CC BY
29
The Anxieties Detection from Tweets about Distance Learning with Negative Sentiments
Yulia Yu. Dyulicheva
Abstract— The study of the mental health of society and the automatic detection of various types of disorders based on text analysis is one of the priority areas of e-health. The empirical investigation of the various anxieties and their causes in the pandemic period connecting with distance learning introduction based on an analysis of 71475 tweets is performed in the paper. The analysis of hashtags of tweets with negative sentiment and proposed patterns based on part of speech recognition allows us to extract regularities that describe sensations and causes of emotions associated with the appearance of the affective state. We created a dictionary of terms describing affective states based on the most frequent words with negative sentiment. Particular attention was paid to feature engineering using positive words from positive tweets and negative words from negative tweets, taking into account parts of speech. The proposed approach to feature engineering made it possible to reduce the dimension of the feature space while maintaining the quality of the classification (84%).
Keywords—mental health, POS-tagging, random forest, sentiment analysis, TF-IDF
I. Introduction
The forced transition to distance learning during the COVID-19 pandemic gave rise to new challenges for studying the media data, in particular, students' feedback on the quality of teaching, used distance technologies, and studying the impact of students' feedback on the reputation formation of the teachers and Universities [1]. There are many investigations that are aimed at the extraction of the social average opinion and moods from comments on social media platforms. The most studied social media platform is Twitter where people in short messages called tweets can express their feelings, emotions, and relationship to events in their own lives and society. The limited number of characters in messages gave birth to a special manner of communication expressed in emoticons, abbreviations, scraps of words, words with grammatical errors, slang, etc. [2]. These peculiar properties of tweets require special methods of text preprocessing, vectorization, and recognition. Natural Language Processing (NLP) methods together with machine learning opened new perspectives for social media data mining and some success was achieved. For example, opinion mining of political leaders based on sentiment analysis with extraction of the most influenced words on sentiments and entities [3]; the learning of social moods and dynamics of opinions changing about vaccination and COVID-19 vaccines with the extraction of frequency words, n-grams [4] and topics detection with extraction such interesting regularities from
tweets as the trust prevailed over negative emotions associated with fear, anticipation, etc [5]; the investigation of dependencies between tweets (quantity, polarity) and prices of cryptocurrencies [6]; the detection and understanding of toxic behavior of users through tweets analysis [7]; considering the tweets as data for the development of recommendation systems with personalized propositions for users [8]. So, tweets are a powerful resource for understanding society's relationship to different social events and a platform for tools development that are aimed at managing people's opinions, for example, through liars and fakes. The scholars used extraction of parts of speech, frequencies of words, most influenced opinion words, and the different levels of sentiment analysis detailed sentiments by words, sentences, documents, and features for tweets investigation [9].
Analyzing the students' feedback based on tweets and sentiment analysis techniques using the lexicon-based approaches is the most investigated direction in educational data mining. The mining of tweets allows for characterizing the quality of education. Oyekunle and Abdulkareem revealed the students' dissatisfaction with CDT exams and the causes of their negative emotions such as high-level stress, feeling of unwellness, waiting time for results, presence of queues, and people who made it difficult to concentrate [10]. Such regularities extracted from tweets can be used for exams organization improvement next time and developing strategies to reduce stressing states of students. Clustering of students' opinions and their causes can be aimed at detecting target groups with personalized methods of psychological assistance and training. The grouping of students' tweets based on vectorization (n-grams, BOW) and clustering methods (k-Means, Brown clustering algorithm) are considered in [11], [12].
