Political agenda is used to draw the attention away from the health crisis

  • Later coverage included Federal Government, media, politics, and international affairs beyond health, while the study found distinct political and COVID-19 topics alongside topics focused exclusively on politics.The authors report that fortnights 6 and 7 contained only COVID-19 topics, before this broader mix appeared.Source: §4.2 · pp. 21–22

4.3. The infodemic: a very specific scenario

  • The authors describe an infodemic setting in which some topics lasted longer than others, while their clusters could be generated despite insufficient lexicon coverage.They also state that restricting the analysis to a specific scenario and model hinders interoperability with other components.Source: §4.3 · pp. 22–23
  • The authors argue that their Markov-chain-based method is more effective than LDA and LSA, which they say require more lexicon and corpus.They also describe examining relationships between relative volumes and temporal dynamics across datasets or individual Twitter accounts.Source: §4.3 · p. 23

5. Discussion and Conclusion

The Brazilian fact-checking outlets addressed distinct topics, and their agendas did not necessarily mirror the country’s situation. The analysis also found politics and the pandemic intertwined, while identifying limits to the study’s scope and the method’s broader applicability.

5. Discussion and Conclusion

  • The study offers first insights into the fake-news agenda through fact-checking content posted by specialized organizations on Twitter.Lupa and Aos Fatos covered the infodemic and other topics relevant to addressing autocratic populism in Brazil.Source: §5 · p. 23
  • The two organizations did not necessarily debunk the same topics, and their agendas did not necessarily reflect the country’s situation.Only a few topics lasted more than one fortnight, and topics covered in one fortnight did not necessarily continue into the next.Source: §5 · p. 24
  • The method could help news organizations identify topics rising or falling in social-media popularity and explore ways to address them.The authors also suggest identifying categories with sharp temporal increases could inform an agenda and support cooperation between fact-checking initiatives.Source: §5 · p. 24
  • Politics and the pandemic were intertwined in varying degrees, while the political agenda shifted toward political actors and institutions and away from the health crisis.The authors say an extreme right-wing political agenda was used to legitimize the leader’s actions and shift focus from the crisis.Source: §5 · p. 24
  • The analysis used two major fact-checking organizations because smaller initiatives did not produce enough tweets during the study period.The authors recommend replication with a larger dataset that includes more organizations and countries.Source: §5 · p. 24
  • Whether the method generalizes beyond tweets and to datasets of different sizes remains an open question.The experiment clustered topics in limited text corpora, but the authors say application to other contexts warrants further study.Source: §5 · pp. 24–25
Loading 2012.11004v1…