Artículo Científico / Scientific Paper

 

https://doi.org/10.17163/ings.n29.2023.10

 

pISSN: 1390-650X / eISSN: 1390-860X

SENTIMENTAL ANALYSIS OF COVID-19 TWITTER DATA USING DEEP LEARNING AND MACHINE LEARNING MODELS

 

ANÁLISIS SENTIMENTAL DE LOS DATOS DE TWITTER DE COVID-19 UTILIZANDO MODELOS DE APRENDIZAJE PROFUNDO Y APRENDIZAJE MÁQUINA

 

Simran Darad1,* , Sridhar Krishnan2

 

Received: 15-10-2021, Received after review: 01-12-2022, Accepted: 16-12-2022, Published: 01-01-2023

 

Abstract

Resumen

The novel coronavirus disease (COVID-19) is an ongoing pandemic with large global attention. However, spreading fake news on social media sites like Twitter is creating unnecessary anxiety and panic among people towards this disease. In this paper, we applied machine learning (ML) techniques to predict the sentiment of the people using social media such as Twitter during the COVID-19 peak in April 2021. The data contains tweets collected on the dates between 16 April 2021 and 26 April 2021 where the text of the tweets has been labelled by training the models with an already labelled dataset of corona virus tweets as positive, negative, and neutral. Sentiment analysis was conducted by a deep learning model known as Bidirectional Encoder Representations from Transformers (BERT) and various ML models for text analysis and performance which were then compared among each other. ML models used were Naïve Bayes, Logistic Regression, Random Foest, Support Vector Machines, Stochastic Gradient Descent and Extreme Gradient Boosting. Accuracy for every sentiment was separately calculated. The classification accuracies of all the ML models produced were 66.4%, 77.7%, 74.5%, 74.7%, 78.6%, and 75.5%, respectively and BERT model produced 84.2%. Each sentiment-classified model has accuracy around or above 75%, which is a quite significant value in text mining algorithms. We could infer that most people tweeting are taking positive and neutral approaches.

En este artículo, aplicamos técnicas de aprendizaje automático para predecir el sentimiento de las personas que usan las redes sociales como Twitter durante el pico de COVID-19 en abril de 2021. Los datos contienen tweets recopilados en las fechas entre el 16 de abril de 2021 y el 26 de abril de 2021, donde el texto de los tweets se ha etiquetado mediante la formación de los modelos con un conjunto de datos ya etiquetado de tweets de virus de corona como positivo, negativo y neutro. El análisis del sentimiento se llevó a cabo mediante un modelo de aprendizaje profundo conocido como Representaciones de Codificadores Bidireccionales de Transformers (BERT) y varios modelos de aprendizaje automático para el análisis de texto y el rendimiento, que luego se compararon entre sí. Los modelos ML utilizados son Bayes ingenuas, regresión logística, bosque aleatorio, máquinas vectoriales de soporte, descenso de gradiente estocástico y aumento de gradiente extremo. La precisión de cada sentimiento se calculó por separado. La precisión de clasificación de todos los modelos de ML producidos fue de 66.4%, 77.7%, 74.5%, 74.7%, 78.6% y 75.5%, respectivamente y el modelo BERT produjo 84.2%. Cada modelo clasificado de sentimiento tiene una precisión de alrededor o superior al 75%, que es un valor bastante significativo en los algoritmos de minería de texto. Vemos que la mayoría de las personas que tuitean están adoptando un enfoque positivo y neutral.

Keywords: COVID-19, coronavirus, Twitter, tweets, sentiment analysis, tweepy, text classification

Palabras clave: COVID-19, coronavirus, Twitter, tweets, análisis de los sentimientos, ttweepy, clasificación de texto

 

 

 

 

 

 

 

 

 

1,*Data Science and Analytics, Toronto Metropolitan University, Canada. Corresponding author ✉: sdarad@ryerson.ca.

