COVID-19 has brought about significant economic and social disruption, and misinformation thrives during this uncertain period. In this paper, we apply state-of-the-art rumour detection systems that leverage both text content and user metadata to classify COVID-19 related rumours, and analyse how users, topics and emotions of rumours difer from non-rumours. We found that a number of interesting insights, e.g. rumour-spreading users have a disproportionately smaller number of followers compared to their followees, rumour topics largely involve politics (with an abundance of party blaming), and rumours tend to be emotionally charged (anger) but reactions towards rumours exhibit disapproving sentiments.
about hydroxychloroquine has lead to the death of a man in Arizona.4 COVID-19, a novel disease that was first identified Social media provides a perfect platform for misin China, is an ongoing pandemic that has brought information propagation as they are largely unregabout significant impact to global economy and cre- ulated. To identify misinformation or fake news, ated hitherto unseen social disruption. Since late we may rely on general fact-checking websites,5 Feburary 2020, the pandemic has come to dominate or COVID-19 specific ones. 6 However, due to the both traditional news and social media platforms,1 evolving circumstances of a pandemic it is unlikely and misinformation such as fake news, conspiracy fact-checking or debunking websites will have the theories and rumours thrive during these uncertain capacity to keep themselves up-to-date. times [1]. As such, early detection of potentially malicious
For example, in Italy we saw rumours being spread rumours and understanding what or how rumours to blame the outbreak on migrants and refuges by are being spread during a crisis is an important task making the implicit connection between migration/ [4]But what is a “rumour”? We adopt a widely used movement with the spread of the virus.2 Hydrox- definition which defines it as a story or a statement ychloroquine, a drug that was rumoured to be a with unverified truthful value [5]. COVID-19 treatment despite lacking robust scien- In this paper, we seek to understand what sorts tific evidence about its efectiveness [ 2, 3], is an- of COVID-19 rumours are being spread on Twitter. other popular topic on social media.3 These rumours To this end, we train state-of-the-art rumour deteccan have serious consequences, e.g. misinformation tion systems on out-of-domain labelled rumour data and apply them to COVID-19 related tweets to deTitle of the Proceedings: Proceedings of the CIKM 2020 Workshops tect rumours. We analyse several characteristics October 19-20, Galway, Ireland that diferentiate rumours from non-rumours in this eEmdiatoilrs: so3f7t9h5e5P3r3o@cesetduidnegns:t.Srtmefiat.nedCuo.naura(dL,.ITlairaina)T;iddi COVID-19 data, such as their propagation patterns, ∗Corresponding author: xiuzhen.zhang@rmit.edu.au (X. users, topics, and emotions. Our rumour detection Zhang∗); jeyhan.lau@gmail.com (J.H. Lau) systems leverage both message content and user orcid: 00©0020-2000C0o1py-r5ig5ht5f8or-3th7is9p0ap(eXrb.y Zitshaaunthgor∗s). Use permitted under characteristics, and our analyses reveal a number of CPWrEooUrckReshdoinpgs IhStpN:/c1e6u1r3-w-0s.o7r3g WCCreESatU.iovRerCgo)mmWonosrLkicesnhseoApttribuPtiorno4c.0eInetderinnatgiosnal (CC(CBYE4U.0)R.- interesting insights. For example, rumour-speaders 1https://www.vox.com/recode/2020/3/12/21175570/corona 4https://edition.cnn.com/2020/03/23/health/arizona-corona virus-covid-19-social-media-twitter-facebook-google. virus-chloroquine-death/index.html.
2https://time.com/5789666/italy-coronavirus-far-right-sal 5E.g. https://www.snopes.com/ and https://www.factcheck. vini/. org/.
