Exploring Occupation Differences in Reactions to COVID-19 Pandemic on Twitter.

Coronavirus disease 2019 (COVID-19) pandemic-related information are flooded on social media, and analyzing this information from an occupational perspective can help us to understand the social implications of this unprecedented disruption. In this study, using a COVID-19-related dataset collected with the Twitter IDs, we conduct topic and sentiment analysis from the perspective of occupation, by leveraging Latent Dirichlet Allocation (LDA) topic modeling and Valence Aware Dictionary and sEntiment Reasoning (VADER) model, respectively. The experimental results indicate that there are significant topic preference differences between Twitter users with different occupations. However, occupation-linked affective differences are only partly demonstrated in our study; Twitter users with different income levels have nothing to do with sentiment expression on covid-19-related topics.

Introduction

The outbreak of coronavirus Disease 2019 (COVID-19) is rapidly spreading worldwide and causing a profound effect on various aspects of society. At the time of writing, Johns Hopkins University reported more than 21,056,850 confirmed cases of COVID-19 globally, including 762,293 deaths and 13,100,902 recovered.1 To slow down the spread of virus, social distancing and self-isolation have been implemented globally, and social media has become an important channel for people to post their opinions and attitudes. The outbreak of the COVID-19 pandemic caused a rapid increase in COVID-19-related information on social media platforms, including YouTube, Facebook, Twitter, etc. (Abd-Alrazaq, Alhuwail, Househ, Hamdi, & Shah., 2020). Extensive research has shown that social media is a popular source of data in understanding public concerns and attitudes, and it is an important way to support crisis communication between the public and the government (Jordan et al., 2018; Shah & Dunn, 2019).

Occupation is one of the noteworthy demographic variables of Twitter users (Sloan, Morgan, Burnap, & Williams, 2015). In a recent study, Kern, McCarthy, Chakrabarty, and Rizoiu (2019) automatically assessed the personality of different occupations based on the tweets and found that personality was related to distinctive occupations. In addition, several studies have found an association between personality and perceptions of the COVID-19 situation (Carvalho, Pianowski, & Goncalves, 2020; Zajenkowski, Jonason, Leniarska, & Kozakiewicz, 2020). The above research may imply that Twitter users engaged in different occupations may have different topic preferences and sentiment expressions, when faced with COVID-19 pandemic. This motivates us to provide a comprehensive analysis based on Twitter data from the perspective of occupation differences, which helps policymakers understand the fine-grained public concerns.

Previous research has been focused on gender-linked affective or topic differences in social media platforms (Thelwall & Thelwall, 2020; Thelwall & Vis, 2017; Vegt & Kleinberg, 2020). Few works have been devoted to occupational differences in response to COVID-19, and the research mainly focuses on a particular occupation, including world leaders (Rufai & Bunce, 2020) and college students (Duong, Luo, Pham, Yang, & Wang, 2020). Additionally, the income level may be another factor that affects topic preference and sentiment expression. The purpose of this study is to demonstrate that Twitter users with different occupation respond differently to COVID-19 and different income levels affect topic preferences and sentiment expressions.

Related Works

When it comes to discovering topics from Twitter, a majority of studies utilize the LDA technique and its improved versions. Abd-Alrazaq et al. (2020) utilized LDA to identify the main topic in the language of COVID-19 tweets. Guo, Vargo, Pan, Ding, and Ishwar (2016) conducted empirical research on two Twitter datasets for the 2012 presidential election, and they concluded that the LDA generated meaningful results. A lot of extended version of the LDA was used for topic analysis on social media platforms. N-gram LDA technique was used by E. H.-J. Kim, Jeong, Kim, Kang, and Song. (2016) to investigate the topic coverage of Twitter and news publications about the Ebola virus. Yan, Guo, Lan, and Cheng (2013) proposed a bi-term topic model to capture the word co-occurrence directly and enhance the topic quality. Sasaki, Yoshikawa, and Furuhashi (2014) implemented a Twitter–TTM model based on the topic tracking model that is competent in online inference, and the model can capture the dynamic topic trends on Twitter. To improve the interpretation of the topic model, especially for short text documents, Alkhodair, Fung, Rahman, and Hung (2018) developed a new model that combined Twitter–LDA, WordNet, and hashtags and assessed the effectiveness of the model against two real Twitter datasets.

