This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research, is properly cited. The complete bibliographic information, a link to the original publication on https://www.jmir.org/, as well as this copyright and license information must be included.
Due to the urgency caused by the COVID-19 pandemic worldwide, vaccine manufacturers have to shorten and parallel the development steps to accelerate COVID-19 vaccine production. Although all usual safety and efficacy monitoring mechanisms remain in place, varied attitudes toward the new vaccines have arisen among different population groups.
This study aimed to discern the evolution and disparities of attitudes toward COVID-19 vaccines among various population groups through the study of large-scale tweets spanning over a whole year.
We collected over 1.4 billion tweets from June 2020 to July 2021, which cover some critical phases concerning the development and inoculation of COVID-19 vaccines worldwide. We first developed a data mining model that incorporates a series of deep learning algorithms for inferring a range of individual characteristics, both in reality and in cyberspace, as well as sentiments and emotions expressed in tweets. We further conducted an observational study, including an overall analysis, a longitudinal study, and a cross-sectional study, to collectively explore the attitudes of major population groups.
Our study derived 3 main findings. First, the whole population’s attentiveness toward vaccines was strongly correlated (Pearson
This study tracked the year-long evolution of attitudes toward COVID-19 vaccines among various population groups defined by 8 demographic characteristics, through which significant disparities in attitudes along multiple dimensions were revealed. According to these findings, it is suggested that governments and public health organizations should provide targeted interventions to address different concerns, especially among males, older people, and other individuals with low levels of education, low awareness of news, low income, and light use of social media. Moreover, public health authorities may consider cooperating with Twitter users having high levels of social influence to promote the acceptance of COVID-19 vaccines among all population groups.
Since the emergence of the COVID-19 pandemic in 2019, human health and life have been gravely jeopardized globally. Governments and public health agencies worldwide primarily implemented the following 2 measures to control this pandemic: (1) nonpharmaceutical preventive methods, such as social distancing [
Traditionally, developing a new vaccine from scratch is a complex process, which takes considerable time to accomplish. The main procedures of traditional vaccine development include preclinical studies (about 2-4 years); phase I, II, and III trials (about 5-7 years total); and manufacturing and approval (about 1-2 years) [
To identify COVID-19 vaccine-related literature, we searched the World Health Organization COVID-19 database [
Monthly statistics on newly published papers related to COVID-19 vaccines.
In order to better understand the current research state of public attitudes toward COVID-19 vaccines, we filtered papers with the term “attitude” in the titles, and “cross-sectional” or “longitudinal” in the titles or abstracts, and retrieved 85 relevant papers. Then, we identified the data collecting methods used in these studies manually, and discovered that there were primarily 2 types as follows: survey (81 papers, 95%) and data mining (4 papers, 5%). Studies involving surveys mainly adopted a cross-sectional design to investigate public attitudes toward COVID-19 vaccination over a short period, while studies involving data mining employed either a longitudinal or cross-sectional design, but rarely both. We conducted a detailed literature review of COVID-19 vaccine-related studies involving these 2 frequently used analysis methods (cross-sectional and longitudinal).
A cross-sectional study analyzes data collected from a population or a predefined subset at a single point in time. Many studies have used this method, primarily through surveys, to explore populations’ attitudes toward receiving COVID-19 vaccination and the factors that affect these attitudes. For example, Lazarus et al [
A longitudinal study is a method that observes some specific variables over an extended period of time. Many studies have applied this method to track trends in population attitudes toward COVID-19 vaccines based on data mining, as data mining can process long-term and large-scale data. Pullan et al [
Considering the above-mentioned limitations, we combined cross-sectional and longitudinal analyses in this work to study the online attitudes toward COVID-19 vaccines based on data mining results of tweets. By doing this, we can not only track long-term evolution, but also discern the disparities of attitudes among various population groups. In addition, this work explored the correlation between the whole population’s attentiveness toward vaccines and official COVID-19 statistics, and analyzed the abrupt influences of some major vaccine-related events. These findings can be used as a guide to assist governments and public health organizations in monitoring the trends of different population groups and relieving the low sentiments of specific groups. It is worth mentioning that the method proposed in this study can be easily reutilized to track the attitude evolution of population groups toward any other public health events. The source code developed in this study has been publicly released via GitHub for follow-up research [
The remainder of this paper is organized as follows. The Methods section first introduces the data collection and preprocessing procedures, and then presents the structure of a 2-step methodology with its essential design details. The Results section analyzes the mining outcomes from multiple dimensions. Finally, the Discussion section concludes the work.
