CoChat
Sign Up
Cited by 375Sentiment Analysis and Opinion MiningCommunication and COVID-19 ImpactMisinformation and Its Impacts

Sentiment Analysis of COVID-19 tweets by Deep Learning Classifiers—A study to show how popularity is affecting accuracy in social media

Open access
Chakraborty, Koyel, Bhatia, Surbhi, Bhattacharyya, Siddhartha, Platos, Jan, Bag, Rajib, Hassanien, Aboul Ella·2020·Applied Soft Computing
DOI 10.1016/j.asoc.2020.106754PMC PMC7521435
Open this paper in CoChat — free account

Abstract

COVID-19 originally known as Corona VIrus Disease of 2019, has been declared as a pandemic by World Health Organization (WHO) on 11th March 2020. Unprecedented pressures have mounted on each country to make compelling requisites for controlling the population by assessing the cases and properly utilizing available resources. The rapid number of exponential cases globally has become the apprehension of panic, fear and anxiety among people. The mental and physical health of the global population is found to be directly proportional to this pandemic disease. The current situation has reported more than twenty four million people being tested positive worldwide as of 27th August, 2020. Therefore, it is the need of the hour to implement different measures to safeguard the countries by demystifying the pertinent facts and information. This paper aims to bring out the fact that tweets containing all handles related to COVID-19 and WHO have been unsuccessful in guiding people around this pandemic outbreak appositely. This study analyzes two types of tweets gathered during the pandemic times. In one case, around twenty three thousand most re-tweeted tweets within the time span from 1st Jan 2019 to 23rd March 2020 have been analyzed and observation says that the maximum number of the tweets portrays neutral or negative sentiments. On the other hand, a dataset containing 226,668 tweets collected within the time span between December 2019 and May 2020 have been analyzed which contrastingly show that there were a maximum number of positive and neutral tweets tweeted by netizens. The research demonstrates that though people have tweeted mostly positive regarding COVID-19, yet netizens were busy engrossed in re-tweeting the negative tweets and that no useful words could be found in WordCloud or computations using word frequency in tweets. The claims have been validated through a proposed model using deep learning classifiers with admissible accuracy up to 81%. Apart from these the authors have proposed the implementation of a Gaussian membership function based fuzzy rule base to correctly identify sentiments from tweets. The accuracy for the said model yields up to a permissible rate of 79%.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper analyzes COVID-19 tweets and argues that highly retweeted/popular tweets skew negative or neutral while unique tweets are more positive or neutral, with word clouds and word frequencies showing little actionable guidance. It reports up to 81% accuracy for a Logistic Regression/Tfidf trigram model on DATA_SET 1 and 79% F1 for a Gaussian fuzzy rule-based sentiment model.

Key findings

  • In DATA_SET 1 (226,673 tweets), TextBlob labeled 98,995 positive, 53,414 negative, and 74,264 neutral tweets; AFINN labeled 66,481 positive, 101,607 negative, and 58,585 neutral tweets. In the most-retweeted DATA_SET 2 (22,985 tweets), TextBlob labeled 6,728 positive, 3,597 negative, and 12,660 neutral tweets; AFINN labeled 4,604 positive, 9,273 negative, and 9,108 neutral tweets.
  • Best classifier accuracy was 81% for DATA_SET 1 using Logistic Regression with trigrams under the Tfidf vectorizer, and 75% for DATA_SET 2 using Logistic Regression with bigrams under the Tfidf vectorizer.
  • The proposed Gaussian membership function fuzzy rule-based system achieved F1-score 79%, precision 82.13%, recall 84.56%, sensitivity 91%, and mean absolute error 0.16, compared with the triangular membership system’s F1-score 78%, precision 80.45%, recall 82.61%, sensitivity 93%, and mean absolute error 0.19.
  • A non-parametric independent t-test on positive and negative sentiment values yielded t-value = 2.578 and p-value = 0.035, which the authors use to support their claim that social media was not useful enough during the COVID-19 outbreak.

Methods

  • The study used two Twitter datasets: DATA_SET 1 with 226,668 unique tweets from December 2019 to May 2020 collected via keywords such as “corona,” “covid,” “sarscov2,” “covid19,” and “coronavirus”; DATA_SET 2 with approximately 23,000 tweets having at least 1,000 retweets from 1 January 2020 to 23 March 2020.
  • Tweets were cleaned, labeled using SentiWordNet, TextBlob, and Afinn, then transformed with Count Vectorizer and Tfidf Vectorizer using unigrams, bigrams, and trigrams; classifiers included MultinomialNB, BernoulliNB, LinearSVC, AdaBoost, RandomForest, and LogisticRegression, with Doc2Vec models DBOW, DMC, and DMM also tested.
  • A Mamdani-style fuzzy rule-based model was proposed using Gaussian membership functions for positive score, negative score, and sentiment output, with seven inference rules and centroid defuzzification; VADER was used to label data for this fuzzy model.

