1.1. Background

Sentiment analysis has gained significant importance in the current era due to the global utilization of various social media platforms as a means for individuals to express their opinions. Individuals engage in self-expression by posting comments on various social media platforms such as Instagram, Facebook, Twitter, and others [8]. The individuals discuss many subjects, such as films, commodities, prevailing fashion, political affairs, and technological advancements. The individual’s viewpoint can be categorized as good, negative, neutral, or spam. Multiple polarity scales can be utilized to predict scores [9]. There should be a process to further expound upon these comments, which is sentiment analysis.

Sentiment analysis [10], a component of Natural Language Processing, is commonly referred to as opinion mining due to its focus on examining users’ emotions, feelings, and opinions. Sentiment analysis serves a broader range of applications beyond opinion prediction, encompassing the identification of fraudulent tweets [11]. The user’s text does not contain any information. The categorization methodologies employed in this study consist of three levels: feature, document, and sentence [12].

The application of sentiment analysis necessitates the utilization of not one but two separate methodological approaches. The first dictionary or lexicon is based on traditional methods, whereas the second dictionary or lexicon is based on techniques used in machine learning [13–17]. The ability of machine learning techniques to process and successfully evaluate large datasets in an automated manner has led to their increased prevalence in today’s technological landscape. Several different algorithms based on machine learning are utilized to perform sentiment analysis, and these approaches have shown promising results. Techniques such as the Naive Bayes classification, Bootstrap aggregation, Linear Discriminant Analysis, Decision Trees, and Support Vector Machines are some examples of this type of methodology [18]. Methodologies for machine learning (ML) can be broken down into one of three categories: unsupervised, semi-supervised, or supervised [19]. When it comes to sentiment analysis, supervised learning algorithms are extremely popular and have a high level of regard. The implementation requires the extraction of sentences and features at several different levels, including feature level and sentence level. The feature set consists of a bag of words as well as tags for the various parts of speech, unigrams, bigrams, and n-grams.

The limitations that are imposed on the processing capabilities of machine learning have led to the development of a more sophisticated strategy known as deep learning (DL). In 2006, G.E. Hinton was the first person to suggest the idea of "Deep Learning." He did so by introducing a representation of a Machine Learning job that specifically pertains to Deep Neural Networks [13]. When he did this, he also created the concept of "Deep Learning." The functional model of the human brain is used as the basis for the construction of the Artificial Neural Network (ANN) [20]. This model describes how several neurons work together to facilitate the processing of information. The capacity of Deep Learning models to provide training for both supervised and unsupervised categories is one reason why these models are used. Word representation, sentence categorization, and text synthesis are three examples of the natural language processing tasks that have been shown to benefit significantly from the utilization of neural networks. Deep learning has played an important role in the expansion of computational capabilities as well as the development of the field of software engineering [16]. Because of its widened computing capabilities, the system is able to successfully evaluate and manage huge amounts of data utilizing several neural networks, such as the Deep Forward Neural Network, which has led to a significant improvement in accuracy [21, 22]. Convolutional Neural Networks (CNN) [23], Recurrent Neural Networks (RNN), Deep Belief Networks (DBN), Long Short-Term Memory (LSTM), and Bidirectional LSTM (BiLSTM [24]), amongst other methods, are all utilized in Deep Learning for the purpose of doing sentiment analysis. F-measure, recall, and precision are the objective measures used by machine learning and deep learning algorithms to analyze the effectiveness of a classifier. These three criteria help select the best appropriate model to address a specific issue and determine how well it addresses it.

