Bilingual Emotion Analysis on Social Media throughout the COVID19

Pandemic in Portugal

Alina Trifan

, S

ergio Matos

, Pedro Morgado

and Jos

e Lu

ıs Oliveira

DETI/IEETA, University of Aveiro, Portugal

School of Medicine, University of Minho, Portugal

Keywords:

Sentiment Analysis, Social Media, COVID-19.

Abstract:

This paper presents preliminary work on the topic of emotion analysis on Twitter, in the context of the coro-

navirus pandemic in Portugal. We collected, curated and analyzed covid-related tweets of users in Portugal

in order to understand the evolution of the six basic emotions reﬂected in these tweets. We analyzed tweets

written in both English and Portuguese. In this ﬁrst step of our work we correlate this information with key

events of the evolution of the pandemic in Portugal during March, which was the most critical period in Por-

tugal. We do so in an attempt to estimate the online manifestation of the psychological toll that this pandemic

has on the overall well-being status of the general population. Our ﬁndings show that the sentiment analysis

of covid-related tweets is consistent with our hypothesis that negative emotions would intensify as the pan-

demic progressed. The preliminary results obtained stand as proof of concept that the analysis of real-time

tweets or other social media messages through sentiment analysis can be an important tool for behavioural and

well-being tracking.

1 INTRODUCTION

The focus of attention of health care providers around

the world for the last several months has been the

problem of the new coronavirus (COVID-19) and its

spread. In addition to efforts at various levels to pre-

vent the spread of the disease and other worrisome

conditions, special attention should be paid to men-

tal health and care. According to similar epidemics

and pandemics, in such cases, serious concerns such

as fear of death can arise among patients, and feelings

of loneliness and insecurity can develop. Moreover,

people who are quarantined lose face-to-face connec-

tions and traditional social interventions, which sig-

niﬁcantly lowers personal and mental well-being.

We present in this short paper preliminary work

on the analysis of social media posts for inferring

the prevalence and evolution of the six basic human

emotions in Portugal, during the ﬁrst four and a half

months of the pandemic. Our hypothesis is that neg-

ative emotions, and consequently a decrease in per-

sonal well-being and possibly mental health, would

be predominant and that their intensity would fol-

low the evolution of the COVID-19 cases in Portugal.

We are interested in understanding this evolution and

performing sentiment analysis as an initial step of a

broader goal of monitoring mental health and well-

being among social networks users in these challeng-

ing times. This paper is structured in ﬁve more sec-

tions. We shortly discuss next the current background

in sentiment analysis, mental health and well-being

monitoring in the context of COVID-19 pandemic.

In Section 3 we detail the process of data collection

and curation. We present our analysis methodology

in Section 4 and the results obtained so far in Sec-

tion 5. Finally, we conclude the paper and discuss

future research steps in Section 6.

2 BACKGROUND

In the current context, global attention has largely

been focused on the infected patients and the front-

line responders, with some marginalised populations

in society having been overlooked (Ho et al., 2020).

Previous research has revealed a profound and broad

spectrum of psychological impact that outbreaks can

inﬂict on people (Hall et al., 2008; M

uller, 2014).

Several recent publications in the area of mental

health try to raise alarm ﬂags with respect to the im-

portance of ﬁnding new methods for responding to

430

Trifan, A., Matos, S., Morgado, P. and Oliveira, J.

Bilingual Emotion Analysis on Social Media throughout the COVID19 Pandemic in Portugal.

DOI: 10.5220/0010244204300434

In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 5: HEALTHINF, pages 430-434

ISBN: 978-989-758-490-9

mental health needs during the pandemic time. Xiang

et al. (Xiang et al., 2020) claim that the mental health

needs of patients with conﬁrmed COVID-19, patients

with suspected infection, quarantined family mem-

bers, and medical personnel have been poorly han-

dled throughout this time. Similarly, Yao et al. (Yao

et al., 2020) express their concerns with regards to

the effect of the pandemic on people with mental

health disorders. The obvious solution for Wind et

al. (Wind et al., 2020) to continue mental health care

within a pandemic is to provide mental health care at

a warm distance by video-conferencing psychother-

apy and internet interventions. This idea is further

supported by a study among Chinese citizens (Gao

et al., 2020), in which Gao et al. assess the prevalence

of mental health problems and examine their associa-

tion with social media exposure. Their ﬁndings high-

light that there is a link between the two, which sug-

gests that the government need to pay more attention

to mental health problems and their online manifesta-

tion. A portuguese study also found that individuals

previously receiving psychotherapeutic support ben-

eﬁt if they did not interrupt the process as a conse-

quence of the outbreak (Moreira et al., 2020).

