Content uploaded by Sunny Rai
Author content
All content in this area was uploaded by Sunny Rai on May 20, 2022
Content may be subject to copyright.
Identifying Human Needs through Social Media:
A study on Indian cities during COVID-19
Sunny Rai1, Rohan Joseph1, Prakruti Singh Thakur2and Mohammed Abdul Khaliq3
1Mahindra University, Hyderabad, India.
2Arizona State University, Tempe, USA
3University of Stuttgart, Stuttgart, Germany
sunny.rai@mahindrauniversity.edu.in,rohan18545@mechyd.ac.in
psthakur@asu.edu,st181091@stud.uni-stuttgart.de
Abstract
In this paper, we present a minimally-
supervised approach to identify human needs
expressed in tweets. Taking inspiration from
Frustration-Aggression theory, we trained
RoBERTa model to classify tweets expressing
frustration which serves as an indicator of un-
met needs. Although the notion of frustration
is highly subjective and complex, the findings
support the use of pretrained language model
in identifying tweets with unmet needs. Our
study reveals the major causes behind feeling
frustrated during the lockdown and the sec-
ond wave of the COVID-19 pandemic in India.
Our proposed approach can be useful in timely
identification and prioritization of emerging
human needs in the event of a crisis.
Accepted @ SocialNLP,NAACL 2022.
1 Introduction
India reported its first case of COVID-19 from
Kerala in the month of Jan, 2020 (Andrews et al
.
,
2020). Several control measures including restric-
tions on international travel, screening of flight
passengers, and institutional quarantine were un-
dertaken shortly after to combat the transmission.
The Government of India (GoI) imposed a nation-
wide lockdown
1
on Mar 25, 2020 as a preventive
measure to curb the spread of COVID-19. Lock-
down is an emergency protocol that restricts non-
essential movement of people as well as goods.
This lockdown was eventually extended till May
31, 2020, making it one of the longest lockdowns
imposed during the pandemic. This resulted in a
huge gap in demand and supply of goods (Mahajan
and Tomar,2021), increased stress (Rehman et al
.
,
2021) and mass exodus of migrant workers from
cities due to lack of earning opportunities in the
informal economy (Das and Kumar,2020).
1
‘Coronavirus in India: 21-day lockdown begins; key high-
lights of PM Modi’s speech’, Business Today (Mar 25, 2020).
Available at Link
Amidst the growing panic, Twitter emerged as
the go to platform to express one’s feelings and
needs such as travel, food,hospital beds, oxygen,
cremation and funds
2
. An overwhelming number
of tweets seeking support, lack of timely response
and inadequate after-care are a few motivating fac-
tors behind this study. We particularly study the
tweets from metropolitan Indian cities posted dur-
ing the COVID-19 pandemic. The main contribu-
tions are as follows:
•
Using topic modeling and minimal supervi-
sion, we create a dataset of tweets labelled
with needs as described in Maslow’s The-
ory of Motivation (Maslow and Lewis,1987).
This dataset with tagged needs is available for
public research3.
•
Taking inspiration from Frustration-
Aggression theory (Dollard et al
.
,1939), we
finetuned a state of the art neural language
model, RoBERTa (Liu et al
.
,2019) to detect
the unmet needs.
The rest of the paper is organized as follows.
Section 2describes the prior work pertinent to the
research presented here. We present our approach
to gather the needs from Twitter in Section 3. We in-
troduce a RoBERTa based classifier to detect unmet
needs in Section 4. We discuss the social impact of
the proposed work in Section 5and list down the
limitations in Section 6. We conclude our work in
Section 7.
2 Background
Understanding human needs is a widely researched
domain by state agencies as well as commercial
2
Reuters:
https://graphics.reuters.
com/HEALTH-CORONAVIRUS/INDIA- TWITTER/
oakpekqlrpr/
3https://github.com/AxleBlaze3/Covid_
19_Tweets_with_Tagged_Needs
Figure 1: Block diagram for the proposed approach
organizations (Costanza et al
.
,2007). Prior re-
search has shown that fulfilled needs have a pos-
itive impact on a person’s feelings of well-being
(Ryff and Keyes,1995). From stockpiling basic
household items during the initial phase of the pan-
demic to embracing digital technologies such as
zoom, the market witnessed quite a shift in con-
sumer needs since the outbreak of COVID-19 (Bec-
dach et al
.
,2020;Mehta et al
.
,2020). Identifying
one’s true needs, however, is a challenging task.
Yang and Li (2013) took inspiration from Maslow’s
theory of motivation to predict consumer’s needs
and purchasing behavior using social media. Ko
et al
.