Sentiment analysis for Twitter textual data is also a hot topic. Two main methods have been utilized to classify the sentiment polarity: supervised machine learning method and rule-based method (Hutto & Gilbert, 2014; Kim & Hovy, 2014). Tang et al. (2014) developed a new method that learned sentiment from specific word embedding and outperformed the previous top-performing system. Ren, Wang, and Ji (2016) used SVM for sentiment classification, and LDA was adopted for the improvement of word embedding. Recently, a large pretrained model was adopted for the classification task. BERT and RoBERTa were used by Duong et al. (2020) to examine the Twitter user's sentiment expression. Supervised machine learning methods are powerful for sentiment classification, but large numbers of annotation data are needed for model training, which is costly and time consuming. Rule-based methods, including sentiwordnet (Esuli & Sebastiani, 2006) and Valence Aware Dictionary and sEntiment Reasoning (VADER) (Hutto & Gilbert, 2014), are also widely used on Twitter. Pandarachalil, Sendhilkumar, and Mahalakshmi (2014) proposed an unsupervised method that incorporated SenticNet, SentiWordNet, and SentislangNet, and it performed well on large scales. Chaithra (2019) presented a hybrid approach combining VADER and Naive Bayes to use the comments of the mobile unboxing videos to predict the sentiment, and the result concluded that VADER has improved the performance of the Naive Bayes classifier in predicting the sentiment of the comments.

To our knowledge, no previous study that adopted topic modeling and sentiment analysis to investigate the occupation distinctions in reactions to COVID-19 pandemic, which will be addressed in this study. As a result of a trade-off between cost and accuracy, LDA and VADER methods are applied in the study.

Material and Methods

Data Collection

Using the Twitter IDs provided by Chen, Lerman, and Ferrara (2020) on the repository,2 we collected 8716,289 tweets from May 1 to 15, 2020. In this study, we focus only on the English-language tweets and Twitter user populations of those who can identify occupational information from their self-description field. Since the self-description field is an open text in which Twitter users can choose to write whatever they like, it's essential to design an appropriate method for occupation extraction. Inspired by Kern et al. (2019), regular expression matching was used to extract occupation information and we obtained 15,984 job titles from their research. The difference is that the job titles are aligned into major occupation groups according to the O*net3 alternative titles data.4 For example, the Assignment Editor and Morning News Producer are two job titles that belong to the Producers and Directors group. Furthermore, only the major occupation for each Twitter user was reserved during the process of occupation extraction. In our subsequent studies, we only considered the major occupation group as a research object. After removing retweets and filtering out corporate Twitter users with more than 50 tweets, we acquired 622,687 unique COVID-19-related tweets, belonging to 373,773 users, representing 800 occupations.

Due to the diversity of occupational categories and the restrictions of space, we selected nine occupations from different income levels. The classification criteria on occupation type and income levels are lacking, so we incorporate the salary data from the Bureau of Labor Statistics5 and information from Wall Street News (Suneson, 2019) to generate the following classification standard: Occupations in the high-income level group with incomes over $100,000 per year; occupations in the medium-income level group with incomes $30,000 to $100,000 per year; occupations in the low-income level group with incomes less than $30,000 per year. The nine occupations encompass a variety of social occupations, including technical occupations, managerial occupations, service occupations, etc. Of these nine occupations, the smallest one included 1,121 unique Twitter users. For balance, we randomly sampled 1,121 unique Twitter users from each occupation. The selected occupations are shown in Table 1 .

Table 1

The Distribution of Occupations of Twitter Users

Income Level	Occupation	Occupation Abbreviations	Number of tweets
High	Computer and Information Research Scientists	CIRS	1,652
	Marketing Managers	MM	1,700
	Dentists, General	DEN	1,835
Medium	Management Analysts	MA	1,909
	Business Teachers, Postsecondary	BTP	1,736
	Financial Analysts	FA	1,841
Low	Farmworkers, Farm, Ranch, and Aquacultural Animals	FFRAA	1,876
	Production Workers, All Other	PW	1,750
	Landscaping and Groundskeeping Workers	LGW	1,988

Method

Many researchers have utilized Latent Dirichlet Allocation (LDA) to identify topics from social media data (Alkhodair et al., 2018; Asghari, Sierra-Sosa, & Elmaghraby, 2020; Giannetti, 2018). LDA is a three-level hierarchical Bayesian model (Blei, Ng, & Jordan, 2003), it is an unsupervised machine learning technique used to create a representation of documents by topic, where each topic consisted of a set of words. In this study, we employed an LDA algorithm from the Python Gensim library.6

To obtain a clean corpus, we conducted data preprocessing at first. Then, we used regular expressions to remove URLs, HTML tags, and Twitter user mentions. Next, we also removed punctuation, stop words, and nonprintable characters from tweets. Finally, all tweets are lower-cased, tokenized, and lemmatized. It is well known that phrases are more meaningful than individual tokens. Hence bigrams and trigrams are created and added to the corpus. Before the corpus was fed into the LDA model (Blei et al., 2003), we used two kinds of document representation methods, bag of words (BOW) and term frequency–inverse document frequency (TF-IDF) (Salton & Yu, 1975), to represent tweets. The number of topics that need to be determined before running the LDA and it is a hyperparameter, so CV coherence measure is used to fine-tune our LDA topic model to obtain the optimal topic number (Röder, Both, & Hinneburg, 2015).