2.1. Machine learning and deep learning algorithms for sentiment analysis

There exists quite a lot of literature in the current state of the art which has focused on the application of machine learning or deep learning methods for sentiment analysis on Twitter [25] examined sentiment in large-scale social data. Social media platforms are becoming more popular for expressing opinions on many topics. Integrated CNN, LSTM, and DNN (Deep Neural Network) are used in this article to provide a new sentiment analysis strategy. ICNN-LSTM-DNN is used for social data sentiment analysis and opinion mining during the 2019 Indian general election. This study tests the algorithm using twitter (now X) data. The above strategy was 89% accurate. On the other side, Gopi et al. [18] studied the Twitter (now X) sentiment analysis. Data polarity is determined in sentiment analysis. Polarity definitions are often asked. Positive, negative, or neutral data is used in polarity opinion mining. In this study, researchers predicted sentiment using a multi-polarity scale. The scale ranged from -5 to +5. Another prominent research by Ali et al. [26] identified fraudulent tweets. Twitter (now X) is the most popular microblogging network, allowing people to share their opinions on a variety of topics with substantial weight and effect. Numerous bogus accounts influence the topic’s conversation. These accounts must be distinguished from legitimate and authorized ones. The classifier flags fake accounts using a Custom Rule-Based algorithm. Other machine learning methods including Multi-Layer Perceptron (MLP), Decision Tree, and Random Forest are used and contrasted. The researchers used Twitter (now X) API (application programming interface) user data. A data frame was created from up to 200 tweets per person. The approach outperformed MLP, Decision Tree, and Random Forest with 97% accuracy on the same dataset. The current body of research is replete with numerous studies on sentiment analysis.

In the landscape of COVID-19 research, multiple articles delve into different facets, including prediction, diagnosis, and sentiment analysis. Indumathi et al. [27] proposed a machine learning algorithms to predict COVID-19 affected zones in Virudhunagar district, achieving an impressive accuracy of 98.06%. Ayalew et al. [28] introduce a unique detection and classification approach (DCCNet) for COVID-19 using chest X-ray images, exhibiting a remarkable 99.9% training accuracy and 98.5% test accuracy. Prasad et al. [29] suggest a cloud-based image analysis approach (CIA-CVD) for COVID-19 vaccine distribution, integrating deep learning, machine learning, digital image processing, and cloud solutions. Misra et al. [30] present a Parallel Ensemble Transfer Learning based Framework for COVID (PETLFC) utilizing chest X-ray images, demonstrating superior performance compared to sequential ensemble approaches. Ayalew et al. [31] employ a Convolutional Neural Network (CNN) for COVID-19 detection from chest X-ray images, achieving high training and test accuracy. Salau et al. [32] roposes a Support Vector Machine (SVM) method with a discrete wavelet transform (DWT) algorithm, attaining a detection rate of 98.2%. Furthermore, Demilie and Salau [33] address the pervasive issue of hate speech detection on social media platforms, emphasizing the need for advanced research and optimal approaches in this challenging domain. Gupta et al. [34]contribute to the sentiment analysis domain, mining the sentiments of Indian citizens regarding the nationwide lockdown during the COVID-19 outbreak, with a notable accuracy of 84.4%. Jayasurya et al. [35] conduct sentiment analysis on the topic of COVID-19 vaccination, employing 14 different machine learning classifiers and revealing insightful temporal and spatial analyses of textual data. Collectively, these articles underscore the interdisciplinary efforts required to comprehensively combat and understand the multifaceted impacts of COVID-19, utilizing advanced technologies and methodologies across various domains, from healthcare to social analysis. A comprehensive summary of these studies is provided in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Current state-of-the-art around sentiment analysis.

https://doi.org/10.1371/journal.pone.0315407.t001

2.2. Recent advances with Large Language Models

Large Language Models (LLMs) and transformers, including BERT, RoBERTa, LLMA 2, and GPT [5, 49], have emerged as essential tools in the domain of fake news detection, demonstrating substantial contributions to improving detection accuracy. Szczepański et al. [50] provided an in-depth study of BERT’s performance, introducing an innovative approach to enhance explainability in fake news detection. Their model utilized Local Interpretable Model-Agnostic Explanations (LIME) and Anchors, enabling users to understand the rationale behind BERT’s classification decisions. This advancement emphasized the need for transparency in AI-driven systems, highlighting that high accuracy alone is insufficient in critical real-life applications, such as disinformation detection. Through experimentation with fake news datasets, including tweets and headlines, they demonstrated that these explainability techniques could be seamlessly integrated into BERT models without affecting performance. Their findings showed that explainability enhances user trust, particularly in high-stakes environments, where understanding decision-making processes is as important as the outcome itself.

RoBERTa, an optimized version of BERT, further advanced the field by refining pretraining on larger datasets and for extended durations. Pavlov et al. [51] explored RoBERTa’s effectiveness in detecting fake news specifically related to COVID-19. By fine-tuning RoBERTa on a dataset comprising real and fake COVID-19-related tweets, they achieved notable improvements in accuracy, recall, and F1 score when compared to standard BERT models. Their research revealed that RoBERTa’s ability to capture subtle linguistic nuances in misinformation made it particularly effective in distinguishing real from fake COVID-19 news. This study demonstrated that RoBERTa, due to its extended training and larger data exposure, not only enhances detection performance but also improves the robustness of the model in real-world, noisy social media environments. The comparative results with BERT underscore the advantages of extended pretraining in improving model generalization across diverse datasets.