(2020) used Korean twitter and blogs to dis-
cover customer’s unmet needs through Hierarchical
Concept Search Space algorithm. Their approach
aimed to facilitate idea generation for home appli-
ances.
More recently, Yang et al. (2021) advocated the
use of Weibo
4
to identify unmet non-COVID-19
healthcare needs. Suh et al
.
(2021) studied the
transition in needs during COVID-19 through the
search queries on Bing. The product type in search
queries were manually marked with the needs as
described in Maslow’s theory of motivation to au-
tomate the task of need identification. Their results
affirmed a human tendency to first satisfy basic
needs such as food and shelter before exploring
advanced needs such as creativity and love.Jolly
et al
.
(2020) performed a psychometric analysis of
tweets posted in response to official bulletins on
COVID-19 by state agencies, revealing the causal-
ity between bulletins and the feeling of medical
emergency on Twitter.
Prior studies (Saha et al
.
,2020;Guntuku et al
.
,
2020;Mendoza et al
.
,2010) have consistently
demonstrated the efficacy of social media platforms
4Weibo.com
such as Twitter in capturing the feelings of society
at scale. However, this unfurls the challenge of
annotating the posts with their expressed needs. In
this paper, we propose an approach to automate
labeling of tweets with their expressed needs. We
also build a model to detect unmet needs from
tweets. To the best of our knowledge, this is the
first study on the needs expressed through tweets
from Indian cities during the pandemic.
3 Identifying Human needs from Tweets
We illustrate the different components of the pro-
posed model in Fig. 1. The block diagram repre-
sents the steps of the proposed approach that are,
(a) extracting key topics from Twitter discourse, (b)
manual mapping of the topics to the expressed hu-
man needs, (c) mapping tweets to needs assigned
to their dominant topics and (d) detection of unmet
needs from tweets. The components in green rep-
resent the use case of detecting unmet needs and
categorization when given a live stream of tweets.
3.1 Tweets Collection
The GoI officially declared the nationwide lock-
down on Mar 25, 2020. The second wave of
COVID-19 peaked in the mid of May, 2021. Tak-
ing into account the baseline considered by Suh
et al
.
(2021) and the number of COVID-19 cases
5
in India, we set the duration of study from Dec 1,
2019 to Jun 30, 2021, comprising a total of nine-
teen months. The first three months that is from
Dec 1, 2019 to Feb 28, 2020 is the baseline period
that serves as an indicator of pre-COVID-19 needs
pattern. We here assume that a tweet does not need
to be marked with hashtags related to COVID-19
to have a need affected or emerged due to ongoing
5
WHO, Coronavirus disease 2019 (COVID-19) Situation
Report – 39. Feb 28, 2020. LINK
Figure 2: Areas covered during Tweet Collection
COVID-19 pandemic.
Using snscrape
6
, we extract Indian tweets posted
between Dec, 2019 and June, 2021. We set the
parameter “geocode” of the form [latitude, longi-
tude, radius] to the (latitude, longitude) of cities
namely Nagpur, Bangalore, Jaipur, Kolkata and
Patna with the radii as 500km, 400km, 350km,
50km and 100km respectively in an attempt to en-
compass representative metropolitan cities situated
in different parts of India. The covered region is
depicted in Fig. 2.
As a pre-processing step, we removed duplicate
and non-English tweets. We also filtered out tweets
having less than twelve words. The word limit
threshold was decided empirically after analysing
the content of tweets. These tweets were either
related to marketing/greetings such as good morn-
ing,happy birthday or comments on the original
tweets without substantial semantic content of its
own. We have a total of
1.4M
unique tweets for
further study.
3.2 Mapping Tweets to Human Needs
In this research, we study only expressed needs in
tweets. Bradshaw (1972) defined expressed need
as “the felt need turned into action”. To identify
the different types of expressed need, we take inspi-
ration from Maslow’s Hierarchy of Needs (MHoN)
(Maslow and Lewis,1987) which categorizes the
human needs into five distinct levels namely physio-
logical[
L1
], Safety [
L2
], Love and Belonging [
L3
],
Love and Belonging [
L3
], Esteem [
L4
] and Self-
6https://pypi.org/project/snscrape/
Actualization [
L5
]. Physiological and safety needs
are considered basic needs that need to be satis-
fied first before one begins to explore the advanced
needs related to esteem and self-actualization.
3.2.1 Topic Extraction
Manual annotation of over
1.4M
tweets with their
expressed need is time consuming as well as pro-
hibitively expensive. We therefore employ topic
modeling
7
(Blei et al
.