3https://abcnews.go.com/Health/tracking-hydroxychloro 6E.g. https://www.fema.gov/coronavirus/rumor-control quine-misinf ormation-unproven-covid-19-treatment-ended/s and https://www.defense.gov/Explore/Spotlight/Coronavirus/ tory?id=70074235. Rumor-Control/ tend to have low follower but high followee count, Table 1 rumours tend to talk about politics (mostly party Rumour classification training data. blaming) and are more emotionally charged (e.g. anger), but reactions towards them are also disproportionately more disapproving. We also provide a website7 to share our latest findings and up-to-date rumour tracking data analysis. #source tweets #all tweets #users #rumours #non-rumours
Rumour detection approaches can generally be categorised into text-based or non-text-based meth- 3. Methodology and Data ods. Text-based methods focus on rumour detection using the textual content, which may include the 3.1. Rumour Classification original source document/message and user com- We focus on the detection of rumours vs. non-rumours, ments/replies. Shu et al. [6] introduce linguistic fea- rather than the veracity (truthfulness) of rumours. tures to represent writing styles and other features In other words, truthful, untruthful and unverified based on sensational headlines from Twitter to de- rumours are all rumours in our definition — they extect misinformation. To detect rumours as early as hibit novelty/surprise in terms of content and tend possible, Zhou et al. [7] incorporate reinforcement to be spread by users — while non-rumours are tralearning to dynamically decide how many responses ditional news stories and non-news related converare needed to classify a rumour. sations. The task of rumour detection can therefore
Non-text-based methods utilise features such as be formulated as a binary classification problem, user profiles or propagation patterns for rumour and we explore both textual information and user detection. For example, Gupta et al. [8] propose a metadata as input features. soef mtwi-eseutpseursviinsgedhaanpdp-rcoraacfhtetdofeevaatuluraetsebtahseedcroenditbwileiteyt EaCchonssoiudrecreatsweteoeft issaosusrocceitawteedetws ith= a{ l1a,b e2l, ...i, nd}i-. and user metadata. Castillo et al. [9] leverage user cating the tweet is rumour ( = 1) or non-rumour frceeraegtdiusibtrreialsittisyou.ncFhoalgaloeswbaneinldigenfs/utiunmdtbeienesrtieooxnfpflfooolrrleorwmumeorroseutcroopamrsepsdeleiscxs- (trieop=nlis0e:)s. Eana=cdh{qsuo1ou,trecs2e., .Et.w.a,ce hetr}e.aRacletsiaoocnhtiaosnasissaerreteporrfeetswenreeetaetcsd-, tion [10], where users are categorised based on their with a tuple = ( , ), which includes the fol“support” or “deny” attitudes toward a piece of news. lowing information: is the textual content of the
In terms of emotion analysis on social media, reaction, and the metadata features of the user
Larsen et al. [11] propose using principle compo- who creates the reaction tweet.
nent analysis to predict emotions of tweets, and In terms of rumour classification models, we
exintroduces a real-time system that analyses global plore two methods based on: (
For analysis of COVID-19 on Twitter, Li et al. [13] For the user-based model, it uses a convolutional explore using multi-lingual BERT [14] to analyse network to process user metadata features extracted public mental health using tweets. Sharma et al. [15] from their Twitter profile and a recurrent network present analysis of COVID-19 misinformation based to combine a set of user features in the propagaon news sources from fact-checking sites rather than tion path. We extend the original eight features to automatic classification and contrast analysis of ru- sixteen features.8 We limit the processing of user mours versus non-rumours. features in the propagation path to the first 50 users.
screenname, count of posts and favourite posts; and the binary features are: whether the profile is protected, has URL, profile image, uses default profile and default profile image.
Figu20r2e0-011-1:5Fil2t02e0r-0e2-d01E20n20g-0l2i-1s5h 2T0D2wa0-t0ee3-e01ts20V20o-03l-u15me2020-04-01
To combine both text and user models for rumour detection, we create an ensemble model that takes the output of both models to make the final prediction. As both models produce a probability value for the rumour class in each source tweet, we com- remaining tweets are “reaction tweets”: retweets, pute the mean probability and tune a threshold to replies or quotes).11 separate rumours from non-rumours.9 Figure 1 shows the volume of filtered English tweets over time. We can see there is some traf3.2. Labelled Rumour Data ifc of COVID-19 related tweets from late January 2020, although it doesn’t really pick up until midWe use Twitter15, Twitter16 [18], PHEME [19], and March. We suspect the spike of activity may be SemEval2019 [20] as training data to train our bi- due the World Health Organisation declaring it as a nary rumour classification models. For Twitter15, pandemic on 12th March.12.