Special interests of social media mining are the understanding of students' problems during learning in universities or colleges such as adaptation problems, the causes of stress, and anxieties. Chen et al. proposed the methodology for mining tweets for students problems detection such as learning difficulties and high academic workload, lack of sleep, social adaptation, etc, and used the Naive Bayes algorithm for effective classification of students' problems [13]. Mujahid et al. investigated the tweets about online learning during Covid-19 with the help of sentiment analysis based on TextBlob, VADER, and SentiWordNet, classification by sentiments based on such vectorization of tweets as BoW, TF-IDF, and classification
techniques such as random forest, SVM, AdaBoost, different deep learning models and SMOTE algorithm usage for solving the problem with unbalanced data. LSA usage allowed us to understand the causes of students' and parents' dissatisfaction with online education during Covid-19. Primarily, it was technical problems connecting with networks and the perception of online classes as boring [14]. The tendencies of the positive relationship between pupils and negative relationship between students from universities to distance learning during Covid-19 were demonstrated in [15]. Authors used unigrams and TF-IDF for tweets vector representation and logistic regression for the classification of Arabic tweets by sentiments and achieved 89.9 % of accuracy. The investigation of tweets about distance learning and their distribution in countries by sentiments based on TF-IDF and TextBlob is presented in [16]. The extraction of the different behavioral patterns of students according to their roles in the Twitter community through their activities is considered in [17].
The research of tweets in the education field contains many open questions that are required a comprehensive study. For example, how to extract and evaluate the fear and anxieties of students, how to define criteria for evaluation of the quality of education technologies usage, teachers' competencies, and universities status based on tweets, and how to extract advantages and disadvantages of distance learning and different teaching techniques with the help of tweets, how to evaluate the involvement of students in the learning process and factors that are most influenced on involvement and academic performance.
The goal of the paper is the extraction of regularities connecting with anxieties, worries, stresses, and their causes etc based on tweets analysis during the pandemic period when distance learning was introduced.
II. RELATED WORKS
Nowadays, the most investigated are the sentiment analysis of tweets, especially, the sentiments of tweets about ^vid-19 and the vaccine relationship. The preprocessing stage includes cleaning up tweets from messages by bots, removing noisy data and insignificant words, punctuation, etc, and usage of Porter stemming for stems extraction, the extraction of patterns from tweets based on association rule, and understanding of emotions of tweets about ^vid-19 with the help of NRC are described in [18]. The choice of the feature set is very important for the solving of classification problem by sentiments. There are many approaches for feature set construction for tweets classification and clustering. Chikersal et al. used a feature set based on TF-IDF for N-grams, frequencies of part of speech contained in tweets, and the usage of different emotion lexicon calculations (Sentiment140 Lexicon, Bing Liu Lexicon, NRC Lexicon, SentiWordNet, etc), and classification algorithm such as SVM together with rule-based approach [19]. The effectiveness of BERT deep
model usage with self-attention and transformers architecture together with a fully connected layer for decision making about sentiments of tweets is described in [20].
A. The Tweets Classification by Sentiments with Deep Learning
For deep learning leveraging to classify tweets by emotion classes, researchers use various pre-trained models for word embedding and emoticons processing. The analysis of emoticons and emojis can provide additional insight into the emotional inflection of tweets. Wolny described main types of emoticons that were used typically in tweets and matched them to emotion words [21]. Khalil et al. performed the investigation of Arabic dialects in tweets and handling of emoticons with manually created lexicons, special stemming algorithm ARLSTEM, BiLSTM architecture with 3 BiLSTM layers and 3 Dropout layers between them with Relu activation function and final Dense layer with sigmoid activation function for detection of probabilities of 11 classes of emotions classes and achieved 61.5% (micro F1). Tam et al. [23] used word2vec embeddings and hybrid architecture of neural network such as ConvBiLSTM: CNN layer with Relu activation function, Max-pooling layer, BiLSTM layer, Dense layer with sigmoid activation function and achieved 91.82% (F1). Mujahid et al. [14] used sentiments evaluation based on TextBlob, VADER, SentiWordNet, and SMOTE technique for tweets sentiments balancing, realized CNN with embedding layer, Conv1D, MaxPooling1D, Flatten, Dense layers and softmax activation function, LSTM-CNN with embedding, Conv1D, MaxPooling1D, LSTM and Dense layer with softmax activation function, LSTM with embedding, dropout, LSTM and Denses layers with softmax activation function, Bi-LSTM with embedding, Dropout, Bidirectional, Dense, Dropout and Dense layer with softmax activation function for decision making and achieved 94% (accuracy) and 93% (F1). Das S. and Kolya
A.K. [24] performed the analysis of sentiments of unigrams, bigrams, trigrams, word2vec embeddings, and sequential LSTM architecture and achieved 58.5% (F1). Situla et al. [25] constructed the domain-specific feature sets, FastText, ensemble of Convolutional Neural Networks and achieved 84.46% (validation accuracy).