,2003) to identify the major
topics of discourse for monthwise set of tweets
which we then manually label to a level as de-
scribed in Maslow’s Theory of Motivation.The
number of topic words is set to
20
and the rest of
the parameters were set to default values. The num-
ber of topics is decided empirically after analysing
the coherence score and execution time for a ran-
domly picked sample of three months. We set the
#topics
to
30
after analysing different number of
topics,
#topics ={10,15,20, ...45,50}
. We thus
obtain a set
T
having
570
topics (
30
topics
∗19
months).
3.2.2 Manual Labeling of Topics
We asked a team of three human annotators to
map the extracted topics
t∈ T
to the levels
{L1, L2, L3, L4, L5} ∈
MHoN. Each annotator is
an undergraduate student, aged 19-21 years and
highly proficient in the English language. Two
were male and one was female. Given a topic
t
and few tweets elaborating its context of usage, the
task is to map the topic
t
to either
Li∈
MHoN
or as ‘unclear’. Annotators were encouraged to
choose unclear if they find a topic ambiguous. A
topic is assigned a need level from MHoN only if
all annotators choose the same level.
Out of
570
topics,
59
topics were assigned to
L1
,
150
topics mapped to
L2
,
84
topics to
L3
,
66
topics to
L4
and
95
topics to the level
L5
. Rest
were unclear.
The key categories emerged after mapping top-
ics with needs are provided in Table 1. The physi-
ological need majorly comprised food staples and
beverages,hygiene concerns,mobility. Clearly, the
meaning of safety has evolved and included topics
such as housing,infection,unemployment,domes-
tic violence,market and financial liabilities. Rela-
tionships and concern for loved one’s are discussed
under love and belonging.Esteem covers online
7
Gensim LDA:
https://radimrehurek.com/
gensim_3.8.3/models/wrappers/ldamallet.
html
Table 1: Mapped topics and Need Level in Maslow’s Theory of Motivation
Need #Tweets #Topics Key Topics
Physiological 165114 59 staples
such as food, beverages, apparel, household products,
Hygiene
such as toilet paper,
basic daily services
such as grocery delivery, milk,
bread,rest,medicine,transport
Safety 367774 150 Housing
such as rental/mortgages, evictions,
COVID-19 Safety
such as
masks, quarantine or sanitizers,
Domestic violence
, justice,
Financial
liabilities
such as tax, loans, or bankruptcy, stock market, business
Job
posts/application & unemployment
Love & Belong-
ing
192843 84
Expression of or resources for
mental health
or
emotional issues
such
as anxiety, depression, loneliness, isolation, suicide, nervousness, re-
jection, fear or sadness;
Social media
, Search for
relationships
with
significant others, dating, issues such as divorce or breakup
Esteem 151873 66 Education
or learning materials, University/Schools;
Online class-
room
learning, Examinations; Educational degrees or programs; Knowl-
edge/Skill, zoom meetings, Ideologies/religions
Self-
Actualization
258287 95 Recreational task
s such as self-care, home decor, music etc., parent-
ing, wedding,
Talent/Skill acquisition
,
Life goals
, Charity/Donation,
volunteering
,
Entertainment
such as Netflix, Prime, TV shows, movie,
sports (IPL), News
learning,ideologies, postponed examinations, lack
of internet and smart devices. Self-actualization
comprises recreational tasks such as DIY, sports,
entertainment and skill acquisition.
3.2.3 Mapping Tweets to MHoN
At this step, we have a list of topics mapped with
the need levels as described in MHoN. We also
have the probability distribution of topics for each
tweet. The dominant topic of a tweet is the topic
with the highest probability. A tweet
t
is thus
marked to a need level
Li∈
MHoN on the ba-
sis of the mapping assigned to its dominant topic.
We illustrate this process in Fig. 1where the topic
distribution along with mapped topics are used to
infer the expressed need in tweets. If the tweet has
multiple topics with same probability, we only as-
sign a need level if all dominant topics are marked
to the same need level else it is marked as unclear.
3.3 Analysis
After excluding unclear tweets, we have over
1.1M
tweets marked as expressing a need. Be-
low are few examples
8
that were mapped to their
relevant level and those that were unclear:
“Nobody staying at hotels, So why not convert
them into covid centers ” –Physiological
“When I see some people attacking doctors, i
get scared about the corona situation" –Safety
“Not everyone can work from home. Feeling
kinda unsafe or its just fear of getting sucked up
8
Tweets are rephrased to protect user’s privacy however,
the message remains the same.