Sentiment analysis was also performed in our study. Sentiment analysis was conducted on the cleaned tweets using the VADER, a lexicon- and rule-based sentiment analysis model for social media text (Hutto & Gilbert, 2014). Hutto and Gilbert (2014) compared VADER with multiple methods, including Affective Norms for English Words (ANEW),7 Linguistic Inquiry and Word Count (LIWC),8 SentiWordNet (Esuli & Sebastiani, 2006), the General Inquirer,9 and machine learning techniques based on SVM, Naïve Bayes, Maximum Entropy, and the results show that the VADER model that is superior to these methods. One significant advantage of using VADER is that it can quickly and accurately obtain the sentiment of each tweet on such a large scale. The sentiment scores ranged from −1 to +1, with −1 as the most negative sentiment and 1 as the most positive sentiment; furthermore, when the score is between −0.05 and +0.05, it means neutral sentiment. The sentiment of each tweet was calculated using the library “vaderSentiment” in Python,10 which is the implementation of the python version of VADER model.

Experiment and Result

Selection of the Optimal Number of Topics

All 622,687 tweets were fed into the LDA model, we repeated the experiment on a different number of topics (ranging from 2 to 32) and reported the coherence value for two document representation methods. As shown in Figure 1 , BOW performed better than TF-IDF, and the coherence value selected 14 as the optimal number of topics. Since the words of Topics 3, 5, 7, 8, and 13 are difficult to assign a specific theme, the authors reached a consensus on selecting nine highly relevant topics as the final result, as shown in Table 2 .

Figure 1

Coherence measurement for bag of words (BOW) and term frequency–inverse document frequency (TF-IDF).

Table 2

Topics in COVID-19-Associated Tweets

Topic	Themes	Top 15 most relevant words	Number of tweets
0	Preparation for reopening	social_distancing, reopen, open, school, take, people, back, life, place, economy, say, child, close, measure, guideline	35,336
1	President's lies about COVID-19	trump, american, say, president, response, cdc, america, death, lie, white_house, claim, government, administration, medium, covid	40,838
2	Coronavirus new cases and deaths	case, death, test, new, report, number, positive, total, covid, update, day, state, rate, high, data	50,863
3	*	get, see, come, work, covid, back, hard, time, like, hit, interesting, one, since, bit, well	35,531
4	Free online support	help, support, need, crisis, student, work, time, learn, online, community, free, join, impact, new, team	49,828
5	*	people, know, get, think, covid, like, make, would, good, thing, one, say, die, want, bad	107,280
6	Protests against the stay-at-home order	state, fuck, vote, governor, house, party, game, get, man, right, play, bbc_news, stay-at-home_order, shit, protest	25,435
7	*	day, love, read, great, time, watch, video, one, today, story, thank, good, new, share, stayhome	62,833
8	*	may, week, month, next, plan, ease, due, year, end, last, start, two, restriction, thread, day	29,162
9	The risk caused by COVID-19	health, business, worker, pay, care, government, public, work, need, risk, job, service, food, staff, help	54,488
10	Measures to slow the spread of COVID-19	home, stay, people, work, safe, mask, keep, get, wear_mask, order, face, family, back, protect, life	48,237
11	Research on the vaccine and treatment	vaccine, world, virus, patient, covid, use, disease, could, doctor, study, find, spread, cure, treatment, people	39,989
12	Virus misinformation and fake news	corona, virus, china, india, fight, wuhan, chinese, world, spread, lab, come, country, war, outbreak, govt	21,210
13	*	like, look, question, year, well, say, answer, ask, science, london, wow, would, could, break, first	21,657

Note: * indicates that it is difficult to assign a specific theme to the topic and will not be analyzed in a subsequent section.