The Twitter data used in this study were randomly collected with our self-designed program using the Twitter application programming interface (API) [
So far, there have been a number of publicly released COVID-19 data sets available for scientific research, such as the data set released by Chen et al [
As each tweet in our data set contains a detailed user profile, tweet text, a creation time, a location, statistics, and some other structured data, it can be treated as one online participant with the characteristics of an individual or organization, carrying an attitude for some specific topic. Since English is the most widely spoken language worldwide, we first excluded 925,008,121 non-English tweets by the language attribute of the tweet object, and obtained 524,293,459 English tweets on general content (hereinafter referred to as general content tweets). Then, we excluded 512,781,119 non-COVID-19–related tweets with a filtering pattern composed of 590 COVID-19 keywords and hashtags according to Twitter COVID-19 filtering rules [
Data selection.
We designed a 2-step methodology for this study, as shown in
The structure of the 2-step methodology in our study.
The data mining step plays a fundamental and decisive role in the entire research. We applied natural language processing, image processing, and tag extracting algorithms on tweets to extract users’ real-world characteristics (user type, gender, age, occupation, and location), cyberspace characteristics (account age and follower number), sentiment polarities, and emotion types. One set of mining outcomes from a tweet constitutes one record of a user on a specific date. This step is analogous to the process of conventional questionnaire design and result collection. However, the data mining method can flexibly adjust the demographic characteristics that need to be analyzed, and acquire a stable amount of historical data from any population group during a long time period. This step contains 6 intelligent modules, which are described as follows.
This predictor, implemented with an open-source package of the M3 (multimodal, multilingual, and multiattribute) model [
The structure of the M3 (multimodal, multilingual, and multiattribute) model. DenseNet: dense convolutional network; ReLU: rectified linear unit.
We constructed a word-based convolutional neural network for occupation inference through user information (such as biographies) and tweet text. As shown in
The structure of the occupation predictor. Conv Block: convolution block; 1D-Conv(kernel_size=k): 1D convolution layer with a kernel size of k; 1D-MaxPool: 1D max-pooling layer; Flatten: flatten layer; OCi: the ith occupation category, i∈(1,3).
The new occupation categories in our study and the original occupation categories in the Standard Occupation Classification hierarchy.
New occupation category (OC) | Original occupation category |
OC1 | C2: professional occupations |
OC2 | C1: managers, directors, and senior officials |
OC3 | C4: administrative and secretarial occupations |
In this study, we only focused on the following 2 key attributes in cyberspace: account age and follower number. As a matter of fact, there are plenty of cyberspace attributes recorded in the tweet object, such as verified status and tweet number. The reason why we chose these 2 attributes is that account age can reflect the internet age of a user, and follower number can indicate a user’s influence and usage level of social media to a certain extent. Additionally, Lyu et al [
We used the “geo” field in the tweet object and the “location” field in the profile of a Twitter user to efficiently extract location information, including the continent, country, and state. The extraction process was implemented in 2 steps. First, the location extractor called the Twitter API to query a place by the geocode in the “geo” field [
The sentiment analyzer was implemented with an open-source tool named Valence Aware Dictionary and Emotional Reasoner (VADER) [
The emotion detector was based on an open-source emotion recognition algorithm [
Based on the mining outcomes, we conducted multiple analyses, including overall analysis, a longitudinal study, and a cross-sectional study, to detect the evolution and disparities of attitudes toward COVID-19 vaccines among population groups during the study period. The multiple analyses are shown in
Concretely, in the overall analysis, multiple linear regression and Pearson correlation analysis were used to detect the impacts of official COVID-19 statistics, including confirmed cases, deaths, vaccinations, and reproduction rate, and some major vaccine-related events on the whole population’s attentiveness toward vaccines. In the longitudinal study, to detect the evolution of attitudes among different population groups over time, a series of longitudinal contrasts with different demographic characteristics were displayed and analyzed. To further reveal the attitude patterns among population groups, a cross-sectional study, incorporating the benchmark of the population distribution of the general content tweets, was conducted at 5 vaccine-related events selected from the overall analysis.