Figure 3: Volume of tweets expressing needs from Dec,
2019 to Jun, 2021.
in situation and putting life of my family friends in
danger." –Love and Belonging
“As a teacher I thank pm for cancelling the Std.
12th board examination." –Esteem
“xxx movie are a big hit, an average human
would have to watch his movies multiple times to
understand." –Self-Actualization
“Met her while traveling she was selling fruits
adding on to the income of her parents, with her
impressive salesman skills ." –Unclear
It may be noted that we have not considered
tweets on political topics such as Citizen Amend-
ment Act (Wikipedia contributors,2021b), Na-
tional Register of Citizens (Wikipedia contribu-
tors,2021c) and Farm laws (Wikipedia contrib-
utors,2021a) which were part of public discourse
during the time of study.
We illustrate the time-wise distribution of tweets
tagged with different need levels in Fig. 3. We
have considered two weeks moving average to nul-
lify noisy fluctuations in the data. For the base-
line, we consider the tweets posted in the first
twelve weeks that is, between week-48’2019 to
week-7’2020, to understand the pre-COVID pat-
tern of needs. Lockdown phase is the period be-
tween week-13’2020 to week-23’2020. The first
wave ranges from week-31’2020 to week-41’2021.
The second wave started from week-14’2021 and
ended in week-23’2021. The phases namely base-
line, lockdown,first wave and second wave are
annotated with boxes in Fig. 3. The first peak is
placed in the lockdown period and the second peak
occurred during the second wave of the pandemic.
The volume of needs were slightly higher than pre-
COVID levels during the first wave. There is also
a huge surge in self-actualization and safety needs
starting week-24’2021.
Indian Twitter users voiced the safety need most
often followed by physiological need during the
lockdown. Both needs peak at the same time. A
total of
45%
more tweets expressing basic needs
were posted during lockdown compared to the sec-
ond wave of the pandemic. Over twice the num-
ber of physiological tweets was expressed during
the lockdown when compared to the second wave.
The relatively advanced needs namely love and
Belonging and esteem display a delay during the
lockdown and peak almost 3-4 weeks after the ba-
sic needs. Soon after the lockdown was lifted, the
needs started to return to pre-COVID pattern of
needs.
During the second phase of the pandemic, safety
turned out to be the foremost concern and phys-
iological needs peaked only after a delay of two
weeks. There is no clear precedence for physiolog-
ical needs over advanced needs during the second
wave. Moreover, love and belonging needs stayed
at pre-COVID levels during the second wave unlike
lockdown phase where concern for loved ones was
expressed in large volumes.
Safety has indeed emerged as the dominant con-
cern in the both phases of the pandemic. Lockdown
was a special scenario where essential commodi-
ties were in shortage due to lack of production as
well as black marketing
9
. It is thus not conclusive
from our data if physiological needs always take
precedence over the advanced needs in the event of
a crisis in today’s world.
The most advanced need, self-actualization
surged and ebbed through out the months of our
9
The Hindu “Coronavirus lockdown: Invoke Essential
Commodities Act to curb black marketing, Home Secretary
tells States" (Apr 8, 2020). Available at Link
study without any clear correlation with the dif-
ferent phases of pandemic. The huge surge in
self-actualization and safety needs starting week-
24’2021 is due to large volume of tweets discussing
Indian Premier League 2021 (Wikipedia contribu-
tors,2022) and mass gatherings.
4 Detecting unmet Needs
Unmet needs are widely characterized by frustra-
tion (Dollard et al
.
,1939;Killgore et al
.
,2021).
Through Frustration-Aggression theory, Dollard
et al
.
(1939) defined frustration as an impediment
or blockage in achieving one’s needs or goals. An
impediment to a goal is considered frustration if
and only if the person is actively striving to reach
this goal. We thus hypothesize that an unmet need
can be detected by identifying whether a tweet with
expressed need has frustration or not.
4.1 Approach
Our task is to classify whether a given tweet tagged
with need is expressing frustration or not. We fine
tuned the RoBERTa pretrained model (Wolf et al
.
,
2020) with a learning rate of
2e−5
and dropout of
0.3
for this classification task. For training, we
collected tweets containing the hashtag #frustrated.