B. The Web-services for Tweets Analysis
We used the visualizing of the Twitter sentiment tool from Healey and Ramaswamy and extracted 246 tweets with the keyword "distance learning" from 23 October 2021 to 25 October 2021 [26]. The results of automated analysis of tweets such as the distribution of tweets according to Russell's model of emotional states and tag clouds connected with distance learning in such a period of time are presented in Figure 1 and Figure 2 respectively.
Such web tool allows us to analyze emotions from tweets based on their sentiments according to Russell's model of emotions, consider the cluster of topics and tag clouds, evaluate the frequencies of tweets with the help of heatmap, and investigate the location of authors of tweets and their activities. Each tweet is represented as a circle with such characteristics as color, brightness, size, and transparency that are meant pleasure of tweets (more pleasant tweets are represented with the help of warm colors such as green and its shades, the presentation of unpleasant tweets is presented with blue color and its shades), reliable and confidence of tweet.
Rusell's model of emotions includes a big circle with 8 sectors: from aroused to excited (or synonymously as in our research we consider emotions from active to excited), from excited to pleased (from excited to pleasant), from pleased to relaxed (from pleasant to relaxed), from relaxed to sleepy (from relaxed to subdued), from sleepy to de- pressed (subdued to depressed), from depressed to miserable (from depressed to unpleasant), from miserable to distressed (from unpleasant to stressed), from distressed to aroused (from stressed to active). As you can see from Figure 1 most part of the tweets with high confidence are located in the pleasant-subdued sector and describe such emotions as pleasant, contented, serene, relaxed, calm, and subdued. Cloud of words from tweets in Figure 2 demonstrated the satisfaction of distance learning support and free access to learning resources.
Figure 1. Tweets analysis according to Rusell's model of emotional states
edmatinn ! - ' ""U > i 1 ! il f¿h ¡ft
" B IglWMiiM = director |=3 I i • I ' I«.*- If C « Ma g £ online 1s learning 5 g ume students i = am g I Ilten coursislearn ! - s support A o Sill ssr 5
Figure 2. Tag clouds based on tweets
Unfortunately, such web-service don't allow us to extract the regularities connection with the mental health of students.
III. DATASET AND METHODOLOGY
A. Dataset Description
We used a dataset with 187052 unique tweets about distance learning during Covid-19 from Kaggle platform sharing with license CC0: Public Domain [27] and extracted from it the balanced data with 2 types of sentiments (positive, and negative) based on the VADER lexicon-based approach corrected manually. So, we extracted 37950 positive tweets, and 33525 negative tweets for our research.
The fundamental paper for our research was [28]. Baris Hasdemir scrapped tweets and performed sentiment analysis of tweets about distance learning based on lexicon-based TextBlob. He noted the prevalence of tweets with positive sentiments, and also extracted the adjective for the description of tweets about distance learning with negative sentiments such as boring, terrible, horrible, etc. The classification of tweets by sentiments was performed with the help of Bernoulli Naive Bayes with the best accuracy of 83.21 %.