The strength of the correlation and similarity in this paper using the guide that Evans [
In this section, we explore the possible influencing factors of the whole population’s attentiveness toward vaccines in 2 steps. First, studying the correlation between attentiveness and official COVID-19 statistics. Second, discovering the abrupt influences of some major vaccine-related events during COVID-19. At the end of this section, the data mining results of vaccine-related tweets and general content tweets have been displayed and analyzed in general.
In order to eliminate possible fluctuations in the quantities of tweets captured daily, we used the percentage of vaccine-related tweets in COVID-19 tweets to represent the attentiveness toward vaccines (sometimes referred to as attentiveness for short in the following text) during COVID-19 in this study. Meanwhile, we obtained the global data of COVID-19 statistics from Our World in Data [
The whole population’s attentiveness toward vaccines, and the COVID-19 statistics during the study period. The y-axis on the left is the level of attentiveness, and the y-axis on the right represents the numbers of COVID-19 statistics, which adopt different scales.
Since some variables of the COVID-19 statistics do not exert an immediate influence on the attentiveness toward vaccines, we applied Pearson correlation analysis on different time delays (0≤lag≤30 days) of attentiveness with each of the statistical variables, and found that the optimal lag at which Pearson
Pearson correlation coefficients between COVID-19 statistics and attentiveness toward vaccines with lags that have absolute maximum r values within 30 days.
Variable | Lag days | Pearson |
|
New cases | 5 | 0.5917 | <.001 |
New deaths | 3 | 0.6543 | <.001 |
New vaccinations | 4 | 0.7843 | <.001 |
Total cases | 0 | 0.9093 | <.001 |
Total deaths | 0 | 0.9066 | <.001 |
Total vaccinations | 0 | 0.7433 | <.001 |
People vaccinated | 0 | 0.7364 | <.001 |
Reproduction rate | 10 | −0.4562 | <.001 |
Then, we took the 8 variables of COVID-19 statistics as independent variables (denoted as X
In this formula,
By using the above regression model between the attentiveness Y(
The whole population’s attentiveness toward vaccines, estimated attentiveness by COVID-19 statistics, and labels of 5 major vaccine-related events.
Furthermore, we noticed that some abrupt jumps appeared in the attentiveness curve. An earlier study by Chen et al [
Some major events related to COVID-19 vaccines.
Time | Major event | Highest attentiveness percentage | Lasting days |
July 14, 2020 ( |
“An mRNA vaccine against SARS-CoV-2—preliminary report” [ |
4.27% | 14 |
August 12, 2020 ( |
“Phase I/II study of COVID-19 RNA vaccine BNT162b1 in adults” [ |
4.78% | 12 |
November 9, 2020 ( |
Pfizer and BioNTech announced phase III results [ |
9.78% | 6 |
December 15, 2020 ( |
The first mass vaccination program started globally [ |
17.48% | 21 |
April 10, 2021 ( |
More and more vaccine-related studies and news were reported (see Multimedia Appendix 1). | 18.92% | 46 |
The above data analysis shows that the total online population’s attentiveness toward COVID-19 vaccines was significantly correlated with COVID-19 statistics, including confirmed cases, deaths, vaccinations, and reproduction rate. Besides, attentiveness was also influenced by some vaccine-related events.
Two data mining experiments were conducted on 2 types of tweets as a comparison to infer different latent characteristics. The data mining methods are described in the Data Mining Step of the Methods section. The first experiment was performed on 1,197,763 vaccine-related tweets covering 1 year, and the second experiment was on 100,000 general content tweets selected from 5 major vaccine-related events (20,000 samples at each event) described in
Overall statistics of vaccine-related tweets and general content tweets.