For negative class that is, Not frustrated, we ex-
tracted tweets with hashtags that symbolise sat-
isfaction (ex: #satisfied,#FeelingContent). This
dataset has a total of
13970
tweets with equal num-
ber of instances for positive and negative class. We
provide a representative tweet from each class be-
low:
“HOW fast does one have to be to book a slot
on COWIN? I saw slots available at a hospital; I
selected the time slot; entered the CAPTCHA in
not more than 15 seconds... and still it didn’t book
the slot. And then when I refreshed, all the slots
were gone” - Frustrated
“I did it! ... I officially completed my
undergraduate program and received my bache-
lors degree. may the glory be to God for blessing
me with the gifts to achieve this great milestone” -
Not Frustrated
As a preprocessing step, we remove hashtags and
mentions from the tweet text. We consider
80%
of tweets for training and the rest
20%
is equally
divided for validation and test set. We achieved
an accuracy of
93.4%
on the validation set. We
obtained an accuracy of
92.2%
with a precision of
91% and recall of 93% on the test set.
Figure 4: Tweets predicted as Frustrated
4.2 Performance Evaluation
Out of
1.1M
tweets, our model predicted a total of
792533
tweets as frustrated.
77.36%
of physiologi-
cal needs and
77.5%
of safety needs expressed frus-
tration. Under advanced needs,
54.13%
of love and
belonging,
70.43%
of esteem needs and
62.91%
of
self-actualization needs were marked frustrated.
To evaluate the quality of predictions, we ran-
domly sampled
100
tweets which were annotated
as frustrated or not frustrated by three undergrad
students proficient in English. The majority vote
was considered as the final label. A total of
45
samples were labelled as frustrated out of
100
. The
inter-annotator agreement (fleiss kappa) obtained
for this task was
0.638
indicating its subjective
nature. The trained model achieved an accuracy
of
76%
with a weighted precision of
78%
and
weighted recall of
76%
on this set of annotated
tweets. Below are two example tweets which were
classified as frustrated:
“Please complete the pending projects in Telan-
gana State. Sir please do the needful. There is no
direct train from Karimnagar to Hyderabad”
“Need of hour Free Education Free/Affordable
health care No freebies , let people work".
We observe that the above tweets clearly express
frustration as described in Dollard et al
.
(1939). An-
other point worth noting is the subjectivity when
labelling frustration. Consider the below tweets
predicted as frustrated but annotated as not frus-
trated by human annotators.
“She lost her life in line of duty. She had
been performing her duty in adverse circumstances
amid lockdown.She should be declared "Corona
Warrior"and all benefits and compensation should
be given to her family by the govt.”
Figure 5: Percentage of Frustrated Tweets
“Finally, I am buying an Iphone , twelfth edi-
tion but next year. As i also thought about Iphone
last year.”
Whether the above tweets express frustration or
not, is quite debatable. Therefore, the performance
metrics need to be interpreted accordingly.
4.2.1 Decoding frustration through RoBERTa
On a random sample of
30
tweets predicted as
frustrated, we used integrated gradients method
(Sundararajan et al
.
,2017) to identify the type of
input features that attribute to the prediction to the
class frustrated.
We provide few example tweets from this set in
Fig. 4. Here, the shade of red signifies the impor-
tance of input features in prediction. The greater
the significance, the deeper the hue of red. For in-
stance, the words highlighted with deeper red such
as shortage,oxygen,where,loose,ridicule, and all
led to the classification into frustrated class for the
first tweet in Fig. 4. Likewise for other tweets, the
words namely have,transport,electricity,delay,
infected,ventilators,expensive,treatment are in-
put features that derived the prediction to the class
frustrated. Since these terms intrinsically reflect
constraints or impediments in leading a purpose-
ful life, we may conclude that the model correctly
learned to detect tweets expressing frustration.
4.3 Discussion
We illustrate the week wise percentage of tweets
predicted as frustrated in Fig. 5. At first glance,
Twitter appears to be a land of frustration with
dissatisfaction rate of around
62%
even before
COVID-19. The jump in frustration rate in the
fourth week of December’19 is due to the mulling
over the passing year and eventually settled down
in the next two months of Jan, 2020 and Feb, 2020.
The percentage of frustrated tweets hovered be-
tween
71−74%
during the lockdown, the first wave
and the second wave of COVID-19. Clearly, there
Figure 6: #Frustrated Tweets for Basic &Advanced
Needs
Figure 7: Proportion of Basic and Advanced Needs
marked Frustrated
is an increment of over
4%
in frustration rate when
compared to non-stressful phases of the pandemic.
We illustrate the week wise transition for the
volume of frustrated tweets expressing basic and
advanced needs in Fig. 6. There is a huge jump
in the volume of both categories of tweets. More
tweets expressing frustration due to basic needs
were posted during the lockdown in comparison
to the second wave. The volume of basic tweets
during the first wave remained slightly above the
pre-COVID level.
