4. Pandemic Response and Crisis
Informa<cs: An Impera<ve for
Public Health Messaging
Chun-Ming Lai∗, Yi-Chen Liu∗, Rong-Ching Chang∗,
Jon W. Chapman† and Chu-Hsing Lin∗
∗ Tunghai University
{cmlai,ycliu,rcc,chlin}@thu.edu.tw
† University of California, Davis
jwchapman@ucdavis.edu
6. Background COVID-19 in Taiwan
Face Masks Map (FMM) Inventory status for
all stores selling face
masks
Significant
announcements from
public health officials
7. Research Ques@ons
• RQ1. What does the crisis informatics platform regarding PPE look like?
• RQ2. In the event of pandemic, does the mask rationing (Policy)
efforts affect the total number of query to face masks? Can we
deduce a strong correlation, or even a causality.
• RQ3. Which type of users will be more likely to use FMM?
8. Main Finding
Face Masks Map (FMM)
User Behavior
Significant
announcements from
public health officials,
Ministry of Health and
Welfare (MoHW)
Correlation
9. RESOURCE QUERY PLATFORM
Face Mask Map
Main Components
1. Geolocation identifier
• Input address or allow device
location access
• Display inventory status of
nearby stores
2. Connector to MoHW
• API access to mask inventory
3. Google Analytics (GA)
• Visitor activity analysis
• Network traffic analysis
10. User Session Data Descrip@on
• Focus
• temporal pattern of audiences for active 3-month
• from February 1 to May 1
• total of 3,812,864 users , 10,472,439 sessions
11. User Session Data Description (From GA)
• Definition of terms
• User
• A user is denoted as 𝑈 in GA
• Session
• A session 𝑈!!
" is a group of event records for user 𝑈!
in a given time period (30
minutes)
• Expired after 30-mins inactive and midnight
• Overview of Sessions in specific duration
• Number of Sessions per User 𝑈!"
"
12. CORRELATION BETWEEN CRITICAL EVENTS
AND ONLINE USERS BEHAVIOR
• Pandemic Event
• MacroQuery Temporal Peak (Aggregated Behavior)
• Result
13. Pandemic Event Identification
• Pandemic information https://covid19.mohw.gov.tw/
• Event Identification
• with the label material preparation
• including the hashtag: Face Masks
• Manually checked by at least two authors to clarify that the target influence
would be a substantial number of citizens
14. Pandemic Event Iden@fied
1) E1 (2020-02-06): Name-Based Mask Distribution System1.0 on board, people need to hold their
health insurance cards (HIC) to purchase at pharmacies. Add diversion here.
2) E2 (2020-02-20): One can purchase 4 child face masks within 7 days given a child HIC.
3) E3 (2020-02-27): Adjust the distribution method of children’s masks (1) Each person can hold up
to 3 children’s HIC (2) HIC are not subject to the diversion restriction based on the last number of
Social Security Number.
4) E4(2020-03-05):The purchase of adult masks increased to 3 pieces, meanwhile, children masks 5
pieces, all in 7 days.
5) E5 (2020-03-12): Name-Based Mask Distribution System 2.0 on board. Pre-Order online and pick-
up in- store (pharmacies) were allowed based on SSN through MoHW application.
6) E6 (2020-03-19): Purchase date was not restricted by the last number of SSN.
7) E7(2020-04-09):Adjust the purchase cycle and quantity of masks, 9 for adults and 10 for children
every 14 days. In addition, cancel all odd and even number split restrictions.
8) E8 (2020-04-13): Masks requisition and export ban has been extended to the end of June.
9) E9 (2020-04-22): Name-Based Mask Distribution System 3.0 on board, one is able to pre-order at
every domestic convenience store.
15. MacroQuery Temporal Peak
Temporal peak within a 90-day interval
• a time series 𝑇 = {(𝑡1 , 𝑥1 ), (𝑡2 , 𝑥2 ), … (𝑡𝑛 , 𝑥𝑛 )}, as 𝑥𝑖 represents
the total number of sessions on 𝑡𝑖.
• length of 𝑡𝑖 is 24-hour, from 12:00 am to 11:59 pm.
• Peak condition
• Greater than yesterday and tomorrow:
• ∀𝑖 = 2, 3, … , 𝑛 − 1, if 𝑡𝑖 represents when the peak occurs, then it satisfies that 𝑥𝑖 >
𝑥𝑖 − 1 ∧
𝑥𝑖 > 𝑥𝑖 + 1.
• Greater than an absolute threshold ε
• ε = 1000
• Greater than the average of week
• ̅
𝑥 =
"
#
×(𝑥𝑖 − 5 +
𝑥𝑖 − 4 + ⋯ +
𝑥𝑖 + 𝑥𝑖 + 1)
• Hypothesize: a news requires a day to disseminate
18. DISCUSSION FOR USERS COMPOSITION
• What is the best channel for broadcasting the information regarding
supplements when disaster occurs?