B. Methodology of Research
At first, we labeled tweets as positive and negative using VADER lexicon-based algorithm based on compound score value [29]. The analysis of tweets include the following steps: 1) the hashtag extraction from tweets with negative sentiments based on regular expressions; 2) the extraction of nouns, adjectives, and verbs with negative sentiment from tweets and word clouds construction based on the most frequent of them; 3) the construction of the dictionary of the words with the most negative sentiment taking into account their frequencies; 4) the construction of the patterns based on part of speech for extraction the description and causes of affective states; 5) the feature engineering based on the most positive and negative words of various part of speech from the most positive and negative tweets (the feature sets based on the most frequent positive and negative nouns, adjectives, verbs from positive and negative tweets respectively); 6) the vectorization of tweets based on TF-IDF with the help of different custom vocabularies constructed based on proposed feature engineering, where TF-IDF is computed as
TF- IDF = TFfr log(N1DF)
TFL
where a is the number of times term i occurred in tweet
DF'
d, N is the total number of tweets in corpus, is the number of tweets in which term i occurred [30]; 7) the random forest construction for classification tweets by sentiments; 8) the comparison the results of the approaches for feature engineering for classification tweets by sentiments based on Accuracy, Recall, Precision and F1-score [31]:
TP+TN Accuracy= Tp+m+FP+ FN
Precson=
Recall=
TP
TP+ FP TP
F1- score=2
TP+ FN Precson* Recall
Precis on+ Recall
where TP, TN - number of true positive and true negative values respectively, FP, FN - number of false positive and false negative values respectively.
IV. EXPERIMENTS AND RESULTS
Sentiment analysis of students' feedback is a very important direction in learning analytics as it is connected with the detection of students' negative emotions and anxieties that have a strong influence on students' engagements and forming of their relationships with disciplines, teachers, and universities, etc.
A. The most frequent hashtags from Negative Tweets
We used VADER lexicon-based approach to detect tweets with negative sentiments. The most frequent hashtags from distance learning tweets with negative sentiments after stopwords removing are presented in Figure 3. The most frequent hashtags from tweets with negative sentiments connected with questions about pay, essay, homework and assignment, and problems with biology, maths, statistics, chemistry, and calculus.
Figure 3. The most frequent hashtags from negative tweets about distance learning
B. The Analysis of Negative Words from Tweets with Negative Sentiments
We elicited the negative words from 33525 negative tweets with the help of part of speech recognition using spacy and word cloud python-libraries as shown in Figure 4.
i&^ffence®1p® .dislike Press
-s' 1 w 1 1 V.p..1^*- ...Jt smh threatine . r impose
ii,»loss coward:,. > defeat
■fiatred^funeral^'. f^tp8 l^fVsL'^rTl^ïin I
1: a C Çu ? a t 10£l~Or hateqÄifiß'J:
• «MtlC~»
ridiodoug
disastrous Lk hellish SOrryioose
grOSSrepetitixe
« lt-Mif-i I »«i*uv* '>nii"i;
.inadequate *
fearful Petty1
B SfciSA,
COnT L
— vulnerab
Figure 4. The clouds of the most negative nouns, adjectives,
verbs
The most frequently negative words connecting with various affective states such as nouns danger, anxiety, crisis, hate, worry, defeat, stress, woe, fear, fault, accusation, offense, harm, disgust; adjectives hellish, disastrous, dead, awful, fearful, stressful, eerie and verbs hate, offend, defeat, suffer, complain, threat, kill, condemn, etc. were used for the creation of dictionary with this kind of words.
C. The Patterns Construction based on POS-tagging for Extraction the Description and Causes of Anxieties
We constructed the patterns with the help of part of speech recognition. The pattern based on the investigation of the left and right surroundings of the target word "anxiety" is presented in Figure 5.
NN(S) VBG RB *VBG|DT NN(S) f J* NNP|NN(S)
NN|JJ|VBG-►Anxiety yBG|PRP$|JJ NN(S)
NN VBG
PRP$ NN NN(S)
PDTDT DT JJ NN
VBG|PRP$ NN(S)|RB|NN(S) DT DT
Figure 5. The patterns based on part of speech recognition for target word "anxiety"
For example, the pattern noun NN(S) + gerund or present participle (VBG) + adverb (RB) or the pattern gerund or present participle (VBG) or determiner (DT) + noun (NN) allow us get anxieties causes connecting with the kids staying home or coming school in pandemic period or problems with some students.