2Department of Electrical, Computer, and Biomedical Engineering, Toronto Metropolitan University, Canada.

 

Suggested citation: Darad, S. and Krishnan, S. “Sentimental analysis of COVID-19 twitter data using deep learning and machine learning models,” Ingenius, Revista de Ciencia y Tecnología, N.◦ 29, pp. 108-117, 2023, DOI: https://doi.org/ 10.17163/ings.n29.2023.10.

  

1.      Introduction

There are various kinds of social media platforms that are used by users for many reasons. In recent times, the most used social media platforms for informal communications have been Facebook, Twitter, Reddit, etc. Amongst these, Twitter, the microblogging platform, has a well-documented Application Programming Interface (API) for accessing the data (tweets) available on its platform. Therefore, it has become a primary source of information for researchers working on the Social Computing domain [1]. Social Media platforms such as Twitter are a great resource to capture human emotions and thoughts. During these trying times, people have taken to social media to discuss their fears, opinions, and insights on the global pandemic [2]. For this research, we focused on a dataset that belonged to the Twitter tweets and accessed tweets related to “COVID-19 Pandemic”.

Coronavirus disease 2019 (COVID-19) was first detected in Wuhan, China, in December 2019 and has spread worldwide in more than 198 countries [3]. The outbreak of COVID-19 has a socio-economic impact. The World Health Organization declared it an epidemic on 30 January 2020. Since then, it has spread exponentially, inflicting serious health issues including painful deaths [4]. Large-scale datasets are required to train machine learning models or perform any kind of analysis. The knowledge extracted from small datasets and region-specific datasets cannot be generalized because of limitations in the number of tweets and geographical coverage. Therefore, this paper introduces a large-scale COVID-19-specific English language tweets dataset [5].

The main objective of this work is to predict people’s sentiments during the peak of the pandemic in April 2021. How can we classify coronavirus tweets as positive, negative, and neutral; which tells us about how people are feeling? So, there are two ways to label the tweets that were extracted using the Twitter API with tweepy. The first way is training already labelled data with BERT and various machine learning models, evaluating which model classifier could correctly label the tweets and then using it to label the text of the tweets extracted. The second way is to find the sentiment comes by using an open-source sentiment analyzer pre-built library known as VADER. It automatically predicts the sentiment score of the tweets classifying the tweets with the power of machine learning and using it to make inferences about the extracted tweets. Based on the classification of different tweets, the effort was to be able to provide more insights about the pandemic affecting mental health and people’s reaction about how well they are handling this situation.

1.1.        Literature Review

 

The main aim of this work is to analyze people’s reactions on the global pandemic COVID-19 via tweets and classify them as positive, negative, or neutral. This is done by performing sentiment analysis on the data obtained from Twitter. Several Machine Learning techniques have been used to obtain the results. In this section, we will provide an overview of the papers used as references for this work.

There have been many studies on this in a short span of time. To begin with, the trends of positive, negative, and neutral tweets state-wise and monthwise in India are captured and presented in this paper. Firstly, state-wise analysis is done and then the frequency of Positive, Negative, and Neutral tweets are calculated. From the analysis in this paper, it is observed that people in India were mostly expressing their thoughts with positive sentiments [1]. In another paper, a very large dataset of almost over 310 million tweets is taken into consideration. This study specifies the sentiment scores of the tweets in English language only. And it was observed that a common hashtag was being used in most of the tweets [5]. In another research work, country-wise sentiment analysis of the tweets has been done. This research work has taken into account the tweets from twelve countries gathered from 11th March 2020 to 31st March 2020. The tweets have been collected, pre-processed, and then used for text mining and sentiment analysis. The result of the study concludes that while the majority of the people throughout the world took a positive and hopeful approach, there are instances of fear, sadness and disgust exhibited worldwide [6]. Another research paper in which the BERT model was used to analyze the sentiments behind tweets made by netizens of India. There were several common words that came out in the analysis and based on that the tweets are classified into four sentiments such as fear, sad, anger, and joy. This model was 89% accurate as compared to other models like LR, SVM, LSTM [7]. A short research aimed at analyzing the sentiments and emotions of people during COVID-19 was conducted based on the tweets from March 11th to March 31st, 2020, which gave us the results that the mindsets of people was almost at the same level all around the world [8].