Twitter16 and PHEME, there are originally 4 classes: In terms of pre-processing, we tokenise the tweets truthful rumours, untruthful rumours, unverified with the TweetTokenizer [22] package of NLTK, rumours and non-rumours; we collapse the truthful, and lowercase and lemmatise all words with the untruthful and unverified rumours into the rumour WordNetLemmatizer package, as well as remove class. SemEval2019 focuses on veracity classification digits, non-Latin characters and @usernames. We and as such has only 3 classes (truthful, untruthful also filter stopwords based on an extended NLTK and unverified); they are all treated as the rumour stopword list, which includes COVID-19 specific class. Statistics of the datasets is presented in Ta- stopwords, such as covid19 or coronavirus. Hyperble 1. links are encoded with a special token for rumour classification (Section 4.1) or removed for topic anal
our analyses.10 We use version 4 of the dataset, 4. Results and Analysis which contains tweets from 1st January 2020 to 5th April 2020. The dataset is regularly updated, 4.1. Rumour Classification and collects tweets for several languages (English, French, Spanish and German) based on COVID-19 To assess the quality of the rumour classification keywords. models, we first evaluate the in-domain performance
As we are interested in rumour analyses in En- of Twitter15, Twitter16 and PHEME. For each dataset, glish, we filter the data to keep only source tweets we randomly split the full data in 60%/20%/20% to that are in English (based on Twitter metadata) and create the training, validation and test partitions. also have at least 10 replies (since those with few In-domain classification performance is presented reactions are of little significance for rumour analy- in Table 3 (in-domain performances are those where sis). Table 2 presents some statistics of our filtered “Train” and “Test” are from the same domain).13 dataset. We have approximately 30M tweets postifltering, and 60K of them are source tweets (the 11Quote is similar to retweet, except that it contains some response to the original tweet. Both retweets and quotes are displayed on the user’s home page, while replies are not.
12https://twitter.com/WHO/status/1237777021742338049 13For the ensemble model, we tune the threshold based on the validation set, and ranges from 0.7 to 0.8.
9That is, the ensemble model labels a source tweet as rumour if the mean probability ≥ .
10https://github.com/thepanacealab/covid19_twitter.
the user model, but the ensemble model (“user+text”) performs best.
We next evaluate cross-domain performance. Given a test domain (e.g. Twitter15), we train the rumour classification models using a combination of all outof-domain data (e.g. Twitter16 and PHEME), and assess their accuracy on the test domain. This is an Figure 4: Bigram word cloud. arguably more dificult setting, as there is little or no topic overlap between the diferent domains.
Unsurprisingly, we see a dip in accuracy com- COVID-19 data (Section 3.3).14 pared to the in-domain performance. Encouragingly, In total, out of the 60K source tweets (Table 2) however, with the ensemble model we are still get- 15K are classified as rumours. These rumours (and ting at least 78% accuracy over all domains, sug- non-rumours) will serve as the basis for user, topic gesting that the model is robust for cross-domain and emotion analyses in subsequent experiments. rumour detection.