• What kind of devices do they prefer to use?
➠ Disaster infrastructure maintenance and management.
24. Conclusion
• Examine the usage patterns exhibited by users for crisis informatics in
the midst of COVID-19 in Taiwan.
• Users query peak for appropriate PPE was able to be detected by
official announcement which gave us predictive power.
• Source and medium
• People rely on official website, search engine, and social media to obtain
information in the midst of crisis.
25. Dataset of Propaganda Techniques
of the State-Sponsored Informa9on
Opera9on of the People’s Republic
of China
Rong-Ching Chang
With Chun-Ming Lai, Kai-Lai Chang, Chu-Hsing Lin
Dataset: https://github.com/annabechang/Propaganda_Tech_Twitter_PRC
Twitter: @AnnCC12
26. Propaganda, Variants And Applications
Propaganda
Computational
propaganda
State-backed
propaganda
Information
warfare
Propaganda
techniques
27. Current Work On State-backed Propaganda
Propaganda
Or not
Propaganda
Techniques
Text
Network
MulJ-Modal
Ref: Da San Martino et al. (2020)
28. Current Datasets
Work Granularity
Number Of
Class
Source Language
Number of
Propagandist
Snippets
Multi-
label
Baisa et al.
(2019)
Fragment:
Specific Text
Spans from
News
18
(8 techniques,
10 doc
attributes)
News Czech 7,494 Yes
Da San
Martino et al.
(2019)
Fragment:
Specific Text
Spans from
News
18 News English 7,385 Yes
This Paper Sentences
21
(18 techniques,
3 other
attributes )
Twitter
Simplified
Chinese
9,950 Yes
29. Propaganda Techniques Labels Propaganda Techniques Frequency
1 Presenting Irrelevant Data 13
2 Straw Man 2
3 Whataboutism 2,509
4 Oversimplification 37
5 Obfuscation 12
6 Appeal to authority 50
7 Black-and-white 265
8 Name Calling 2,313
9 Loaded Language 2,609
10 Exaggeration or Minimisation 114
11 Flag-waving 81
12 Doubt 147
13 Appeal to fear or prejudice 141
14 Slogans 37
15 Thought-terminaVng cliché 0
16 Bandwagon 64
17 Reductio ad Hitlerum 83
18 Repetition 60
19 Neutral Political 915
20 Non-Political 6117
21 Meme humor 5
• 8 Name-calling
• Labeling the target with the
inten]on of arousing prejudices or
making an associa]on with
stereotypes.
• 9 Loaded Language
• Using emo]onal words to
influence audience opinions.
30. Source of Data
• Information Operations from Twitter Transparency Center
• https://transparency.twitter.com/en/reports/information-operations.html
31. Source of Data
• Total number of
accounts: 936
• We sampled tweets from
a 744 account batch
32. Using This Dataset With The Propaganda
Technique Labels
• https://github.com/annabechang/Propaganda_Tech_Twitter_PRC
33. Dataset Sample Display
Tweetid Translated Tweet Propaganda Techniques Labels
990189929 836699648 The truth and hypocrisy under the
false democratic face of Guo
Wengui, the clown jumping-bean,
is now undoubtedly exposed!
小丑跳豆郭文贵虚假民主面孔下
的真相和虚伪,如今无疑暴露无
遗!
3,8,9
114879827 6281364480 We must severely punish the
rioters and return Hong Kong to a
peaceful condition
必须严惩暴徒,让香港回归和平
8,9,13,14
34. Language Usage Of The Dataset
Language
Usage
• Total number of tweets
in Chinese: 74,277
• The number of tweets
we sampled: 9,950
• Uren et al. (2019): re-
purposed spam accounts
35. The Labeling Process Conducted
• Manually labelled the first 1,000 tweets
• Observed consistent propaganda techniques
and their alignment toward a certain en=ty
• Iterate:
• Build pre-defined labels according to the
men6oned en66es, their variant-related names
and categories
• Review tweets and pre-defined labels
• Check if en6ty or reference name is on the list
• If not, add it or edit pre-defined rule
• Con6nue
Keyword Category Count
Exiled or anti-
government Chinese
5,406
Hong Kong protest 209
International Geo-
Political
1,718
Taiwan independence 2
36. The Labeling Process Conducted
• Manually labelled the first 1,000 tweets
• Observed consistent propaganda techniques
and their alignment toward a certain entity
• Iterate:
• Build pre-defined labels according to the
mentioned entities, their variant-related names
and categories
• Review tweets and pre-defined labels
• Check if entity or reference name is on the list
• If not, add it or edit pre-defined rule
• Continue
Translated Tweet Propaganda
Techniques Labels
The truth and
hypocrisy under the
false democratic face
of Guo Wengui, the
clown jumping beam,
is now undoubtedly
exposed!