There have been few papers in which the exploratory analysis of the data is performed to obtain the results. For instance, in a research paper, exploratory data analysis was performed for a dataset providing information about the number of confirmed cases on a per-day basis in a few of the worst-hit countries to provide a comparison between the change in sentiment with the change in cases since the start of this pandemic till June 2020 [2]. In this paper, the authors have tried to understand and analyze the tweets around COVID-19 in India and have tried to analyze

 

 

these data using NVIVO processors and word cloud. The study involves the words, hashtag being used and the sentiments involved around these words. The conclusion gives an understanding of high-impact and low-impact words [9]. In this research paper, data is collected from the users who shared their location as ‘Nepal’ between 21st May 2020 and 31st May 2020. The result of the study concluded that while majority of the people of Nepal took a positive and hopeful approach, there are instances of fear, sadness and disgust exhibited too [10].

Since Twitter is a place where individuals can express their views without revealing their identity, this is used as an advantage by many of them to present their opinions either positive negativelyive based on their sentiments. By using various Machine Learning techniques and knowledge from social media, sentiment analysis on COVID Twitter data was performed, which gave us the results as positive or negative. Logistic Regression Algorithm was used to perform the analysis which gave an accuracy up to 78.5% [11].

Data mining was conducted on Twitter to collect a total of 107,990 tweets related to COVID-19 between December 13, 2019, and March 9, 2020. A Natural Language Processing (NLP) approach and the latent Dirichlet allocation algorithm were used to identify the most common tweet topics as well as to categorize clusters and identify themes based on the keyword analysis. The results indicate the main aspects of public awareness and concern regarding the COVID-19 pandemic. First, the trend of the spread and symptoms of COVID-19 can be divided into three stages. Second, the results of the sentiment analysis showed that people have a negative outlook toward COVID-19 [12]. In this paper, our aim is to perform a sentimental analysis of tweets during the COVID-19 pandemic and classify them as positive, negative, or neutral.

After learning about the dataset, the next step was to solve the classification problem. The classification problem in this paper is sentiment analysis. Many of the papers already mentioned earlier [1, 5] performed sentiment analysis on tweets to classify them in three different categories. These research papers provided vital information about how sentiment analysis can be performed for the classification of tweets in the dataset. Creating a classifier was the next step. “The impact of preprocessing on text classification” is a resourceful paper that provided details and leads on how to conduct preprocessing on data and which classifier would be optimal. It mentions that SVM is state-of-the-art pattern classifier and is recommended to be used as the classification algorithm [13]. The papers use Random Forest, Naïve Bayes, SVM, and Random Forest for classification and tells us that Linear SVM provided the best results. Almost 95% accuracy was achieved using this technique. Based on this research, we have decided to use Naïve Bayes, Logistic Regression, Random Forest, SVM, SGD, XGB and BERT.

Before moving further to the dataset, it is important to know about the dataset and learn as much about it as possible. A detailed exploratory analysis of the dataset was conducted using reference from various papers. 

 

2.      Materials and Methods

 

2.1. Material

 

Data for this work is acquired from Twitter using its API and tweepy. Tweepy is an open-source and easyto-use Python package for accessing the functionalities provided by the Twitter API. Tweepy includes a set of classes and methods that represent Twitter’s models and API endpoints, and it transparently handles various implementation details, such as: Data encoding and decoding. Data extraction of tweets from Twitter API is done from date 16th April 2021 to 26th April 2021 containing 2,00,000 tweets to get a bigger dataset and better results.