The proportion of frustrated tweets across basic
and advanced level of needs is illustrated in Fig.
7. We observe that almost
80%
of tweets express-
ing basic needs are unmet irrespective of the time
of the year. Despite the fact that a large number
of basic needs were posted throughout the lock-
down, the dissatisfaction rate remained constant.
It is thus safe to assume that users discuss basic
needs only when these needs are unfulfilled. The
general rate of frustration for advanced needs is
60%
. We also note that as soon as the frustration
due to basic needs reduces, the frustration due to
advanced needs increased by over
10%
. There are
three such peaks in Fig. 7. This does support the
belief that once the basic needs are secured, one
quickly moves to the advanced needs. On analysis,
education with key terms such as board exams, na-
Figure 8: Themes for frustrated tweets: Lockdown
tional level entrance exams, graduation degree was
found to be the dominant concern across each peak.
Another common concern was consumer-centric
services with worries revolving around delayed re-
funds, cancelled travel plans, delayed delivery of
online orders etc.
Moreover, time specific events such as call
against products made in China, Bollywood scan-
dals, football, entertainment were found during
the second peak (week-36’2020 - week-44’2020).
In contrast, the third peak (week-18’2021 - week-
22’2021 discussed the lack of availability of vac-
cines and further called for inclusivity and trans-
parency in distribution.
4.3.1 Key Themes behind frustration
To discover the themes behind the increased vol-
ume of frustrated tweets during lockdown and the
second wave of COVID-19, we used a computer
program called VOSviewer (van Eck and Waltman,
2011) to create a term co-occurrence map for the
tweets labelled as frustrated. Fig. 8and Fig. 9il-
lustrate the oft-discussed terms in frustrated tweets
posted during the lockdown and the second wave
respectively.
Lockdown: Travel concerns due to the imposed
nationwide lockdown are evident from the terms in
cluster
blue
in Fig. 8. Major Indian cities namely
bengaluru, bihar, pune coupled with transportation
choices such as bus, train, vehicle can be seen. We
also note terms such as quarantine, doctor, patient,
office in the same cluster indicating the traveling
problems faced during daily life activities. The
nodes in
green
reveal the challenges faced by logis-
tics and travel industry. Terms such as refund, ticket,
airline, flight, credit reflect the chief complaints by
customers along with bill and other payments.
The nodes in cluster
red
highlight the discus-
sion on digital media and news channels. Growing
Figure 9: Themes for frustrated tweets: Second wave
concern due to increasing toll of infections in the
USA and a sense of anger towards China were ex-
pressed through tweets. Fake news,channel,minor-
ity and economy were also a few topics of online
discussion. The nodes in cluster
yellow
depict the
concerns revolving around closed educational insti-
tutions, payment of fees,online classes and exams.
Second Wave: The usual customer care com-
plaints are depicted in cluster
green
. The nodes
in cluster
red
particularly reveal the frustration
against political parties and elections. There are
also terms such as player, season, ball, game due to
upcoming IPL cricket matches. The anxiety due to
shortage of ventilator, patient, hospital,icu bed and
oxygen cylinder is captured through nodes in clus-
ter
blue
. Words such as refer, friend, help reveal
anxious attempts to locate healthcare through con-
tacts on Twitter. Availability of vaccine and book-
ing of slots were also a cause of frustration amongst
Indians. Education remained a concern during the
second wave as evident from nodes marked in
pur-
ple.
5 Social Impact
Tsao et al
.
(2021) highlighted the paucity of action
driven research on the COVID-19 data. Early de-
tection of human needs will enable public agencies
and independent organizations to provide prompt
support including food supplies,medical care,
transport and timely awareness about the crisis
amongst masses. Our approach can facilitate timely
identification and prioritization of emerging human
needs in the event of a crisis. When coupled with
geo-location tag, the proposed approach can be
customized to retrieve closest support available.
Unmet needs scoping can help in designing public
policies to cater to emerging needs of a society.
During the COVID-19 pandemic in India, people
expressed distinct needs at different stages of each
wave. Public needs on social media can thus serve
as an immediate feedback mechanism for public
agencies to improvise their relief efforts and poli-
cies. Our model to detect unmet needs leverages a
pre-trained neural language model that generalises
well and is capable of transfer learning from previ-
ously labelled data at the start of a crisis. It is thus
easy to extend our approach for other languages
using publicly available pre-trained multilingual
language models.