The similar patterns were generated for target words "fear" and "worry" presented in Table 1.
Table 1. The extracted descriptions and causes of the mental disorders connected with distance learning
The target word The numbe r of tweets The description of target word The causes of of target word
anxiety 177 awful, social, parental, storm, major, zoom, crippling, real, overcome, constant, debilitating, extreme experiencing for kids staying home/coming school/ some students, at bay/ thought, of my kid/life changing/digital connec-tion/ engaging their kids, about this all/weighing our schooling/being in online, teaching in person/my fall class, distance learning, over all this/this stupid online, with our new workshop
fear 167 teachers, nagging, big(est), constant, learning, crippling of death/immigration/ young child/ infection/ being hurt / contagion/ being judged/disease/ corona virus (covid-19)/ intrusion, in anxious educators/ all this/ speaking up, for adverse effect
worry 185 parents, experts, some about distance learning/ quality/ domestic violence / virtual learning/ mana-ging decision/ your child / their mental health / losing kid / online education / hateful self / internet costs / conflicting schedules / teaching online/ writing notes / my friends
As shown in Table 1, people describe anxiety as awful, debilitating, social, extreme and connecting with kids engaging and staying at home, the changing of their life, being online/distance learning. They describe fear as nagging, crippling, constant, and fear of teachers. Their states are caused by fear of death, infection, contagion, being judged, and dealing with anxious teachers. We extracted the parents' and experts' worry about connecting with distance or virtual learning, domestic violence, internet costs, problems with mental health, schedules, communication with friends and writing notes in online mode.
D. The Classification of Tweets about Distance Learning
For the classification of 71475 tweets about distance learning by positive and negative sentiments, we used different approaches for feature engineering. We realized
vectorization of tweets based on TF-IDF with different vocabularies such as all words (vocabulary contains 41376 feature names (words)), positive nouns among positive tweets with negative nouns among negative tweets (nouns based sentiment vocabulary (sentiment_nouns) with 1766 words), positive adjectives among positive tweets with negative adjectives among negative tweets (adjectives based sentiment vocabulary (sentiment_adjs) with 636 words) positive verbs among positive tweets with negative verbs among negative tweets (verbs based sentiment vocabulary (sentiment_verbs) with 791 words) and random forest was used for classification by sentiments.
Table 2. The average score of classification metrics for 10-fold cross validation
Approach Accuracy Recall Precision F1
tf-idf+rf 0.84±0.01 0.72±0.03 0.91±0.01 0.80±0.01
TF-IDF (vocabulary = sentiment_nou ns)+RF 0.84±0.01 0.74±0.02 0.90±0.01 0.81±0.01
TF-IDF (vocabulary = sentiment adj s)+RF 0.82±0.01 0.71±0.02 0.87±0.01 0.78±0.01
TF-IDF (vocabulary = sentiment ver bs)+RF 0.82±0.01 0.73±0.02 0.87±0.01 0.79±0.01
As can be seen from the Table 2, the use of words, taking into account sentiment and parts of speech, can significantly reduce the dimension of the feature space with a slight change in the quality (accuracy, recall precision, f1 -score) of the classification, which is extremely important when classifying big data.
V. Conclusion
Sentiment analysis is a very important direction in data analysis. It allows extracting polarities from social media data and a better understanding of the mood of society and individuals. Sentiment analysis in education opens perspectives in the field of the students' emotions understanding that are connected with their engagement in the learning process and academic performance. We considered tweets with negative sentiments about distance learning and found the absence of sharply negative emotions. The anxieties were connected with homework, zoom usage, increased payment to people who can help with assignments on exact sciences, fear of infection and death, worrying about domestic violence, internet costs, and problems with mental health.