The other dataset is open-sourced and collected from a blog [14] which contains coronavirus tweets with labelled sentiments. The dataset that has been collected for tweets by the blog was a labelled sentiment analysis dataset. This dataset was split into two subsets for training and testing of the various classifiers. The dataset we gathered and fetched from Twitter is unlabelled. 

 

2.1.1.  Descrptive Analytics

 

The dataset contains text fields, so text analysis of the tweets was performed as outlined below. But before that analysis was conducted to learn more about the dataset. Firstly, even before the cleaning process, one should get familiar with the kind of data they’ll be dealing with. This just helps in providing more context and background information to the data scientist. So, after loading the csv file, a few functions were run on the data just to familiarize with it. We get to know the size of the data, the data types of each column, the number of null entries, the distribution of different classes, etc. Next is dropping duplicate rows if any. We then realize that we won’t be needing a few columns in further analysis, so we drop it.

After those preprocessing techniques were applied to the data to clean the tweets. It includes converting the text to lowercase, tokenization, and removal of username tags, retweet symbol, hashtags, white spaces, punctuations, numbers, emoji, and URLs to clean the text. Using this clean text, further text analysis was conducted as outlined below. The analysis was conducted on the dataset we collected from Twitter API with 200,000 tweets (Figure 1).

 

 

Figure 1. Dataset size

 

We look at the information of the dataset. It tells us about the type of field it is and about how many non-null values are present in the dataset, which helps us understand our dataset better (Figure 2).

 

Figure 2. Dataset information

 

With social media, one can never retrieve all the data. There are always some missing values in the dataset. People like to keep few things discreet such as their location and description in case of twitter. Also, some people as we can see are not comfortable of using hashtags see Figure 3.

 

Figure 3. Total null values

 

Then finding out what is term frequency of the words showing the most frequently used words by their

count. We see that “COVID-19” is the most used word (Figure 4).

 

Figure 4. Top words used in tweets

 

To get a closer look at the text contained in the dataset, a visualization of the word cloud was created (Figure 5).

 

Figure 5. Word Cloud for top 50 most used words

 

The word cloud above lists all words with the top 50 most used words. Word clouds are useful in understanding what the users are posting about. Most of the words are related to COVID, and new cases, and it seems like most people posted about vaccines as well (Figure 6).

After looking at an overview of the tweet text in our corpus, let’s move on to hashtags looking for the most trending ones

 

Figure 6. Word Cloud for Hashtags

 

The word cloud above lists all words with extremely used hashtags. Word clouds are useful in understanding what the users are posting about. Most of the words are related to COVID, new cases, and it seems like most people posted about vaccines as well.

Figure 7 shows the location of the people from where most of them are tweeting. We can see a large number of people are tweeting from India and USA, as the time period selected for extracting the tweets

 

 

was during the third wave and the number of cases was higher in those countries.

 

Figure 7. Top 25 locations where tweets originate from

 

Figure 8 shows which verified users tweeted the most about COVID. We can see that almost all of them are news channels tweeting about the latest updates about COVID and the number of cases in their respective countries.

 

Figure 8. Top 20 user-verified tweets

 

After looking at an overview of the data, we clean and preprocess the text of the tweets in our corpus, moving on to do some n-gram analysis. N-grams provide a better context of what the users are posting about as we move to bi and trigrams because these provide the most frequent phrases instead of just words. Figure 9 shows that most frequent unigrams are based on new cases, vaccines, health, pandemic, people, availability and appointments.

 

Figure 9. Top 20 Unigrams

 

A bigram (Figure 10) analysis provides further details trending during that time giving details about availability of vaccine appointments, new cases and second wave.

Figure 10. Top 20 Bigrams

 

A trigram (Figure 11) analysis provides further details on where the COVID new vaccine appointments are available. It seems like most of them are in Walgreens which is an American company that operates as the second-largest pharmacy store chain in the United States behind CVS Health.