6 Limitations
Due to our focus on understanding the pattern of
needs emerged in India during the COVID-19 pan-
demic, we performed rigorous filtering to retain
only those tweets geo-tagged with locations within
India. This significantly reduced the quantity of
tweets gathered for our study. Human needs are in-
nately complex and ever evolving concept. As we
transition from basic to advanced needs, the needs
become more obscure and implicit. To optimize the
time and effort for human annotation, we assumed
that the dominant topic of a tweet would reflect its
need type as discussed in Section 3. This had an
impact on the quality of tweet-need mapping and
resulted in incorrect labeling in some cases.
7 Conclusion
In this paper, we examined the human needs ex-
pressed in Indian cities during the COVID-19 pan-
demic. We described a minimally supervised ap-
proach to annotate tweets with their need level as
in Maslow’s Hierarchy of Needs. This greatly re-
duced the time and human effort without much
impact on the quality of annotation. We observed
a recurring pattern in the needs, indicating pre-
dictability in the emerging needs in the event of a
crisis. The results support the use of pretrained lan-
guage model for the task of unmet needs detection.
In future, we will extend the proposed model to
detect needs in regional languages. We will further
work upon incorporating theories better suited to
capture advanced psychological needs.
References
MA Andrews, Binu Areekal, KR Rajesh, Jijith Krish-
nan, R Suryakala, Biju Krishnan, CP Muraly, and
PV Santhosh. 2020. First confirmed case of COVID-
19 infection in India: A case report. The Indian Jour-
nal of Medical Research 151, 5 (2020), 490.
Camilo Becdach, Brandon Brown, Ford Halbardier,
Brian Henstorf, and Ryan Murphy. 2020. Rapidly
forecasting demand and adapting commercial
plans in a pandemic. URL https://www. mckinsey.
com/industries/consumer-packaged-goods/our-
insights/rapidly-forecasting-demand-and-adapting-
commercial-plans-in-a-pandemic (2020).
David M Blei, Andrew Y Ng, and Michael I Jordan.
2003. Latent dirichlet allocation. the Journal of ma-
chine Learning research 3 (2003), 993–1022.
Jonathan Bradshaw. 1972. The concept of social need.
New society 30 (1972).
Robert Costanza, Brendan Fisher, Saleem Ali, Caro-
line Beer, Lynne Bond, Roelof Boumans, Nicholas L
Danigelis, Jennifer Dickinson, Carolyn Elliott,
Joshua Farley, et al. 2007. Quality of life: An ap-
proach integrating opportunities, human needs, and
subjective well-being. Ecological economics 61, 2-
3 (2007), 267–276.
Ritanjan Das and Nilotpal Kumar. 2020. Chronic crisis:
migrant workers and India’s COVID-19 lockdown.
South Asia@ LSE (2020).
John Dollard, Neal E Miller, Leonard W Doob, Or-
val Hobart Mowrer, and Robert R Sears. 1939. Frus-
tration and aggression. (1939).
Sharath Chandra Guntuku, Garrick Sherman, Daniel C
Stokes, Anish K Agarwal, Emily Seltzer, Raina M
Merchant, and Lyle H Ungar. 2020. Tracking men-
tal health and symptom mentions on Twitter during
COVID-19. Journal of general internal medicine
35, 9 (2020), 2798–2800.
Baani Leen Kaur Jolly, Palash Aggrawal, Amogh
Gulati, Amarjit Singh Sethi, Ponnurangam Ku-
maraguru, and Tavpritesh Sethi. 2020. Psychome-
tric Analysis and Coupling of Emotions Between
State Bulletins and Twitter in India during COVID-
19 Infodemic. arXiv preprint arXiv:2005.05513
(2020).
William DS Killgore, Sara A Cloonan, Emily C Taylor,
Ian Anlap, and Natalie S Dailey. 2021. Increasing
aggression during the COVID-19 lockdowns. Jour-
nal of affective disorders reports 5 (2021), 100163.
Taehoon Ko, Ilsun Rhiu, Myung Hwan Yun, and Sung-
zoon Cho. 2020. A Novel Framework for Identify-
ing Customers’ Unmet Needs on Online Social Me-
dia Using Context Tree. Applied Sciences 10, 23
(2020), 8473.
Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Man-
dar Joshi, Danqi Chen, Omer Levy, Mike Lewis,
Luke Zettlemoyer, and Veselin Stoyanov. 2019.
Roberta: A robustly optimized bert pretraining ap-
proach. arXiv preprint arXiv:1907.11692 (2019).
Kanika Mahajan and Shekhar Tomar. 2021. COVID-19
and Supply Chain Disruption: Evidence from Food
Markets in India. American journal of agricultural
economics 103, 1 (2021), 35–52.