Given these results, we next train an ensemble than14tWhee sthetretshheotlhdrseswheoludsedtoin0.t8h5e, wcrhoiscsh-disommaairngienxaplleyrihmigehnetsr model on all datasets (Twitter15+Twitter16+PHEME), to improve precision. Note that the COVID-19 data does not inand use it to classify tweets on our filtered English clude user metadata, so we crawl them using the oficial Twitter API. #WuhanVirus, #MOG, #OneVoice1, #FoxNews, #DemocratsAreDestroyingAmerica, #KAG2020, #ChinaVirus, #Hydroxychloroquine, #IWillStayAtHome, #ChinaLiedPeopleDied, #MasksNow, #TheMoreYouKnow, #TheResistance, #StopAiringTrump, #VoteRedToSaveAmerica, #WuhanHealthOrganisation, #CCP_is_terrorist, #DemCast, #BillGates, #TrumpIsTheWORSTPresidentEVER, #TrumpOwnsEveryDeath, #5G trump, pelosi, bill, democrat, fox, gop, american, blame, president, briefing, joe, lie, hoax, medium, fail, governor, response, china, vote, drug, hydroxychloroquine nancy pelosi, chinese chinese, jared kushner, chinese communist, trump response, held accountable, trump supporter, trish regan, speaker pelosi, joe biden, bill gate, china lie, task gown, deep state, blame trump, fox business #BREAKING, #StaySafe, #CoronaUpdate, #CoronavirusLockdown, #IndiaFightsCorona, #CoronaOutbreak, #DonaldTrump, #COVID19PH, #COVID19Pandemic, #covid19australia, #TakeResponsibility, #21daylockdown, #CoronavirusPandemic, #Covid19usa, #StayHomeStaySafe, #StayAtHome, #coronapocalypse, #flu, #Italia, #COVID19OhioReady, #COVID_19uk, #masks, #china, #StrongerTogether positive, confirm, total, india, march, symptom, health, minister, due, nigeria, lockdown, update, death, infect, old, donate, day, negative, cancel, wash, hand, social, hour, announce, today, data, stay, worker, isolation, quarantine bring total, march march, year old, total number, relief fund, number conifrm , patient positive, prime minister, travel history, premier league, wash hand, hubei province, first death, cruise ship, health condition, social care 3% 4% 3% 3% ❤ 4% 5% 6%
15https://www.theatlantic.com/health/archive/2020/03/cor
onavirus-pandemic-herd-immunity-uk-boris-johnson/608065/.
reactions over time for rumours and non-rumours. china lie), (5) status reports (death toll and death
Although rumours tend to attract more reactions rate), (6) healthcare (doctor nurse and health worker);
in the first 24 hours, we see a convergence after 48 (7) panic buying (toilet paper16); and others.
hours. To better understand the topical diference
between rumours and non-rumours, we compute
log4.3. Topic Analysis likelihood ratio [24] of unigrams, bigrams, hashtags
and display the most salient words in Table 5.17
To understand the popular topics discussed in Twit- To ease readability, we highlight some of the salient
ter, we first present a bigram wordcloud in Figure 4. words in the table. For rumours, US politics is one of
We see several broad topics: (
17We include both source tweets and reactions to contruct the rumour and non-rumour “corpora”, and use NLTK’s BigramAssocMeasures to compute the loglikelihood ratio. To decide whether a word is salient for rumour or non-rumour, we look at its normalised frequency. #TrumpIsTheWORSTPresidentEVER). Unsurprisingly, (Figure 6(a)), anger dominates all hashtags, although Fox News (#FoxNews and fox) are associated with #ChinaVirus source tweets are substantially “anrumours.18 China is another topic, and the hash- grier” (68%!). Anger in non-rumour source tweets tags/bigrams suggest blaming (#ChinaVirus, (Figure 6(b)) is a little more toned down; interest#CCP_is_terrorist, #WuhanHealthOrganisation and ingly the dominant emotion for the global lockdown china lie). We also see also some of the well-known (#CoronavirusLockdown) is more positive than negCOVID-19 rumours/hoaxes: #Hydroxychloroquine, ative (41% “thumbs up” vs. 33% “angry”). #BillGates,19, and #5G.20 Moving over to the emoji distribution for reac