 

Figure 11. Top 20 Trigrams

 

2.2. Methods

 

The aim of this study is to train text from the labelled tweets that could automatically assess if the unlabelled tweet gathered is positive, negative, or neutral. After training the models on labelled twitter data, models were applied to data extracted to label the sentiments and compare the results of different algorithms. The second method of tweets labelling is done by using NLTK VADER inbuilt python package based on lexicons.

In this work, the response is labelling the tweets as positive, negative or neutral. The dataset gathered contains a lot of information on the user such as name, description, followers, friends and many more but only the text of the tweet was used to label the data from training the existing labelled data.

 

2.2.1.      Experimental Design 1

 

It is very difficult to label the sentiments for COVID-19 data because of the words used to represent the situation. For example, if there are new cases, there is a tweet saying “I am tested Corona positive” which ML technically would label as positive. So, there is a huge uncertainty in predicting the sentiments of the pandemic. Therefore, we apply two different techniques to understand the sentiments.

 

 

a)      Text Processing

The dataset called “coronavirustweets” contained labelled data of tweets showing the sentiment as extremely positive, positive, neutral, negative and extremely negative. Narrowing down the categorical labels to only three-class classifications, there is neutral, negative combined with extremely negative, and positive combined with extremely positive to achieve greater accuracy. The text from the original tweet needs to be pre-processed to train and test the data by removing punctuations, stop words, spaces, emoticons and stemming the data.

The preprocessing of the text data is an essential step as it makes the raw text ready for mining. The objective of this step is to clean text irrelevant to search the sentiment of tweets such as punctuation(.,?,”etc.), special characters (,%,&,$, etc.), numbers (1,2,3, etc.), Twitter handle, links(HTTPS: / HTTP:) and stop words which don’t mean anything in context to the text.

Stop words are those words in natural language that have very little meaning, such as “is”, “an”, “the”, etc. To remove stop words from a sentence, the text can be divided into words and then remove the word if it exists in the list of stop words provided by NLTK.

 

b)      Randomization

 

The dataset was randomly divided into two sets stratifying with sentiment values of the dataset, one for training with 80% data and another for testing with 20% data.

 

c)      Vectorizing the tweets

 

Before we implement different ML text classifiers, we need to convert the text data into vectors. It is crucial as the algorithms expect data in some mathematical for rather than textual form. Count Vectorizer counts the number of times a word appears in the document (in each tweet). This process helps in converting the text data as we understand it, to numerical data, that is easier for the computer to understand.

 

d)      Classifiers

 

After vectorizing the tweets, we are all set to implement classification algorithms. There are three types of sentiments so we must train our models so that they can give us the correct label for the test dataset. We have built different machine learning models such as Naive Bayes, Logistic Regression, Random Forest, Support Vector Machine, Stochastic Gradient Descent and Extreme Gradient Boosting along with BERT, a deep learning model. Ensemble Classifier such as bagging and boosting are applied on the dataset as well to minimize any over-fitting by the classifiers.

 

We use the accuracy score to measure the performance of the model (precision score, recall and confusion matrix are also calculated). Precision score, recall and confusion matrix let us know how correctly labelled the actual values are. BERT(bi-directional Encoder Representation of Transformers) is a technique developed by Google based on the Transformers mechanism. In our sentiment analysis application, our model is trained on a pre-trained BERT model. BERT models have replaced the conventional RNN based LSTM networks which suffered from information loss in large sequential text [15]. The results from paper explained that a language model that is bi-directionally prepared can have a more profound feeling of language setting and stream than single directional models. In contrast to directional models that enable sequential reading of text input (right to left or left to right), the transformer encoder recognizes the total sequence of words at once. Thus, it is considered bidirectional, but it is a non-directional model with higher accuracy than other established models [7].

 

e)          Labelling new tweets

 

Since our collected data is not labelled, we save and load our trained models with pickle. This allows us to save our model to a file and load it later in order to make predictions. We can then apply them to label the data we extracted and preprocessed.