In addition to detection, GPT models have been pivotal in simulating fake news generation, as demonstrated by Dhiman et al. [52]. Their work introduced GBERT, a hybrid deep learning framework that combines GPT’s generative capabilities with BERT’s deep contextual understanding. By fine-tuning both models on real-world benchmark datasets, the authors achieved superior accuracy (95.30%), precision (95.13%), and F1 score (96.23%). The generative aspect of GPT allowed the creation of high-quality, diverse fake news samples, which were then used to train the detection model, thus enhancing its ability to handle adversarial examples. This approach demonstrated the utility of combining generative and discriminative models to address the evolving challenge of misinformation, particularly in the era of LLMs. The statistical tests conducted in the study revealed that the hybrid model not only improved performance but also increased robustness against various types of fake news, establishing a promising direction for future research in the field.

Su et al. [49] identified another critical issue with LLM-generated text—bias in fake news detectors. Their research revealed that existing detection models tend to misclassify human-written fake news as genuine while disproportionately flagging LLM-generated content as fake. This bias was attributed to the distinct linguistic patterns produced by LLMs, which detectors could easily identify. To mitigate this, the authors employed adversarial training using LLM-paraphrased genuine news, significantly improving detection accuracy across both human- and machine-generated news. Their introduction of two novel datasets, GossipCop++ and PolitiFact++, consisting of human-validated and LLM-generated news, provides a valuable resource for future studies. This work highlights the importance of addressing bias in fake news detection, particularly as LLMs become more adept at generating human-like content.

The recent advancements in fake news and fake review detection have further demonstrated the utility and limitations of transformer-based models. Qin and Zhang’s study [53] explores the challenge of generalization in BERT fine-tuning for fake news detection, a common issue that arises due to overfitting on the training data. Their research identifies that BERT’s fine-tuning often leads to reduced generalization capability, making it difficult to detect unseen fake news accurately. To address this, they propose a novel adversarial fine-tuning strategy, FGM-FRAT, which treats fine-tuning as a mini-max optimization problem. By applying adversarial training on the word vector space and incorporating feature regularization, the model enhances its robustness against unseen data. Extensive experiments on benchmark datasets, including FakeNewsNet and KaggleFakeNews, demonstrated that the proposed method improves generalization accuracy from 61% to 73%, significantly outperforming conventional fine-tuning methods. This shows that while BERT-based models have shown great promise, careful fine-tuning strategies are essential to extend their applicability to diverse datasets and new scenarios.

In contrast, Mohawesh et al. [54] tackle the issue of fake review detection by combining RoBERTa with an LSTM layer to build a more semantic- and linguistic-aware model. Their approach focuses on overcoming the limitations of traditional machine learning models, which primarily rely on linguistic features and often fail to capture deeper semantic meanings within text. The authors integrated RoBERTa’s transformer capabilities with LSTM, the model can process elaborate temporal dependencies within the data. The authors further enhance the model’s performance by incorporating Shapley Additive Explanations (SHAP) and attention mechanisms to ensure transparency in classification decisions. Their experiments on the OpSpam and Deception datasets demonstrated that this hybrid model outperformed state-of-the-art methods, achieving 96.03% and 93.15% accuracy, respectively. The use of SHAP and attention mechanisms also provided clearer insights into why specific reviews were classified as fake, adding an important layer of interpretability. This highlights the critical role that hybrid models and explainability techniques play in ensuring robust and transparent detection systems, especially in high-stakes contexts like consumer behavior.

In summary, the integration of Large Language Models like BERT, RoBERTa, and GPT into fake news detection has significantly advanced the field, offering both improved performance and new methodological insights. The works of Szczepański et al. [50], Pavlov et al. [51], Dhiman et al. [52], and Su et al. [49] demonstrate that these models enhance detection accuracy through deep contextual understanding and robust training, and on the other side, address critical issues such as explainability and bias.