The already available scientiﬁc evidence previ-

ously introduced strongly suggests that a shift in men-

tal health care provision towards online prevention

and treatment in the near future is needed. As such,

the widespread use of social media combined with

the rapid development of computational infrastruc-

tures to support big data, and the maturation of natural

language processing and machine learning technolo-

gies offer exciting possibilities for the improvement

of both population-level and individual-level mental

health (Conway and O’Connor, 2016).

3 DATA COLLECTION

This work builds on prior research contributions, both

national and international, that enabled the collection

of both social media and clinical statistics data. For

social media data, we collected four and a half months

of Twitter

posts whose content contained coron-

avirus related vocabulary, as explained next. The data

collection process followed the pipeline recently pub-

lished by Chen et. al (Chen et al., 2020). According to

Twitter’s Terms and Conditions, Chen et. al have re-

leased Tweet IDs, which are unique identiﬁers tied to

speciﬁc tweets containing coronavirus related terms.

The collection of ids is available on GitHub

www.twitter.com

https://github.com/echen102/COVID-19-TweetIDs

Based on this dataset, we queried the Twitter API

and obtained the complete dataset. For each tweet

in the collection we retrieved tweet content (text,

URLs, hashtags) and authors’ metadata. Follow-

ing authors’ suggestions, we used Hydrator

, Twit-

ter’s search API

and Twarc

for retrieving the data.

All tweets included in this collection contain coron-

avirus related vocabulary, both as hashtags or men-

tions/keywords in the text of the tweet. We collected

tweets posted from 21st of January to 31st of May,

2020. More details on the collection process can be

found in the original publication (Chen et al., 2020).

We ﬁltered the collected tweets based on the

tweet’s geo-location or user’s location. When users

post tweets from a GPS-enabled device or they tag

a location in their tweet, this information can be re-

trieved as longitude and latitudine coordinates. Un-

fortunately, only roughly 1–3% of Twitter messages

are geocoded (Paul and Dredze, 2017). Because the

number of geo-located tweets in the corpora we col-

lected is indeed limited, we used the location de-

ﬁned on the user’s proﬁle for extracting tweets of Por-

tuguese users. Using Tweepy

we ﬁlter for user loca-

tions that match a dictionary of city names in Por-

tugal. We ﬁlter out false positives by removing lo-

cations that also match names of cities or states in

other countries, namely Brazil (e.g. a match for Porto

Alegre, a city in Brazil, replaces a match for the Por-

tuguese city Porto). We ended up with a list of 76898

distinct users and 117772 distinct tweets.

For clinical statistics and the identiﬁcation of key

events of the manifestation of the COVID-19 pan-

demic in Portugal we relied on an open-source re-

spository maintained by Data Science for Social Good

Portugal

, an open community of data scientists that

tackle relevant and current societal issues. This repos-

itory contains daily updates of the statistic informa-

tion released by the Portuguese Ministery of Health

regarding COVID-19 cases in Portugal.

4 DATA ANALYSIS

Psychologist Paul Eckman identiﬁed six basic emo-

tions that he suggested were universally experienced

in all human cultures. These emotions were happiness

(joy), sadness, disgust, fear, surprise, and anger (Ek-

man, 1992). Happiness is often classiﬁed as a posi-

https://github.com/DocNow/hydrator

https://developer.twitter.com/en/docs/tweets/search/api-

reference/get-search-tweets

https://github.com/DocNow/twarc

https://www.tweepy.org/

https://github.com/dssg-pt/covid19pt-data

Bilingual Emotion Analysis on Social Media throughout the COVID19 Pandemic in Portugal

431

tive emotion, while the remaining ﬁve basic emotions

are classiﬁed as negative. We studied the basic emo-

tions reﬂected in Portuguese user tweets at the word

level, using Natural Language Processing (NLP), both

in English and in Portuguese. Words with basic dis-

crete emotions were tabulated, counted and their fre-

quency was measured.

We ﬁrst ﬁltered tweets by language and addressed

only tweets written in English or Portuguese. In the

preprocessing step we performed tokenization, stem-

ming and lemmatization, lowercase conversion and

stopwords removal for each of the two languages.

To this purpose we chose Python

as a programming

language and the Natural Language Toolkit

as NLP

framework. For the tweets written in English, we

used Empath (Fast et al., 2016), a lexicon mined from

modern text on the web, using a combination of deep

learning and crowdsourcing, to retrieve the counts and

frequencies of the six basic emotions. Each of these

emotions corresponds to an Empath lexical category,

which is formed by the emotion word and a collection

of other similar words that convey the same emotion.