Characteristic | Vaccine-related tweets (N=1,197,763) | General content tweets at 5 major time points (N=100,000) | ||
|
|
|
<.001 | |
|
Individual | 1,079,105 (90.09) | 94,560 (94.56) |
|
|
Organization | 118,658 (9.91) | 5440 (5.44) |
|
|
|
|
<.001 | |
|
Male | 661,511 (61.30) | 50,156 (53.04) |
|
|
Female | 417,594 (38.70) | 44,404 (46.96) |
|
|
|
|
<.001 | |
|
≤18 | 157,395 (14.59) | 35,036 (37.05) |
|
|
19-29 | 254,920 (23.62) | 32,142 (33.99) |
|
|
30-39 | 202,451 (18.76) | 11,112 (11.75) |
|
|
≥40 | 464,339 (43.03) | 16,270 (17.21) |
|
|
|
|
<.001 | |
|
OC1 | 385,276 (35.70) | 20,102 (21.26) |
|
|
OC2 | 347,257 (32.18) | 31,176 (32.97) |
|
|
OC3 | 346,572 (32.12) | 43,282 (45.77) |
|
|
|
|
<.001 | |
|
North America | 274,565 (40.12) | 21,202 (37.45) |
|
|
Europe | 179,683 (26.26) | 13,132 (23.20) |
|
|
Asia | 106,713 (15.60) | 12,388 (21.88) |
|
|
Africa | 67,805 (9.91) | 7320 (12.93) |
|
|
Oceania | 39,414 (5.76) | 1250 (2.21) |
|
|
South America | 15,794 (2.31) | 1292 (2.28) |
|
|
Antarctica | 301 (0.04) | 26 (0.05) |
|
|
|
|
<.001 | |
|
United States | 214,606 (31.36) | 18,163 (32.09) |
|
|
United Kingdom | 106,337 (15.54) | 6334 (11.19) |
|
|
India | 51,698 (7.56) | 2969 (5.24) |
|
|
Canada | 42,554 (6.22) | 1538 (2.72) |
|
|
Australia | 20,477 (2.99) | 474 (0.84) |
|
|
|
|
<.001 | |
|
<5 | 478,914 (44.38) | 58,341 (61.70) |
|
|
5-10 | 349,595 (32.40) | 26,333 (27.85) |
|
|
≥10 | 250,596 (23.22) | 9886 (10.45) |
|
|
|
|
<.001 | |
|
<500 | 619,808 (57.44) | 55,745 (58.95) |
|
|
500-5000 | 368,780 (34.17) | 32,742 (34.63) |
|
|
≥5000 | 90,517 (8.39) | 6073 (6.42) |
|
|
|
|
<.001 | |
|
Overall (–1 to 1), mean (SD) | 0.0161 (0.4591) | 0.1215 (0.4566) | <.001 |
|
Negative (–1 to −0.05), n (%) | 411,990 (38.18) | 41,555 (43.95) |
|
|
Neutral (–0.05 to 0.05), n (%) | 292,273 (27.08) | 29,987 (31.71) |
|
|
Positive (0.05 to 1), n (%) | 374,842 (34.74) | 23,018 (24.34) |
|
|
|
|
<.001 | |
|
Fear | 528,667 (48.99) | 21,703 (22.95) |
|
|
Joy | 313,423 (29.04) | 31,819 (33.65) |
|
|
Surprise | 153,807 (14.25) | 28,690 (30.34) |
|
|
Sadness | 42,124 (3.90) | 8940 (9.45) |
|
|
Anger | 23,814 (2.21) | 2570 (2.72) |
|
|
Disgust | 17,270 (1.60) | 838 (0.89) |
|
aUnder location (continent and country) characteristics, the total number of vaccine-related tweets with location information was 684,275, and the total number of general content tweets with location information was 56,610.
bThe top 5 countries with the most vaccine-related tweets are selected.
From
In the next 2 sections, we further analyzed these long-term and multicharacteristic data in longitudinal and cross-sectional studies.
In this section, we analyzed the attitude evolution of population groups from the following 2 aspects: attentiveness and sentiments toward vaccines.
Based on the data mining outcomes, we obtained the daily attentiveness toward vaccines of population groups by calculating the percentages of vaccine-related tweets in COVID-19 tweets with different characteristics. As shown in
Attentiveness toward vaccines of different population groups. The vertical dashed lines represent the time points of 5 major vaccine-related events as follows: July 14, 2020 (
Furthermore, except for 2 location characteristics, population groups under the 6 demographic characteristics exhibited consistent differences throughout the study period. For example, males always had a higher level of attentiveness than females. In contrast, there was no specific pattern for population groups under the 2 location characteristics. In particular, North America almost had the highest attentiveness among all continents during the pandemic, while Europe and Asia surpassed it in some periods. Moreover, the United States almost had the highest attentiveness among all countries, while the United Kingdom and India surpassed it in some periods.
In summary, most of the population groups had very strong similarities with the whole population regarding attentiveness, that is, attentiveness increased over time with some local peaks, indicating that they might be affected by the vaccine-related events as well. Moreover, there existed consistent group differences in the evolution of attentiveness under the demographic characteristics, except for the 2 location characteristics.
As shown in
Sentiments toward vaccines of different population groups. OC: occupation category.