2.3. Fake news resistant society through media literacy and artificial intelligence tools

Creating a society resistant to fake news necessitates a comprehensive approach to media literacy, emphasizing critical thinking, verification skills, and awareness of information sources. Media literacy education empowers individuals to critically evaluate content, discerning between credible information and misinformation. In this context, the integration of Artificial Intelligence (AI), particularly through sentiment analysis, can significantly enhance media literacy efforts. Sentiment analysis, a subset of AI, involves analyzing texts to determine the sentiment behind them, offering insights into the biases and emotional undertones that may indicate biased or manipulative content. By using AI tools that apply sentiment analysis, educators and learners can more effectively identify the emotional manipulation often associated with fake news. This not only facilitates a deeper understanding of how information is constructed to influence public opinion but also equips individuals with the analytical tools to question and verify the authenticity of the content they encounter. Consequently, the use of AI through sentiment analysis in media literacy programs represents a promising avenue for fostering a society more resilient to the pervasive challenge of fake news, by enhancing the analytical capabilities essential for discerning the reliability and bias of information sources.

Scheibenzuber et al. [55] discuss the development and testing of an undergraduate online problem-based course designed to combat fake news illiteracy, especially during the Covid-19 pandemic. This course, developed in response to the surge in misinformation ("infodemic") and the shift to emergency online learning, aimed to enhance students’ ability to critically assess the credibility of news. It incorporated insights from fake news research and the problem-based learning approach. The study, involving 102 students, found positive student feedback on online communication and feedback aspects of the course, though some suggested improvements for task descriptions. Significant improvements were observed in students’ ability to assess the credibility of fake news, with academic achievements reflecting a good to very good standard. The conclusion underscores the potential of problem-based online courses as effective environments for improving fake news literacy, even in emergency learning contexts.

Mi and Apuke [56] explore the role of social media knowledge in combating fake news among Nigerian social media users. Their structural equation modeling analysis reveals that altruism and self-promotion significantly influence fake news sharing behavior. Moreover, individuals with higher social media knowledge are less likely to engage in fake news sharing. The study concludes that enhancing social media knowledge can effectively reduce the spread of misinformation, contributing to both theoretical understanding and practical approaches to curbing fake news dissemination.

Jang and Kim [57] investigate the third-person effect in the context of fake news, finding that individuals believe fake news has a greater impact on others than on themselves or their in-group. The study identifies partisan identity, social undesirability, and external political efficacy as predictors of this perception. Interestingly, those with a stronger third-person perception favor media literacy interventions over media regulation to combat fake news. This research highlights the varied responses to fake news and underscores the importance of media literacy as a preferred solution.

Li et al. [58] examine the public’s engagement with fake news rebuttals on social media during the COVID-19 pandemic. Using the fake news rebuttals engagement index (FNREI) and time-varying parameter vector auto-regressive model (TVP-VAR), the study analyzes data from 19,603 Weibo posts. Findings reveal that public engagement with COVID-19 related fake news rebuttals varies over time and is influenced by major pandemic events. This research provides insights into how public engagement with fake news rebuttals evolves and offers suggestions for enhancing fake news governance effectiveness.

Beauvais [59] explores the reasons behind individuals’ belief in fake news, particularly during the COVID-19 pandemic and US presidential elections. The study identifies media and social network ecosystems, cognitive and psychological factors, and sociological influences as key determinants. It suggests that high digital literacy can protect against believing in fake news and recommends measures for journalism, education, fake news detection, fact-checking, and social network regulation to counteract misinformation. This study emphasizes the multifaceted approach needed to understand and combat the belief in fake news.

Apuke and Omar [1] examine the factors influencing the sharing of fake news about COVID-19 on social media, using a Nigerian sample. Altruism emerged as the most significant predictor, with motivations related to information sharing, socialization, and information seeking also contributing to the spread of misinformation. The study concludes with implications for understanding the motivations behind fake news sharing, suggesting targeted interventions to reduce its spread.

Yang and Tian [60] focused into the third-person perception (TPP) regarding COVID-19 fake news susceptibility on social media. They find that communal engagement increases TPP, mediated by perceived knowledge and self-efficacy, but not by negative emotion. The study suggests that while social media can propagate fake news, it also fosters a belief among users in their own immunity to misinformation, highlighting the complex role of social media in the dissemination of fake news.