For the tweets written in Portuguese, a psychiatrist

member of our research team deﬁned a semantic clus-

ter of Portuguese words that are associated to each of

the six basic emotions. These semantic clusters are

presented in Table 1. For both languages, we used the

lexicon words’ frequency in a tweet to measure the

strength of a speciﬁc emotion. For example, if two

words in a tweet were coded as joy or two occurences

of terms belonging to the joy semantic cluster, and

one word was coded as fear, joy was counted twice

and fear was counted once. These counts were then

normalized with respect to the total number of tokens

in a tweet.

5 RESULTS

We present preliminary results on the sentiment anal-

ysis of the evolution of the six basic emotion of Por-

tuguese Twitter users, along with statistical informa-

tion regarding the corpus that we curated. We con-

sider this curation of tweets written by Portuguese

users an important scientiﬁc contribution by itself.

We will publish the list of tweets ids, annotated with

the language in which the tweet was written, as open

source upon paper acceptance. We want, on one hand,

to encourage further exploration of this data for gain-

ing general well-being insights in Portugal and on the

other hand, to serve as possible comparation base for

www.python.org

https://www.nltk.org/

Table 1: Portuguese language semantic clusters deﬁned for

each of the six basic emotions.

Emotion Semantic cluster

Joy (alegria) feliz; felicidade; es-

petacular; esperanc¸a;

expetativa; fant

astico;

wow; alegria

Sadness (tristeza) triste; deprimido; de-

pressivo; tristeza

Disgust (nojo) nojo; contaminac¸

ao; re-

pulsa; cont

agio

Fear (medo) medo; ansioso; preocu-

pado; apreensivo; ner-

voso;

Surprise (surpresa) surpreendido; surpresa;

inesperado

Anger (raiva) wtf; merda; pqp; fdx;

revoltado; zangado; irri-

tado; enervado

Table 2: Number of tweets per month, global numbers and

tweet counts discriminated by language - Portuguese (PT),

English (EN) and other languages. It is important to note

that in January we only retrieved posts from the 21st of Jan-

uary to 31st of January.

Month # PT # EN # Other

January 2679 2262 79

February 17277 3312 1175

March 29158 11663 1745

April 10022 6401 1006

May 12407 5733 941

similar studies going on in other countries.

Table 2 overviews the number of tweets posted by

Portuguese users, by month, from 21st of January to

31st of May. For each month, we indicate the num-

ber of tweets written in Portuguese and in English,

as well as the total number of tweets written in other

languages. Among the tweets written in other lan-

guages than Portuguese and English, Spanish, Franch

and Italian were the most predominant ones.

Just by looking at the total number of tweets per

month we can see that March was the most critical

month of the pandemic in Portugal. The ﬁrst COVID-

19 positive cases were conﬁrmed by the end of Febru-

ary but in a small number. Partial lockdown started in

Portugal by mid-March and it slowly progressed into

total lockdown by the end of the month. The num-

ber of tweets related to COVID-19 published in this

period speaks for the increased interest and concern

that this pandemic had during the month of March.

Another important remark is that the ratio between

the number of tweets written in Portuguese and the

number of tweets written in English during February

HEALTHINF 2021 - 14th International Conference on Health Informatics

432

and March is superior to the ones in April and May,

which might suggest an increased focus on the de-

velopment of the national COVID-19 situation, rather

than worldwide. The main hypothesis of the present

study is that the tweets collected would reﬂect more

negative emotions throughout the pandemic, up to its

peak, which in Portugal it is estimated to have hap-

pened around 23rd-25th of March 2020. The evolu-

tion of COVID-19 cases in Portugal, both in terms of

infected population and number of suspected cases is

presented in Fig. 1.

Figure 1: COVID-19 cases in Portugal throughout the

month of March.

We examined the patterns of the tweets and the

evolution of the basic emotions to see whether they

were consistent with our expectations. We performed

this analysis for all the tweets collected so far. How-

ever, due to the late-breaking nature of this research,

we will focus here on the month of March, the most

critical one so far in Portugal.

We present the number of tweets per day written

by Portuguese users during the month of March in

Fig. 2. Figure 3 shows the normalized evolution of the

six basic emotions in covid-related tweets written in

English by Portuguese users over the month of March

2020. As we can see, the negative emotions are pre-

dominant with little joy or positive sentiment present

in these tweets. Moreover, it is important to note the

spike in fear around the period when home isolation

started. On the 7th of March the ﬁrst COVID-19 cases

were found among student population and in the fol-

lowing day 2 universities were closed.