Limitations

  • The paper states that the work “does not have the features to attend multilingual tweets,” which is listed as future work.
  • The paper states that further improvement may be possible by incorporating more membership grades in the fuzzy model.
  • Evident from the text: the analysis is limited to Twitter data over specific pandemic periods, and the datasets have unequal numbers of positive, negative, and neutral tweets, which the authors address by also using a Random Forest classifier to create balance.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1016/j.asoc.2020.106754

Cites

55 of 55 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 7 cited sources checked · 0 supported · 0 flagged

  • Findingnone given

    The implemented model gives an accuracy of above 75% on both datasets used.

    “The implemented model gives an accuracy of above 75% on both the datasets used for this purpose.”
    Quote in the paperNot a substantial claim
  • Findingown data

    The proposed Gaussian membership function based fuzzy rule base model achieved accuracy up to 79%.

    “The accuracy for the said model yields up to a permissible rate of 79%.”
    Quote in the paperSubstantial claim
  • Findingown data

    People have tweeted mostly positive regarding COVID-19.

    “The research demonstrates that though people have tweeted mostly positive regarding COVID-19”
    Quote in the paperSubstantial claim
  • Findingnone given

    Netizens were busy engrossed in re-tweeting negative tweets.

    “yet netizens were busy engrossed in re-tweeting the negative tweets”
    Quote in the paperSubstantial claim
  • Findingown data

    The proposed deep learning classifier model achieved admissible accuracy up to 81%.

    “The claims have been validated through a proposed model using deep learning classifiers with admissible accuracy up to 81%.”
    Quote in the paperSubstantial claim
  • Classified as probably not a substantial claim. 22
  • The quote only partly states this claim. 3
  • May combine two claims. 9

  • Probably not a substantial claim: may not pass S4 (based).

    Findingnone given

    Negative or neutral tweets are being re-shared most.

    “but the tweets those are negative or neutral in nature are being re-shared most.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). The quote only partly states this claim.

    Findingown data

    The unique dataset shows people uniquely are spreading positive vibes pertaining to the pandemic.

    “The unique dataset has hence been analyzed and presented to support the claim that though people uniquely are spreading positive vibes pertaining to the pandemic”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Findingnone given

    People initially shared positive and neutral data, but further users re-tweeted tweets that are negative in nature.

    “As a contribution to the society, this paper has demonstrated with facts that how it is shocking that though people are initially sharing positive and neutral data but further users are re-tweeting those tweets which are negative in nature.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    For DATA_SET 1, there is a maximum number of positive and negative tweets.

    “For DATA_SET 1, there is a maximum number of positive and negative tweets”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Findingnone given

    Analyzing the posts, most social media posts/tweets on COVID-19 have misled people by posting false data and figures.

    “analyzing the posts, it has been seen that most of them have misled people by posting false data and figures.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The Mean Absolute Error value indicates the efficiency of the proposed system, although the Log Loss measure is lower for the Triangular counterpart.

    “In addition, the Mean Absolute Error value also stands as an indicator of the efficiency of the system, although the Log Loss measure is lower for the Triangular counterpart.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Findingnone given

    The most retweeted users in Twitter are @realDonalTrump, @WHO etc.

    “It has been seen that the most retweeted users in Twitter® are @realDonalTrump, @WHO etc.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). The quote only partly states this claim.

    Findingnone given

    Unique tweets result in positive sentiment.

    “While the unique tweets result in positive”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Findingown data

    The WordClouds of the sentiments display words that do not prove any efficiency in representing a viable solution during emergencies.

    “These too display words that do not prove any efficiency in representing a viable solution during emergencies.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The WordCloud visualization adequately represents the content and discourse of the tweets.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingreasoning

    The time taken for the model to train DATA_SET 1 is much more compared to the time taken for DATA_SET 2, due to its larger size.

    “Obviously, the time taken for the model to train DATA_SET 1 is much more compared to the time taken for DATA_SET 2 due to its larger size.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    Tweets from WHO failed to provide precise information that could be retrieved to better deal with the disease.

    “Tweets from WHO were also analyzed but they also failed to provide precise information which could be retrieved to better deal with the disease.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    DATA_SET 1 is ten times the size of DATA_SET 2, and yet the model exhibits more or less the same behavior for both datasets.

    “DATA_SET 1 is ten times the size of DATA_SET 2 and yet the model exhibits more or less the same behavior for both the datasets.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The model's accuracy on the testing data is up to 81% for DATA_SET 1 and 75% for DATA_SET 2.