As shown in Table 2, we assigned the potential themes to nine topics based on the top 15 most relevant words. Topic 0 discussed the preparation for reopening, with words such as “social distancing,” “reopen,” “school,” and “guideline” indicating this overarching theme. Topic 1 was primarily about U.S. President Donald Trump's lies about the COVID-19. Topic 2 was about the coronavirus new cases and deaths, which were identified in tweets that mentioned the rapid growth in the number of confirmed cases. Topic 4 talked about free online support, and Twitter users on the topic mainly discussed joining a local free online support team to provide useful pandemic-related information. A potential theme identified in Topic 6 was protests against the stay-at-home order, with the words “fuck” and “shit” showing the feelings for the order. Topic 9 primarily referred to the risk caused by COVID-19, which can affect business, jobs and food. Topic 10 was measures to slow the spread of COVID-19, including wearing masks and staying at home. The theme of research on the vaccine and treatment was identified in topic 11. Topic 12 was related to virus misinformation and fake news, and Twitter users debated whether the novel coronavirus originated in Wuhan's laboratory and spread around the world.

Distribution of Topics among Different Occupations

Figure 2 shows the topic distribution of Twitter users engaged in different occupations. Topics 3, 5, 7, 8, and 13 are grouped into a separate group (other Topics).

Figure 2

Topic distribution of Twitter users engaged in different occupations.

Overall, there is a significant difference in topic concern between Twitter users with different occupations, while Twitter users at different income levels showed a slight tendency toward some topics. For high-income level groups, Topic 2 is more concerned with Computer and Information Research Scientists (CIRS) and Dentists, General (DEN) and Marketing Managers (MM) cares more about Topics 4 and 9, and it indicates the topic preference for Twitter users with difference occupations. In addition, Topic 4 also attracts more attention to high-income and medium-income occupations than to low-income occupations. Twitter users in low-income occupations are more interested in Topic 1 than Twitter users in high-income and medium-income occupations. Compared with other topics, all of Twitter users who engaged in these nine occupations showed less curiosity about the virus misinformation and fake news (Topic 12).

Topic-based Sentiment Analysis among Different Occupations

To understand occupation differences in sentimental responses to COVID-19, we analyze the sentiment distribution toward the different topics in Figure 3, Figure 4 .

Figure 3

Sentiment distribution on the topics 0, 1, 2, and 4.

Figure 4

Sentiment distribution on the topics 6, 9, 10, 11, and 12.

Overall, Twitter users in different occupations express different sentiments on different topics, but the sentiment expressed by Twitter users of different occupations toward a topic seems to have nothing to do with their income levels. Overall sentiment trends are almost identical for Twitter users with different occupations for most of the topics. For example, regardless of the Twitter user's occupation, negative sentiment ratio is greater than positive sentiment ratio toward Topic 1 (President's lies about the COVID-19), that is, U.S. President Donald Trump admitted to journalist Bob Woodward that he played down the severity of the pandemic11 ; the number of negative tweets posted on Twitter on this topic is very alarming. Moreover, online support groups offered medical support to the public during the COVID-19 pandemic, enhancing the public's ability to self-protection (Gong, Xu, Cai, Chen, & Wang, 2020), and thus about 58% of tweets positively responded to the Topic 4 (free online support). At the same time, differences exist in emotional expression toward Topic 2, that is, CIRS, MM, and DEN respond more positively sentiment toward the coronavirus new cases and deaths, whereas, Financial Analysts (FA), Farmworkers, Farm, Ranch, and Aquacultural Animals (FFRAA), Production Workers (PW), and Landscaping and Groundskeeping Workers (LGW) are more likely to express negative feelings on this topic. Additionally, for Topic 6, the positive sentiment ratio of CIRS and LGW is a little higher than that of other occupations, and MA tends to be subject to less positive sentiment on the topic of protesting against the stay-at-home order.

Conclusion and Future Works

In this study, we collected 622,687 tweets from 800 occupations and selected nine occupations with different income levels as topic for the research. We found that there was a significant difference in topic concern between Twitter users with different occupations, but sentiment expression differences only existed toward Topic 2, and income level seems to have nothing to do with emotional expression. These findings only partly demonstrate the hypothesis that Twitter users engaged in different occupations have different sentiment expressions. Furthermore, our study finds significant occupation differences in topic preference.

As a short paper, there are also a few limitations. First, we only compare nine occupations, and since the sample size is not big enough, it is necessary to investigate all the occupations and may provide more convincing results. Second, supervised machine learning can apply to topic discovery and sentiment analysis to improve the topic quality and classification accuracy, it may provide a better analytical result.

There are some related studies that can be implemented in the future. First, this article mainly focuses on sentiment differences toward topics among groups of Twitter users clustered by occupations; a more nuanced exploration should be conducted on emotional difference, which could provide us with an in-depth understanding of people's actions. Second, the topics posted on Twitter have been continuously changing with the development of the COVID-19 pandemic, so merely examining the topic preferences of Twitter users for a given period of time may be insufficient. Future researchers can therefore conduct the research throughout the lifetime of the pandemic and explore that whether the phenomena of topic preferences of Twitter users with different occupations still exist.