In particular, the sentiments of organizations were more positive than individuals. Females were sometimes a bit less positive than males in the development phase of vaccines, but were more positive than males after the inoculation started. Among the 4 age groups, the sentiments of people aged ≥40 and ≤18 years were almost the lowest during the study period. Among the 3 categories of occupations, the sentiments of OC1 were the highest, while those of OC3 were always the lowest. Under the 2 location characteristics, South America exhibited a different sentiment trend compared with other continents. Asia among the continents and India among the countries had the highest sentiments, and both showed downward trends. Among the 3 account age groups, the group of <5 years almost had the lowest sentiments. Among the follower number groups, the group of <500 followers nearly had the lowest sentiments, while the group of ≥5000 followers had the highest sentiments, except for the period before
In summary, the sentiments differed among population groups and fluctuated a lot in the early period of vaccine development, which suggested that different populations might hold different and immature views at the beginning. After June 2021, there were downward trends in all populations, indicating that populations might become less positive toward vaccines than before.
In the previous section, we mainly focused on the long-term evolution of population attitudes toward vaccines among COVID-19–related tweets, ignoring the general population distribution. Actually, the sizes of population groups vary greatly with respect to population characteristics. Thus, it is meaningful and essential to investigate the attentiveness ratios toward vaccines among different population groups under the benchmark of the general population distribution. Therefore, in this section, we conducted 5 cross-sectional analyses at 5 major vaccine-related events, by applying the odds ratio (OR) to represent the attentiveness ratio toward vaccines of each population group. Due to the complexity of attentiveness among continents and countries under the 2 location characteristics, we only analyzed the 6 demographic characteristics.
As shown in
Attentiveness odds ratios toward vaccines at 5 major vaccine-related events.
Characteristic | Major vaccine-related events | ||||||||||
|
|
|
|
|
|
||||||
|
Individual | 1 (refb) | 1 (ref) | 1 (ref) | 1 (ref) | 1 (ref) | |||||
|
Organization | 1.51 (1.20-1.89) | 1.46 (1.22-1.76) | 2.01 (1.70-2.39) | 1.44 (1.28-1.61) | 1.98 (1.79-2.19) | |||||
|
|
|
|
|
|
||||||
|
Male | 1 (ref) | 1 (ref) | 1 (ref) | 1 (ref) | 1 (ref) | |||||
|
Female | 0.75 (0.66-0.86) | 0.79 (0.71-0.87) | 0.66 (0.60-0.73) | 0.82 (0.77-0.88) | 0.61 (0.57-0.66) | |||||
|
|
|
|
|
|
||||||
|
≤18 | 1 (ref) | 1 (ref) | 1 (ref) | 1 (ref) | 1 (ref) | |||||
|
19-29 | 1.68 (1.55-1.82) | 1.49 (1.42-1.56) | 1.85 (1.78-1.93) | 2.31 (2.23-2.39) | 2.13 (2.05-2.20) | |||||
|
30-39 | 3.55 (3.26-3.86) | 2.31 (2.19-2.44) | 4.48 (4.30-4.66) | 4.39 (4.23-4.55) | 5.22 (5.03-5.42) | |||||
|
≥40 | 5.26 (4.89-5.67) | 3.70 (3.53-3.87) | 7.34 (7.08-7.60) | 7.12 (6.90-7.35) | 8.81 (8.53-9.10) | |||||
|
|
|
|
|
|
||||||
|
OC1 | 1 (ref) | 1 (ref) | 1 (ref) | 1 (ref) | 1 (ref) | |||||
|
OC2 | 0.52 (0.44-0.63) | 0.62 (0.55-0.70) | 0.66 (0.61-0.72) | 0.52 (0.48-0.57) | 0.59 (0.55-0.64) | |||||
|
OC3 | 0.37 (0.30-0.44) | 0.44 (0.39-0.49) | 0.42 (0.39-0.46) | 0.41 (0.38-0.45) | 0.42 (0.39-0.45) | |||||
|
|
|
|
|
|
||||||
|
<5 | 1 (ref) | 1 (ref) | 1 (ref) | 1 (ref) | 1 (ref) | |||||
|
5-10 | 1.31 (1.10-1.55) | 1.34 (1.20-1.49) | 1.67 (1.54-1.81) | 2.28 (2.12-2.46) | 2.25 (2.08-2.44) | |||||
|
≥10 | 1.93 (1.54-2.40) | 1.62 (1.40-1.87) | 2.88 (2.62-3.17) | 3.71 (3.40-4.04) | 3.28 (3.01-3.58) | |||||
|
|
|
|
|
|
||||||
|
<500 | 1 (ref) | 1 (ref) | 1 (ref) | 1 (ref) | 1 (ref) | |||||
|
500-5000 | 0.97 (0.82-1.14) | 1.02 (0.92-1.13) | 1.04 (0.96-1.13) | 1.07 (0.99-1.14) | 1.02 (0.95-1.10) | |||||
|
≥5000 | 1.32 (1.00-1.74) | 1.20 (1.00-1.44) | 1.69 (1.48-1.92) | 1.43 (1.27-1.61) | 1.40 (1.23-1.59) |
aOR: odds ratio.