    “Finally, the aforesaid model is estimated on the testing data and the accuracy yields up to 81% and 75% accuracy for DATA_SET 1 and DATA_SET 2, respectively.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). The quote only partly states this claim.

    Findingown data

    In DATA_SET 1, TEXTBLOB labeled 98,995 tweets positive, 53,414 negative and 74,264 neutral.

    “| No. of positive tweets | 98 995 | 66 481 | 6728 | 4604 |”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The sentiment labeling methods (TEXTBLOB, AFINN, VADER) correctly capture the sentiment of the tweets.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Social media cannot be considered as a trusted source.

    “The work done in the paper also relates to the title as how instead of the magnanimous popularity of social media, it is leading to the dwindling of tweets and for which the same cannot be considered as a trusted source.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationown data

    People showed confused behavior in social media during the pandemic outbreak and portrayed emotions likewise.

    “The figures clearly depict that people showed confused behavior in social media during the pandemic outbreak and portrayed emotions likewise.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    The proposed Gaussian system exhibits less sensitivity towards change of inputs, thereby indicating the stability of the system.

    “Moreover, the proposed system exhibits less sensitivity towards change of inputs, thereby indicating the stability of the system.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

    Premisecitation

    Twitter is one of the most trendy micro blogging sites, which is considered as a crucial depository of sentiment analysis.

    “Twitter® is one of the most trendy micro blogging sites, which is considered as a crucial depository of sentiment analysis [4].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [4] FÅ. Nielsen, A new ANEW: Evaluation of a word list for sentiment analysis in microblogs. arXiv preprint arXiv:1103.2903. 15 Mar.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    TextBlob produces normalized scales in contrast to Afinn.

    “It is generally observed that TextBlob produces normalized scales in contrast to Afinn [38].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [38] E. Kouloumpis, T. Wilson, J. Moore, Twitter sentiment analysis: The good the bad and the omg!, in: Fifth International AAAI conference on weblogs and social media. (2011, July).

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    The remarkably high number of neutral tweets emotes uncertainty over the state of the disease.

    “hence emoting uncertainty over the state of the disease.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    This work does not have the features to attend multilingual tweets.

    “This work does not have the features to attend multilingual tweets, which could be considered as a probable future work in this direction.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductcitation

    K-fold Cross Validation is used as a resampling technique to check the steadiness of the model.

    “Here, K-fold Cross Validation [57] is also used as a resampling technique to check the steadiness of the model.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [57] Mikolov, 2013, Distributed representations of words and phrases and their compositionality, Adv. Neural Inform. Process. Syst.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    In general, LinearSVC classifiers prove efficient for text data classification cases.

    “In general, LinearSVC classifiers prove efficient for text data classification cases.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The Gaussian membership function depends on two parameters: the mean of the data and its standard deviation.

    “Gaussian membership function (MF) [41] depends on two of its parameters, the mean of the data and its standard deviation.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [41] Hameed, 2011, Using Gaussian membership functions for improving the reliability and robustness of students’ evaluation systems, Expert Syst. Appl. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Premisenone given

    For self-created analyser models, there is no perfect classification algorithm that exists.

    “It has been observed that for self-created analyser models, there is no perfect classification algorithm that exists.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Naïve Bayes classifier models are very effective in making predictions for sentiment analysis and are based on Bayes' Theorem.

    “The Naïve Bayes classifier models are very effective in making predictions for sentiment analysis and are based on Bayes’ Theorem [55].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [55] https://en.wikipedia.org/wiki/Naive_Bayes_classifier, Anonymous, retrieved 12th September, 2020.

      Not checked in the free run.

26 of 26

  • The paper relies on this without support.

    Misleading information on social media causes depression, stress, anxiety and directly affects mental health.

    “it is now becoming a source of depression, stress and anxiety because of misleading information posted on social media.”
    “The mental health is directly affected because of the rapid spread of false information on social media.”

  • The paper relies on this without support.

    People's tweets and sentiments reflect the actual mental/emotional state of social media users.

    “We have implemented various state-of-the-art deep learning classifiers to extract the actual sentiment of social media users during this pandemic.”
    “Sentiment analysis is the gathering of people’s views regarding any event happening in real life.”

  • The paper relies on this without support.

    Accurate and not popular messages are being seen, shared or followed at par with the outbreak of Corona Virus in the world.

    “complete unique tweets and the tweets that have at least 1,000 re-tweets based on the logic that accurate and not popular messages are being seen, shared or followed at par with outbreak of Corona Virus in the world, have been considered in this paper.”

  • The paper relies on this without support.

    The mental and physical health of the global population is directly proportional to the pandemic disease.