Looking at non-rumours, the topics are very dif- tions towards rumour tweets (Figure 7(a)), we see ferent: they are mostly related to health advice (#Coro- anger in all hashtags, but some of the other emonavisuLockdown, #StayHomeStaySafe and wash hand) tions are rather curious, e.g. “thumbs up” (approval) and status updates (total number, number confirm ), for #Hydroxychloroquine, and “googly eyes” (attenand more neutral/positive in tone (#StrongerTogether tion drawing) for #BillGates. Unsurprisingly though, and #coronapocalypse). Politics is rare, although we reactions for all non-rumour hashtags (Figure 7(b)) see prime minister, which may be related to UK poli- are dominated by “prayers” and approval emojis tics. Another interesting non-rumour topic observed (“thumbs up” and “biceps”), suggesting that despite here is the cruise ship outbreaks (cruise ship). the general doom and gloom atmosphere of COVID19, there is still a sense of positivity. 4.4. Emotion Analysis To understand the public sentiment during the COVID- 5. Conclusion 19 crisis, we explore using an emotion prediction system to classify the emotion of tweets in our data. We explored an ensemble model combining textWe experiment with DeepMoji [25], a Bi-LSTM with based and user-based rumour detection models to attention model trained on a large number of emoji classify COVID-19 related rumours on Twitter. We occurrences in tweets. We use their pre-trained presented quantitative evaluation to demonstrate model to label our data with 63 predefined emojis. its robustness in cross-domain rumour detection,
Figure 5 illustrates the distribution of emojis for analyse the users, topics and emotions of rumours source and reply tweets in rumours and non-rumours. vs. non-rumours, and found a number of insights. Looking at the emotions of source tweets (Figure 5(a) and (b)), “anger” dominates both rumours and non- Acknowledgements rumours, but substantially more in rumours than non-rumours (54% vs. 34%). Non-rumours also see This work is partially supported by the Australian more “thumbs up” (encouragement), although the Research Council Discovery Project DP200101441. diference is less severe (25% vs. 18%).
For reply tweets (Figure 5(c) and (d)), we see a similar distribution for the top-3 emotions (“anger”, References “thumbs up” and “mask face”), but the interesting observation here is the emojis for the rest (left half of the pie chart): the reply tweets for rumours display disapproving sentiments (e.g. “punch” and “frown”), while that of non-rumours are generally positive and encouragement in tone (“pray”, “love” and “biceps”).
We next present the emoji distribution for some of the salient hashtags for the source and reaction tweets in Figure 6 and 7 respectively, to see how public attitude towards diferent topics vary across rumours and non-rumours. For rumour source tweets
18https://www.nytimes.com/2020/03/31/opinion/coronavir us-fox-news.html.
19https://www.bbc.com/news/52847648.
20https://www.reuters.com/article/uk-factcheck-coronavir us-5g/false-claim-coronavirus-is-a-hoax-and-part-of-a-wider5g-and-human-microchipping-conspiracy-idUSKBN22P22I. [4] B. Wang, J. Zhuang, Crisis information distri- [15] K. Sharma, S. Seo, C. Meng, S. Rambhatla, bution on Twitter: a content analysis of tweets A. Dua, Y. Liu, Coronavirus on social media: during Hurricane Sandy, Natural hazards 89 Analyzing misinformation in Twitter conversa(2017) 161–181. tions, arXiv preprint arXiv:2003.12309 (2020). [5] G. W. Allport, L. Postman, The psychology of [16] L. Tian, X. Zhang, Y. Wang, H. Liu, Early detecrumor. (1947). tion of rumours on Twitter via stance transfer [6] K. Shu, A. Sliva, S. Wang, J. Tang, H. Liu, Fake learning, in: European Conference on Infornews detection on social media: A data min- mation Retrieval, Springer, 2020, pp. 575–588. ing perspective, ACM SIGKDD Explorations [17] Y. Liu, Y.-F. B. Wu, Early detection of fake news Newsletter 19 (2017) 22–36. on social media through propagation path clas[7] K. Zhou, C. Shu, B. Li, J. H. Lau, Early rumour sification with recurrent and convolutional netdetection, in: Proceedings of the 2019 Con- works, in: Thirty-Second AAAI Conference on ference of the North American Chapter of the Artificial Intelligence, 2018.