3,8,9
Named Entity Keyword Category:
Exiled or anti-
government Chinese
Name Variants 8 Name Calling
37. Multi-label Propaganda Technique
Classification
• Bidirecconal Encoder Representacons from Transformers (BERT)
• bert-base-chinese pre-trained model provided by Huggingface
• Accuracy: 0.80352
• F1 Score (Micro): 0.85431
• F1 Score (Macro): 0.20803
38. Some Of The Possible Ways You Can Use This State-
Sponsored Propaganda Technique Dataset
Cross-
Platform
Political vs
Non-
Political
Multimodal
State-Level
Stance
Detection
Propaganda
Detection
Multilingual
39. References
• Giovanni Da San MarVno, Seunghak Yu, Alberto Barrón-Cedeno, RosVslav Petrov, and Preslav Nakov. 2019. Fine-grained
analysis of propaganda in news arVcles. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language
Processing and the 9th InternaVonal Joint Conference on Natural Language Processing (EMNLP-IJCNLP). AACL, Hong Kong,
5640–5650.
• Giovanni Da San MarVno, Stefano Cresci, Alberto Barrón-Cedeño, Seunghak Yu, Roberto Di Pietro, and Preslav Nakov.
2020. A survey on computaVonal propaganda detecVon. arXiv e-prints.
• Vít Baisa, Ondřej Herman, and Ales Horak. 2019. Benchmark Dataset for Propaganda DetecVon in Czech Newspaper Texts.
In Proceedings of the InternaVonal Conference on Recent Advances in Natural Language Processing (RANLP 2019).
INCOMA Ltd, Varna, 77–83.
40. Online Sentiment and Reactions To
Posts on Facebook:
Leading Indicators for State-level
Covid-19 Confirmed Cases
Rong-Ching Chang, Chun-Ming Lai, Chu-Hsing Lin and Kai-Chih Pai
Tunghai University, Taiwan
ICLR Workshop on AI for Public Health
41. Main Contributions
Examines the relaconship between online sencment and the
number of local COVID-19 confirmed cases
Invescgates public online sencment instead of user-level
sencment
Focuses on Facebook instead of Twieer
1
2
3
42. Delayed COVID-19 Test Results for Weeks
“In September, 42% of those
tested had to wait at least 3
days before getting their
results; the corresponding
percentage in April was
56%.”
Chwe, Hanyu, et al. "The COVID States
Project# 17: COVID-19 test result times."
(2021).
43. Why is COVID-19 Surveillance Important?
Delayed COVID-19
Test Results
Delayed Number of
Confirmed Cases
Reported
Delayed Policy
Response
More COVID-19
Spread
More Confirmed
Cases
Lack Of Medical
Resources
44. Social Media for COVID-19 Surveillance
• Shen et al. (2020) found reporting of COVID-19 related symptoms on
Weibo posts has a significant predictive power up to 14 days before
the official statistics.
• Gharavi et al. (2020) observed a temporal lag between the number of
reports of symptoms on Twitter and officially-reported positive cases.
45. Research Ques@on
Are large-scale collections of public sentiments and post responses on
Facebook a suitable data source for the number of state-level Covid-19
confirmed cases?
We have found that the senZment and reacZons to posts on public pages and groups from Facebook
can be a leading factor in predicZng state-level future confirmed cases 24 to 70 days in advance.
47. Data Collection
California, USA
25 January 2020 to 20 December 2020
Over-performing
Score
Message
The number for
each reaction
Aggregacon of
reaccon
COVID-19
Confirmed Cases
49. Conclusion
We have found that the sentiment and reactions to posts on
public pages and groups from Facebook can be a leading factor in
predicting state-level future confirmed cases 24 to 70 days in
advance.
50. References
• Hanyu Chwe, et al. "The COVID States Project# 17: COVID-19 test result times." (2021).
• Icons from www.flaticon.com
• Cuihua Shen, Anfan Chen, Chen Luo, Jingwen Zhang, Bo Feng, and Wang Liao. Using reports of symptoms and diagnoses
on social media to predict covid-19 case counts in mainland china: Observational infoveillance study. Journal of medical
Internet research, 22(5):e19421, 2020.
• Erfaneh Gharavi, Neda Nazemi, and Faraz Dadgostari. Early outbreak detection for proactive crisis management using
twitter data: Covid-19 a case study in the us. arXiv preprint arXiv:2005.00475, 2020.
• CrowdTangle Team. Crowdtangle. Facebook, Menlo Park, California, United States, 2020.
51. Conclusion
• Sociality with Programming and Algorithms
• Humanity
• Social as a responsibility
• Facebook Destroys Friendships with 3B reactions and likes
• Computing Friendship