References
[1] Z. Kastrati, F. Dalipi, A.S. Imran, K.P. Nuci and M.A. Wani, "Sentiment Analysis of Students' Feedback with NLP and Deep Learning: A Systematic Mapping Study," in Applied Sciences, vol. 11 (3986), 1, pp. 1- 23, Apr. 2021.
[2] K. Gimpel, N. Schneider, B. O'Connor, D. Das, D. Mills, J. Eisenstein, M. Heilman, D. Yogatama, J. Flanigan and N.A. Smith, "Part-of-speech tagging for twitter: Annotation, features, and experiments," in Proc of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies: short papers (Association for Computational Linguistics. Portland, Oregon, USA, vol.2, pp.42-47, 2011.
[3] X. Zhou, X. Tao, J. Yong and Z. Yang, "Sentiment Analysis on Tweets for Social Events," in Proc of the 2013 IEEE 17th International Conference on Computer Supported Cooperative Work in Design (IEEE, Whistler, BC, Canada,), vol. 13710533, pp. 557- 562, 2013.
[4] V. Raghupathi, J. Ren and W. Raghupathi, "Studying Public Perception about Vaccination: A Sentiment Analysis of Tweets," in Int J Environ Res Public Health, vol. 17(10): 3464, May. 2020.
[5] J.C. Lyu, E.L. Han, G.K. and Luli, "COVID-19 Vaccine-Related Discussion on Twitter: Topic Modeling and Sentiment Analysis," in J Med Internet Res, vol. 23(6). 2021.
[6] E. Sasmaz and F.B. Tek, "Tweet Sentiment Analysis for Cryptocurrencies," in Int Conference on Computer Science and Engineering (IEEE, Ankara, Turkey), vol. 2. 2021.
[7] T.Wijesiriwardene,H.Inan,U.Kursuncu1,M.Gaur, V.L. Shalin, K. Thirunarayan, A. Sheth and I. Budak, "ALONE: A Dataset for Toxic Behavior among Adolscents on Twitter," in arXiv:2008.06465, pp. 113. 2020.
[8] A. M. Khattak, R. Batool, F. A. Satti, J. Hussain, W. A. Khan, A. M. Khan and B. Hayat, "Tweets classification and sentiment analysis for personalized tweets recommendation," in Complexity. 2020.
[9] V.A. Kharde and S. Sonawane, "Sentiment Analysis of Twitter Data: A Survey of Techniques," in Int Journal of Computer Applications, vol. 139, 5, pp.5-15. 2016.
[10] R.A. Oyekunle and R.O. Abdulkareem, "Sentiment Analysis of Students' Tweets on Unilorin CBT Examinations," in Int Journal of Applied Business and Information Systems, vol. 3(2), pp. 66-75. 2019.
[11] H. Al-Rubaiee and K. Alomar, "Clustering Students' Arabic Tweets using Different Schemes," in Int Journal of Advanced Computer Science and Applications, vol. 8(276). 2017.
[12] E.F. Felipe Bravo-Marquez and B. Pfahringer, "From Unlabelled Tweets to Twitter-specific Opinion Words," in Proc of 38th Int ACM SIGIR Conference on Research and Development (New York, USA: ACM, Santiago de Chile), pp. 743-746. 2015.
[13] X.Chen,M.Vorvoreanu and K.Madhavan, "Mining Social Media Data for Understanding Students' Learning Experiences," in IEEETrans on Learning Technologies, vol. 7(246). 2014.
[14] M. Mujahid, E. Lee, F. Rustam, P.B. Washington, S. Ullah, A.A. Reshi and I. Ashraf, "Sentiment Analysis and Topic Modeling on Tweets about Online Education during COVID-19," in Applied Sciences, vol. 11(1). 2021.
[15] M. Aljabri, S.M.B. Chrouf, N.A. Alzahrani, L. Al- ghamdi, R. Alfehaid, R. Alqarawi, J. Alhuthayfi and N. Alduhailan, "Sentiment Analysis of Arabic Tweets Regarding Distance Learning in Saudi Arabia during the COVID-19 Pandemic," in Sensors, vol. 21(1). 2021.