Association for Computational Linguistics: Hu- [18] J. Ma, W. Gao, K.-F. Wong, Detect rumors in man Language Technologies, Volume 1 (Long microblog posts using propagation structure and Short Papers), 2019, pp. 1614–1623. via kernel learning, in: Proceedings of the [8] A. Gupta, P. Kumaraguru, C. Castillo, P. Meier, 55th Annual Meeting of the Association for Tweetcred: Real-time credibility assessment of Computational Linguistics (Volume 1: Long content on Twitter, in: International Confer- Papers), 2017, pp. 708–717. ence on Social Informatics, Springer, 2014, pp. [19] E. Kochkina, M. Liakata, A. Zubiaga, All-in228–243. one: Multi-task learning for rumour verifica[9] C. Castillo, M. Mendoza, B. Poblete, Informa- tion, arXiv preprint arXiv:1806.03713 (2018). tion credibility on Twitter, in: Proceedings [20] G. Gorrell, E. Kochkina, M. Liakata, A. Aker, of the 20th international conference on World A. Zubiaga, K. Bontcheva, L. Derczynski, wide web, ACM, 2011, pp. 675–684. Semeval-2019 task 7: Rumoureval, determin[10] X. Liu, A. Nourbakhsh, Q. Li, R. Fang, S. Shah, ing rumour veracity and support for rumours, Real-time rumor debunking on Twitter, in: in: Proceedings of the 13th International WorkProceedings of the 24th ACM International shop on Semantic Evaluation, 2019, pp. 845– on Conference on Information and Knowledge 854.
Management, ACM, 2015, pp. 1867–1870. [21] J. M. Banda, R. Tekumalla, G. Wang, J. Yu, T. Liu, [11] M. E. Larsen, T. W. Boonstra, P. J. Batterham, Y. Ding, G. Chowell, A large-scale COVIDB. O’Dea, C. Paris, H. Christensen, We feel: 19 Twitter chatter dataset for open scientific mapping emotion on Twitter, IEEE journal research–an international collaboration, arXiv of biomedical and health informatics 19 (2015) preprint arXiv:2004.03688 (2020). 1246–1252. [22] S. Bird, E. Klein, E. Loper, Natural Language [12] N. Farruque, C. Huang, O. Zaiane, R. Goebel, Processing with Python, 1st ed., O’Reilly MeBasic and depression specific emotion identifi- dia, Inc., 2009. cation in Tweets: multi-label classification ex- [23] S. Vosoughi, D. Roy, S. Aral, The spread of periments, in: The 20th International Confer- true and false news online, Science 359 (2018) ence on Intelligent Text Processing and Com- 1146–1151.
putational Linguistics (CICLing), 2019. [24] T. E. Dunning, Accurate methods for the statis[13] I. Li, Y. Li, T. Li, S. Alvarez-Napagao, D. Gar- tics of surprise and coincidence, Computacia, What are we depressed about when we tional linguistics 19 (1993) 61–74. talk about COVID19: Mental health analysis [25] B. Felbo, A. Mislove, A. Søgaard, I. Rahwan, on tweets using natural language processing, S. Lehmann, Using millions of emoji occurarXiv preprint arXiv:2004.10899 (2020). rences to learn any-domain representations for [14] J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, detecting sentiment, emotion and sarcasm, in: BERT: Pre-training of deep bidirectional trans- Proceedings of the 2017 Conference on Empirformers for language understanding, in: Pro- ical Methods in Natural Language Processing, ceedings of the 2019 Conference of the North Copenhagen, Denmark, 2017, pp. 1615–1625. American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), Minneapolis, Minnesota, 2019, pp. 4171–4186.