6 CONCLUSIONS AND FUTURE

DIRECTIONS

Social media reﬂects the lives of a population and

their attitudes and as such, social media sentimen-

tal analysis and behavioural tracking has the poten-

tial of becoming an important tool into turning pub-

lic health provision more personalized and efﬁcient.

Figure 2: Number of tweets per day during March. Tweets

written in English in blue and tweets written in Portuguese

in orange. There are two clear spikes in the number of Por-

tuguese tweets in the beginning of the month, with the ﬁrst

COVID-19 suspected cases and a second spike around the

period when home isolation was decreted in Portugal.

Figure 3: Evolution of the basic emotion frequencies in

tweets written in English and Portuguese during March. We

also include the evolution as categories of negative emotion

and positive emotion.

We presented in this paper the evolution of the six ba-

sic emotions over Twitter, during the COVID-19 pan-

demic in Portugal, which can be correlated with the

disease evolution. We found the emotional patterns

of the tweets largely consistent with our expectations,

with more negative tweets posted in the beginning of

the pandemic, up to its peak.

As future work, we are interested in analyzing this

dataset into more detail in order to leverage mental

health and well-being status knowledge of the social

media users. As such, we will explore psycholinguis-

tic features and previously trained models on social

media corpora of mental-health related issues in or-

der to better understand the impact of this pandemic.

ACKNOWLEDGEMENTS

This work was supported by the Integrated Pro-

gramme of SR&TD SOCA (Ref. CENTRO-01-0145-

Bilingual Emotion Analysis on Social Media throughout the COVID19 Pandemic in Portugal

433

FEDER-000010), co-funded by Centro 2020 pro-

gram, Portugal 2020, European Union, through the

European Regional Development Fund and by the

EU/EFPIA Innovative Medicines Initiative 2 Joint

Undertaking under grant agreement No 806968.

REFERENCES

Chen, E., Lerman, K., and Ferrara, E. (2020). Tracking

social media discourse about the covid-19 pandemic:

Development of a public coronavirus twitter data set.

JMIR Public Health and Surveillance, 6(2):e19273.

Conway, M. and O’Connor, D. (2016). Social media, big

data, and mental health: current advances and ethical

implications. Current opinion in psychology, 9:77–82.

Ekman, P. (1992). An argument for basic emotions. Cogni-

tion & emotion, 6(3-4):169–200.

Fast, E., Chen, B., and Bernstein, M. S. (2016). Empath:

Understanding topic signals in large-scale text. In Pro-

ceedings of the 2016 CHI Conference on Human Fac-

tors in Computing Systems, pages 4647–4657. ACM.

Gao, J., Zheng, P., Jia, Y., Chen, H., Mao, Y., Chen, S.,

Wang, Y., Fu, H., and Dai, J. (2020). Mental health

problems and social media exposure during covid-19

outbreak. Plos one, 15(4):e0231924.

Hall, R. C., Hall, R. C., and Chapman, M. J. (2008). The

1995 kikwit ebola outbreak: lessons hospitals and

physicians can apply to future viral epidemics. Gen-

eral hospital psychiatry, 30(5):446–452.

Ho, C. S., Chee, C. Y., and Ho, R. C. (2020). Mental

health strategies to combat the psychological impact

of covid-19 beyond paranoia and panic. Ann Acad

Med Singapore, 49(1):1–3.

Moreira, P. S., Ferreira, S., Couto, B., Machado-Sousa, M.,

Fernandez, M., Raposo-Lima, C., Sousa, N., Pico-

Perez, M., and Morgado, P. (2020). Protective ele-

ments of mental health status during the covid-19 out-

break in the portuguese population. medRxiv.

uller, N. (2014). Infectious diseases and mental

health. Comorbidity of Mental and Physical Disor-

ders, page 99.

Paul, M. J. and Dredze, M. (2017). Social monitoring for

public health. Synthesis Lectures on Information Con-

cepts, Retrieval, and Services, 9(5):1–183.

Wind, T. R., Rijkeboer, M., Andersson, G., and Riper, H.

(2020). The covid-19 pandemic: The ‘black swan’for

mental health care and a turning point for e-health.

Internet Interventions.

Xiang, Y.-T., Yang, Y., Li, W., Zhang, L., Zhang, Q., Che-

ung, T., and Ng, C. H. (2020). Timely mental health

care for the 2019 novel coronavirus outbreak is ur-

gently needed. The Lancet Psychiatry, 7(3):228–229.

Yao, H., Chen, J.-H., and Xu, Y.-F. (2020). Patients with

mental health disorders in the covid-19 epidemic. The

Lancet Psychiatry, 7(4):e21.

HEALTHINF 2021 - 14th International Conference on Health Informatics

434