[16] N. Arambepola, "Analysing the Tweets about Distance Learning during COVID-19 Pandemic using Sentiment Analysis," in Proc of the International Conference on Advances in Computing and Technology , pp. 169-171. 2020.
[17] M. Yoshida, "Investigation of University Students' Behaviour in a Heterarchical Twitter Community," in Education and Information Technologies, vol. 26(1). 2021.
[18] Y. Drias and H. Drias, "Sentiment Evolution Analysis and Association Rule Mining for COVID-19 Tweets," in Journal of Digital Art and Humanities. pp. 3-21. 2021.
[19] P.Chikersal, S.Poria and,E.Cambria, "SeNTU:Sentiment Analysis of Tweets by Combining a Rule-based Classifier with Supervised Learning," in Proc of the 9th International Workshop on Semantic Evaluaion Association for Computational Linguistics, Denver, Colorado. pp. 647-651. 2015.
[20] M. Pota, M. Ventura, R. Catelli and M. Esposito, "An Effective BERT-Based Pipeline for Twitter Sentiment Analysis: A Case Study in Italian," in Sensors, vol. 21(1). 2021.
[21] W. Wolny, "Emotion Analysis of Twitter Data that use Emoticons and Emoji Ideograms," in 25th Int Conference on Information Systems Development. pp. 1 -10. 2016.
[22] E.A.H. Khalil, E.M.E. Houby and H.K. Mohamed, "Deep Learning for Emotion Analysis in arabic Tweets," in J Big Data, vol. 8(136). 2021.
[23] S. Tam, R. B. Said and O. O. Tanriover, "A ConvBiLSTM Deep Learning Model-Based Approach for Twitter Sentiment Classification," in IEEE Access, vol. 9. pp. 41283-41293. 2021.
[24] S. Das and A.K. Kolya, "Predicting the pandemic: sentiment evaluation and predictive analysis from large-scale tweets on Covid-19 by deep convolutional neural network," in Evol Intell. pp. 1- 22. Mar 2021.
[25] C. Sitaula, A. Basnet, A. Mainali and T.B. Shahi, "Deep Learning-Based Methods for Sentiment Analysis on Nepali COVID-19-Related Tweets," in Computational Intelligence and Neuroscience, vol.2021(2158184). 2021.
[26] Healey, Ramaswamy. (2022, Apr 10) Tweet Sentiment Visualization [Online]. Available: https://www. csc2.ncsu. edu/faculty/healey/tweet_viz/tweet_app/
[27] B. Hasdemir. (2022, Apr 10) Tweet about Distance Learning [Online], Available: https://www.kaggle.com/barishasdemir/ tweets-about-distance-learning
[28] B. Hasdemir. (2022, Apr 10) Sentiment analysis on the tweets about distance learning with TextBlob [Online]. Available: https://towardsdatascience.com/ sentiment-analysis-on-the-tweets-about-distance-learning-with-textblob-/cc73702b48bc
[29] C.J. Hutto, E. Gilbert, "VADER: A Parsimonious Rule-based Model for Sentiment Analysis of Social Media Text, " in Proceedings of the Eighth International AAAI Conference on Weblogs and Social Media. pp. 216-225. 2014.
[30] S. Ghosh, M.S. Desarkar, "Class Specific TF-IDF Boosting for Short-text Classification," in WWW '18: Companion Proceedings of the The Web Conference, pp. 1629-1637. 2018.
[31] F. Valencia, A. Gomez-Espinosa and B. Valdes-Aguirre, "Price Movement Prediction of Cryptocurrencies using Sentiment Analysis and Machine Learning," in Entropy, vol.21(589). 2019.
Manuscript received May 6, 2022,
Yulia Yu. Dyulicheva, V.I. Vernadsky Crimean Federal University, 4 Vernadsky Ave., Simferopol, 295007, Crimea (e-mail: [email protected]), https://orcid.org/0000-0003-1314-5367