bref: reference.
To further discern the deep law of attitudes under different demographic characteristics after the inoculation started, we crossed the 2 dimensions of attitude (the attentiveness ratio and sentiment polarity) according to the cross-sectional results at
Four categories of population groups by crossing the 2 dimensions of attitude toward vaccines. OC: occupation category.
From
In this study, we acquired and analyzed a year-long collection of tweets, from June 9, 2020, to July 31, 2021, to discover the evolution and disparities of attitudes toward COVID-19 vaccines among various online population groups. Overall, the whole population’s attentiveness toward COVID-19 vaccines increased over time with some local fluctuations during the study period. This study demonstrated that this attentiveness had a very strong correlation (Pearson
By analyzing the attentiveness evolution toward vaccines under 8 demographic characteristics, we observed that, except for the United States, Australia, and India, all population groups exhibited very strong similarities with the whole population. As for the sentiment evolution toward vaccines, we found that different populations initially held different and fluctuated sentiments in the early stage of vaccine development, and then, the sentiments gradually stabilized and tended to relatively positive levels at the beginning of vaccination, but after June 2021, they all had a considerably downward trend. The research findings of Yan et al [
Furthermore, there are significant attitude disparities toward COVID-19 vaccines across population groups. By crossing the 2 dimensions of attitude (the attentiveness ratio and sentiment polarity), we found that among population groups carrying low sentiments, some have low attentiveness ratios (the 1st category in
We investigated and inferred the internal reasons for the low sentiments, and found some corresponding epidemiological studies that can confirm our inference. For the 1st category of the population, the low sentiments may be derived from the insufficient knowledge and distrust of vaccines based on education status, news awareness, economic conditions, level of social media usage, etc. This finding appears consistent with the finding of Paul et al [
Overall, only paying excess attention blindly cannot effectively allay the diverse public concerns and fears about vaccines. Specialized interventions should be implemented to address these concerns raised by different populations. Some studies in the field of public health are worthy of reference. For example, Brooke et al [
Due to the high complexity of multilingual analysis and the insufficient support for detecting the various characteristics of the population groups, this paper only extracted data in English and key demographic characteristics for analysis. In addition, considering that too many OCs may reduce the accuracy of the occupation predictor, we only divided the occupations into 3 categories, which may have resulted in the loss of some fine-grained information. Despite these limitations, it did not affect the overall findings.
By analyzing large-scale tweets during vaccine development and vaccination, this study tracked the year-long evolution of attitudes toward COVID-19 vaccines among population groups, and offered rich evidence to gain insights about the attitude patterns of the international population on social media. Through well-organized approaches, governments, public health agencies, health care providers, and influential Twitter users can work together to help those populations with low sentiments get through this difficult period. At last, it is worth mentioning that the method applied in this paper can be easily extended to other public health events for multidimensional and large-scale research on the long-term evolution of human responses. The source code developed in this study is available for use at GitHub [
Supplemental information about vaccine-related news between March and April 2021.
Supplemental information about the sentiments toward vaccines at 5 major vaccine-related events.
application programming interface
multimodal, multilingual, and multiattribute
occupation category
odds ratio
Standard Occupation Classification
Valence Aware Dictionary and Emotional Reasoner
This research was supported by the National Natural Science Foundation of China (grants 61876150 and 12026609), the Key Research and Development Program of Shaanxi Province (grant number 2021SF-188), the Science and Technology Program of the City of Xi'an (grant number 20YXYJ0009-12), and the Youth Fund of the Second Affiliated Hospital of Xi'an Jiaotong University (grant number YJ(QN)202004).
None declared.