    “The mental and physical health of the global population is found to be directly proportional to this pandemic disease.”

  • The paper relies on this without support.

    The tweet datasets (DATA_SET 1 and DATA_SET 2) are representative of public sentiment worldwide during the COVID-19 pandemic.

    “One of the major outcomes of this paper is the establishment of the fact from DATA_SET 1 that people worldwide has shown positive sentiments towards the disease.”

  • The paper relies on this without support.

    The sentiment labeling methods (TEXTBLOB, AFINN, VADER) correctly capture the sentiment of the tweets.

    “Step 2: VADER sentiment lexicon is used to label the data into three classes, viz., positive, negative and neutral based on their polarity scores.”

  • The paper relies on this without support.

    The WordCloud visualization adequately represents the content and discourse of the tweets.

    “The majority of the words that have been portrayed in each of the sentiments has been visualized using the WordCloud modules.”

  • The paper relies on this without support.

    The number of retweets reflects the popularity or importance of a tweet.

    “But, as mentioned earlier, the tweets with minimum 1000 retweets were considered for this experiment.”
8 of 8

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Conflicting start dates for the retweeted-tweet dataset

    Minor

    The abstract says the retweeted tweets were collected from 1 January 2019, whereas the methods section says collection began on 1 January 2020.

  • DBOW+DMC logistic regression is not the fastest Table 7 result

    Minor

    The text calls this the best result for execution time, but Table 7 reports 14.12 seconds for it and 0.16 seconds for DBOW+DMC with BernoulliNB.

  • Conflicting sizes for DATA_SET 1

    Minor

    The text gives the first dataset as 226,688 tweets in one passage and 226,668 in another; Table 3 gives a third total of 226,673.

  • The described data split assigns no portion to the test set

    Minor

    The paper says data were divided into training, validation and test sets, but its stated allocation assigns both the 90% portion and the remaining 5% portion to training.

  • Logistic regression does not perform best in all Doc2Vec cases

    Minor

    The claim conflicts with Table 8, which lists LinearSVC or BernoulliNB-not logistic regression-as the reported classifiers for every DATA_SET 1 Doc2Vec model.

5 of 5

Cites

Flagged first, then by citations

Research works printed but not linked: 5

  1. Li, 2020, The impact of COVID-19 epidemic declaration on psychological consequences: a study on active weibo users, Int. J. Environ. Res. Publ. Health, 17, 2032, 10.3390/ijerph17062032 · DOI · Paper page
  2. E. Kouloumpis, T. Wilson, J. Moore, Twitter sentiment analysis: The good the bad and the omg!, in: Fifth International AAAI conference on weblogs and social media. (2011, July).
  3. Q. Le, T. Mikolov, (2014, January). Distributed representations of sentences and documents, in: International Conference on Machine Learning, pp. 1188-1196.
  4. I. Rish, An empirical study of the naive Bayes classifier, in: IJCAI 2001 workshop on empirical methods in artificial intelligence, Vol. 3(22) (2001, August), pp. 41-46.
  5. Olsen, 2004, Cross-sectional study design and data analysis, College Entrance Exam. Board, 26, 2006

Cited by

Sorted by citations

Related

Authors
Chakraborty, K., Bhatia, S., Bhattacharyya, S., Platos, J., Bag, R., & Hassanien, A. E.
Affiliations
  • King Faisal University (SA)
  • Christ University (IN)
  • Cairo University (EG)
Corresponding author
Bhattacharyya, Siddhartha
ORCID
Venue
Applied Soft Computing
Publisher
Elsevier BV
Published
2020-12
Volume
97
Pages
106754
ISSN
1568-4946 (ISSN-L)
DOI
10.1016/j.asoc.2020.106754
PMC
PMC7521435
OpenAlex
W3088352144
Semantic Scholar
13aa8ef82c8a0994da27f3edcf9a0eb90a6fea73
Cited by
375
References
53
Citations by year
2026: 6 · 2025: 24 · 2024: 55 · 2023: 80 · 2022: 115 · 2021: 94 · 2020: 1
FWCI
28.50
Type
journal article
Language
en
Open access
Licence
Locations
  • Applied Soft Computing (publishedVersion)
  • PubMed Central (publishedVersion)
  • PubMed (publishedVersion)
  • DSpace VŠB-TUO (VŠB-TUO) (submittedVersion)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Sentiment Analysis and Opinion Mining
  • Social Sciences › Communication › Communication and COVID-19 Impact
  • Social Sciences › Sociology and Political Science › Misinformation and Its Impacts
Keywords
COVID-19 pandemic, sentiment analysis, Twitter, retweets, tweet popularity, deep learning classifier, Gaussian membership function, fuzzy rule base, public perception of pandemic, World Health Organization

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review