 

f)          Comparing algorithms

 

Obtaining the sentiments of the tweets from different models and saving the csv files of different models, we compare the results of the labelled data.

 

2.2.2.      Experimental Design 2

 

VADER stands for Valence Aware Dictionary and sentiment Reasoner. VADER belongs to a type of sentiment analysis that is based on lexicons of sentimentrelated words. It is a rule-based model for general sentiment analysis, and its effectiveness was compared to 11 typical benchmarks, including Word Count (LIWC), Affective Norms for English Words (ANEW), the General Inquirer, Linguistic Inquiry, Senti WordNet, and machine learning techniques that rely on Support Vector Machine (SVM) algorithms, Naive Bayes, and Maximum Entropy. In this approach, each of the words in the lexicon is rated as to whether it is positive or negative, and in many cases, how positive or negative.

VADER performs well in the analysis of sentiments expressed in social media. These sentiments must be present in the form of comments, tweets, retweets, or post descriptions, and it works well on texts from other domains also. We design our VADER sentiment model, which extracts features from Twitter data, formulates

 

 

 

the sentiment scores, and classifies them into positive, negative, neutral classes.

 

a)      Data Cleaning

 

The dataset extracted from the tweet needs the text to be pre-processed by removing punctuations, stop words, spaces, emoticons and stemming the data.

 

b)      Finding Polarity

 

The compound score (polarity) is computed by summing the valence scores for each word in the lexicon, adjusted according to the rules, and then normalized to be between -1 (most extreme negative) and +1 (most extreme positive).

 

c)      Finding Sentiments

 

After getting the compound scores, the polarity of the tweets is used to categorize them into 3 classes: Positive, Negative and Neutral. Positive Sentiments are those with a score above 0. Negative sentiments from less than 0, and neutral sentiments are having 0.0 polarity. These 3 classes were stored along with the tweets in the dataset called “Sentiments”.

 

3.         Results and discussion

 

3.1.       Results

 

3.1.1.  Experimental Result 1

 

Multi-class classification on different models was applied to train data to find the accuracy of the correct label for the test dataset. I have built different ML models like Naive Bayes, Logistic Regression, Random Forest, Support Vector Machine, Stochastic Gradient Descent and Extreme Gradient Boosting (Figure 12).

We have observed that the Stochastic Gradient Descent classifier gives the best result with accuracy reaching 78.64%. The accuracy is pretty much close to the accuracy of Logistic Regression, and both models can be used to predict the sentiment of unlabelled data. The least accuracy is shown by Naïve Bayes Classifier. It works well with large data. Naïve Bayes works on n-grams, I have tried using different n-grams, but accuracy stays around 65%.

 

Figure 12. Comparison of model accuracies

 

The BERT model performs extremely well in comparison to other ML models. It gives an accuracy score of 84.2%, which is the highest accuracy we got by training and testing the models. BERT is an excellent and different technique, which provides the best accuracy because it is designed to read in both directions at once. This capability, enabled by the introduction of Transformers, is known as bi-directionality. BERT, however, was pre-trained using only an unlabeled, plain text corpus (namely the entirety of the English Wikipedia, and the Brown Corpus). It continues to learn unsupervised from the unlabeled text and improve even as its being used in practical applications (ie Google search). Its pre-training serves as a base layer of "knowledge" to build from. From there, BERT can adapt to the evergrowing body of searchable content and queries and be fine-tuned to a user’s specifications. This process is known as transfer learning [16].

Next using this trained model on our dataset, we see the following results based on the test accuracy (Figure 13).

 

SGD Classifier gives:

Neutral: 65462

Positive: 48785

Negative: 35741

Figure 13. SGD Classifier results

 

Stochastic Gradient Descent is a simple yet very efficient approach to fitting linear classifiers and regressors under convex loss functions. SGD has been successfully applied to large-scale and sparse machine

 

 

learning problems often encountered in text classification and natural language processing, which is why it performs better than all other models (Figure 14).