Abraham Maslow and KJ Lewis. 1987. Maslow’s hi-
erarchy of needs. Salenger Incorporated 14 (1987),
987.
Seema Mehta, Tanjul Saxena, and Neetu Purohit.
2020. The New Consumer Behaviour Paradigm
amid COVID-19: Permanent or Transient? Journal
of Health Management 22, 2 (2020), 291–301.
Marcelo Mendoza, Barbara Poblete, and Carlos
Castillo. 2010. Twitter under crisis: Can we trust
what we RT?. In Proceedings of the first workshop
on social media analytics. 71–79.
Usama Rehman, Mohammad G Shahnawaz, Neda H
Khan, Korsi D Kharshiing, Masrat Khursheed,
Kaveri Gupta, Drishti Kashyap, and Ritika Uniyal.
2021. Depression, anxiety and stress among Indians
in times of Covid-19 lockdown. Community mental
health journal 57, 1 (2021), 42–48.
Carol D Ryff and Corey Lee M Keyes. 1995. The struc-
ture of psychological well-being revisited. Journal
of personality and social psychology 69, 4 (1995),
719.
Koustuv Saha, John Torous, Eric D Caine, Mun-
mun De Choudhury, et al. 2020. Psychosocial ef-
fects of the COVID-19 pandemic: large-scale quasi-
experimental study on social media. Journal of med-
ical internet research 22, 11 (2020), e22600.
Jina Suh, Eric Horvitz, Ryen W White, and Tim Al-
thoff. 2021. Population-scale study of human needs
during the covid-19 pandemic: Analysis and impli-
cations. In Proceedings of the 14th ACM Interna-
tional Conference on Web Search and Data Mining.
4–12.
Mukund Sundararajan, Ankur Taly, and Qiqi Yan. 2017.
Axiomatic attribution for deep networks. In Inter-
national Conference on Machine Learning. PMLR,
3319–3328.
Shu-Feng Tsao, Helen Chen, Therese Tisseverasinghe,
Yang Yang, Lianghua Li, and Zahid A Butt. 2021.
What social media told us in the time of COVID-19:
a scoping review. The Lancet Digital Health 3, 3
(2021), e175–e194.
Nees Jan van Eck and Ludo Waltman. 2011.
Text mining and visualization using VOSviewer.
arXiv:1109.2058 [cs.DL]
Wikipedia contributors. 2021a. 2020 Indian agricul-
ture acts — Wikipedia, The Free Encyclopedia.
https://en.wikipedia.org/w/index.php?
title=2020_Indian_agriculture_acts&
oldid=1025817251. [Online; accessed 9-July-
2021].
Wikipedia contributors. 2021b. Citizenship (Amend-
ment) Act, 2019 — Wikipedia, The Free Encyclo-
pedia. https://en.wikipedia.org/w/index.
php?title=Citizenship_(Amendment)_Act,
_2019&oldid=1032618250 [Online; accessed
9-July-2021].
Wikipedia contributors. 2021c. National Register of
Citizens — Wikipedia, The Free Encyclopedia.
https://en.wikipedia.org/w/index.php?
title=National_Register_of_Citizens&
oldid=1028089793. [Online; accessed 9-July-
2021].
Wikipedia contributors. 2022. 2021 Indian Premier
League — Wikipedia, The Free Encyclope-
dia. https://en.wikipedia.org/w/index.
php?title=2021_Indian_Premier_League&
oldid=1079776521. [Online; accessed 1-April-
2022].
Thomas Wolf, Lysandre Debut, Victor Sanh, Julien
Chaumond, Clement Delangue, Anthony Moi, Pier-
ric Cistac, Tim Rault, Rémi Louf, Morgan Fun-
towicz, Joe Davison, Sam Shleifer, Patrick von
Platen, Clara Ma, Yacine Jernite, Julien Plu,
Canwen Xu, Teven Le Scao, Sylvain Gugger,
Mariama Drame, Quentin Lhoest, and Alexan-
der M. Rush. 2020. HuggingFace’s Transform-
ers: State-of-the-art Natural Language Processing.
arXiv:1910.03771 [cs.CL]
Huahai Yang and Yunyao Li. 2013. Identifying user
needs from social media. IBM Research Division,
San Jose (2013), 11.
Wei-Fa Yang, Danping Zheng, Reynold CK Cheng,
Jingya Jane Pu, and Yu-Xiong Su. 2021. Identify-
ing unmet non-COVID-19 health needs during the
COVID-19 outbreak based on social media data: a
proof-of-concept study in Wuhan city. Annals of
Translational Medicine 9, 18 (2021).