 

LG Classifier gives:

Neutral: 62263

Positive: 49022

Negative: 38703 

Figure 14. LG Classifier results

 

Seeing the results, we observe Logistic Regression gives more Positive and Negative labelled tweets whereas Stochastic Gradient Boosting predicts some of them as Neutral. Even though the accuracy for these both is almost the same, there is different labelling of approximately 3,000 tweets as neutral. Multinomial logistic regression is an extension of logistic regression that adds native support for multi-class classification problems. Logistic regression, by default, is limited to two-class classification problems, which is why SGD is better in accuracy for predicting sentiments.

Ensemble Classifier such as bagging and boosting are applied on the dataset to minimize any over-fitting by the classifiers. But there isn’t any over-fitting of the data because the accuracy obtained by bagging is 72.1% which is around the same whereas accuracy of boosting is 51.4% which is pretty much lower.

A similar analysis has been presented in [17] for the understanding of pandemic anxiety among Twitter users based on keywords. About 900,000 tweets were extracted from Twitter Application programming interface (API) and analysed using Naïve Bayes and logistic regression models. The model accuracy that appeared in short tweets is 91% and 74%, respectively. However, the main limitation of this study is all sentiments depend on the single word “fear” of only USA citizens and they are short tweets [7].

 

3.1.2.  Experimental Result 2

 

VADER sentiment model is an automatic labelling technique in which we formulated the sentiment score by classifying the tweets as positive, negative and neutral. The main difference we observe here is, it gives

 

fewer (around 5000 fewer) neutral tweets and classifies them as positive and negative. We can see that it almost matches our trained labelled models accuracy by showing us the results as follows (Figure 15):

 

Figure 15. VADER results

 

 3.2.     Discusssion

 

This study can be used to analyze the changing sentiments of people from all over the world and check whether there are major shifts in them over the period of time along with the increased supply of vaccinations. It is expected that as the spread of this pandemic will increase for unvaccinated people, the sentiments in the tweets to positive almost entirely as things get back to normal.

A similar analysis was conducted using TextBlob as we did in Experiment 2 but we used VADER. But according to the TextBlob documentation, TextBlob takes advantage of Naïve Bayes (NB) model for classification. NB classifier has been trained on NLTK (Natural Language Tool Kit) to detect the valence of aggregated tweets [10]. As we saw Naïve Bayes gives the least accuracy so Textblob is not accurate for labelling sentiments. Classical ML methods provided an accuracy of high 70%, whereas the deep learning model that uses BERT provided an impressive accuracy rate of 84.2%.

 

4.      Conclusions

 

The results of the study conclude that majority of the people throughout the world took a positive and hopeful approach. However, countries such as India and United States of America have shown signs of biggerscale tweeting due to the third wave as compared to remaining countries.

We used two techniques for our dataset to get the labels, but as we show, there is always some margin of error in text classification. We also show that BERT requires high computational power, GPU and a large time to train on a model. The prediction of any social media text is nearly impossible to give a perfect accuracy score. Through this, we can learn the main issue to help the healthcare providers to identify some kind of mental illness before it’s too late.

 

 

 

Acknowledgements

 

The authors would like to thank Ryerson University for supporting this project.

 

References

 

[1] T. Vijay, A. Chawla, B. Dhanka, and P. Karmakar, “Sentiment analysis on covid19 twitter data,” in 2020 5th IEEE International Conference on Recent Advances and Innovations in Engineering (ICRAIE), 2020, pp. 1–7. [Online]. Available: https://doi.org/10.1109/ICRAIE51050.2020.9358301

[2] M. Mansoor, K. Gurumurthy, A. R. U, and V. R. B. Prasad, “Global sentiment analysis of COVID-19 tweets over time,” CoRR, vol. abs/2010.14234, 2020. [Online]. Available: https://doi.org/10.48550/arXiv.2010.14234

 

[3] H. Drias and Y. Drias, “Mining twitter data on covid-19 for sentiment analysis and frequent patterns discovery,” medRxiv, 2020. [Online]. Available: https://doi.org/10.1101/2020. 05.08.20090464

 

[4] F. Rustam, M. Khalid, W. Aslam, V. Rupapara, A. Mehmood, and G. S. Choi, “A performance comparison of supervised machine learning models for covid-19 tweets sentiment analysis,” PLOS ONE, vol. 16, no. 2, pp. 1–23, 02 2021. [Online]. Available: https://doi.org/10.1371/journal.pone.0245909

 

[5] R. Lamsal, “Design and analysis of a large-scale COVID-19 tweets dataset,” Applied Intelligence, vol. 51, no. 5, pp. 2790–2804, May 2021. [Online]. Available: https://doi.org/10.1007/s10489-020-02029-z

 

[6] A. D. Dubey, “Twitter sentiment analysis during covid-19 outbreak,” SSRN, 2021. [Online]. Available: https://dx.doi.org/10.2139/ssrn.3572023

 

[7] N. Chintalapudi, G. Battineni, and F. Amenta, “Sentimental analysis of COVID-19 tweets using deep learning models,” Infect Dis Rep, vol. 13, no. 2, pp. 329–339, Apr. 2021. [Online]. Available: https://doi.org/10.3390/idr13020032

 

 

[8] M. A. Kausar, A. Soosaimanickam, and M. Nasar, “Public sentiment analysis on twitter data during covid-19 outbreak,” International Journal of Advanced Computer Science and Applications, vol. 12, no. 2, 2021. [Online]. Available: http: //dx.doi.org/10.14569/IJACSA.2021.0120252

 

[9] A. Mitra and S. Bose, “Decoding Twitter-verse: An analytical sentiment analysis on Twitter on COVID-19 in india,” Impact of Covid 19 on Media and Entertainment, 2020. [Online]. Available: https://bit.ly/3YMj1c3

 

[10] B. P. Pokharel, “Twitter sentiment analysis during covid-19 outbreak in nepal,” SSRN, 2020. [Online]. Available: https: //dx.doi.org/10.2139/ssrn.3624719

 

[11] C. R. Machuca, C. Gallardo, and R. M. Toasa, “Twitter sentiment analysis on coronavirus: Machine learning approach,” Journal of Physics: Conference Series, vol. 1828, no. 1, p. 012104, feb 2021. [Online]. Available: https: //dx.doi.org/10.1088/1742-6596/1828/1/012104

 

[12] S. Boon-Itt and Y. Skunkan, “Public perception of the COVID-19 pandemic on twitter: Sentiment analysis and topic modeling study,” JMIR Public Health Surveill, vol. 6, no. 4, p. e21978, Nov. 2020. [Online]. Available: https://doi.org/10.2196/21978

 

[13] A. K. Uysal and S. Gunal, “The impact of preprocessing on text classification,” Information Processing & Management, vol. 50, no. 1, pp. 104–112, 2014. [Online]. Available: https://doi.org/10.1016/j.ipm.2013.08.006

 

[14] S. Gujral, “Sentiment analysis: Predicting sentiment of COVID-19 tweets,” Analytics Vidhya, 2021. [Online]. Available: https://bit.ly/3j9tMVj

 

[15] ——, “Amazon product review sentiment analysis using bert,” Analytics Vidhya, 2021. [Online]. Available: https://bit.ly/3Vad9WE

 

[16] B. Lutkevich. (2022) Bert language model. TechTarget Enterprise Al. [Online]. Available: https://bit.ly/3Wo5Pb4

 

[17] J. Samuel, G. G. M. N. Ali, M. M. Rahman, E. Esawi, and Y. Samuel, “Covid-19 public sentiment insights and machine learning for tweets classification,” Information, vol. 11, no. 6, 2020. [Online]. Available: https://doi.org/10.3390/info11060314