Wei et al. [61] Investigate the moderating role of social media literacy skills in the relationship between rational choice factors and fake news sharing behavior among Nigerian social media users. Trust in social media and status.

The proliferation of informal education through social media presents both unparalleled opportunities and significant challenges. On one hand, it democratizes access to knowledge, allowing lifelong learners from all walks of life to continuously engage with new information. This accessibility is particularly beneficial in an era where continuous learning and adaptation to new knowledge are crucial. However, the openness of these platforms also paves the way for the widespread dissemination of misinformation or fake news. The risk is more pronounced among those who may not belong to the digital native generation, such as older adults, who might lack the digital media literacy skills necessary to differentiate between authentic and fraudulent content. The issue of fake news is not just about the spread of misinformation but also about the potential it has to influence public opinion and behavior based on false premises. This scenario highlights the critical need for enhanced media literacy that includes the ability to critically assess the credibility and authenticity of online content. The integration of Artificial Intelligence (AI) tools, particularly machine learning and deep learning algorithms, in media literacy efforts presents a promising solution. These technologies can analyze vast amounts of data, identify patterns, and utilize sentiment analysis to gauge the authenticity and bias of the content. By using AI, a user alert system can be created to report potential misinformation, thereby acting as a digital literacy tool that empowers users to navigate the vast seas of information with a more discerning eye. Thus, while social media serves as a valuable educational resource, ensuring the reliability of the content consumed requires a concerted effort to foster digital literacy, especially among those not naturally versed in digital technologies. The use of AI and sentiment analysis offers a bridge across this digital divide, equipping users with the means to critically engage with content and become more resistant to the pitfalls of fake news. This approach not only protects individuals but also preserves the integrity of the digital information ecosystem, making it a safer space for educational purposes and informed discussions.

2.4. Motivation and research contribution

This work is driven by the growing global use of the internet and social media for sharing info. These digital places help us connect with people all over the world fast, but it is hard to know if what we read is true. So many lies and fake news spread on them easily. The urgency to address this issue is highlighted by the widespread dissemination of COVID-19-related misinformation, posing significant risks to public health. The research also stresses how important it is to stop false information online so many people get the COVID-19 vaccine. In this situation, sentiment analysis is very important because it allows us to analyze the information in real time. This helps stop fake information from spreading especially when talking about health or public info.

In concluding our discussion on sentiment analysis methodologies, it is noteworthy that Deep Learning has emerged as a prominent choice, surpassing other approaches like Machine Learning and traditional Lexicon-based methods. This is attributed to the superior accuracy demonstrated by Deep Learning when applied to extensive datasets with diverse values, outperforming both ML and Lexicon-based techniques. The existing literature extensively explores sentiment analysis through various machine learning and deep learning algorithms. Each application yields distinct performance indicators; some studies report the use of logistic regression, while others employ Support Vector Machine or LSTM. Despite these findings, there is a clear need for more comprehensive research articles that apply a variety of learning methods to fully comprehend the comparative landscape between different algorithms, especially in the context of evolving deep learning techniques. In light of this, the contributions of our work can be summarized as follows: Deep Learning is recommended to replace lexicon-based methods and establish a platform for evaluating sentiment analysis algorithms in natural language processing (NLP) for textual feature extraction.

Consequently, the research contribution and the objective of this research is to employ five diverse machine learning classifiers, including Support Vector Machine, Logistic Regression, K-Nearest Neighbors, Decision Trees, and Random Forest, alongside three distinct deep learning techniques—CNN, LSTM, and GRU algorithms—to classify whether a Twitter (now X) post is authentic or not.

The article is organized into five sections. The upcoming section will focus into the methodology, explaining the details such as the dataset, pre-processing procedures, and the classification algorithms employed. Section 4 is dedicated to presenting the results and initiating discussions. This involves utilizing classification metrics and confusion matrices to provide a comprehensive analysis. Finally, the conclusion and recommendation section is included at the end of the article.

3.1. Dataset

This study uses the "COVID Fake News Dataset" from kaggle.com. [https://www.kaggle.com/datasets/csmalarkodi/covid-fake-news-dataset The dataset, collected from news articles by scraping various websites and manually annotated by experts, was subsequently published on Kaggle for public access and research use. The dataset used for this work was retrieved on 28-08-2023.There are 6420 tweets, 3360 of which are authentic and 3060 fraudulent. The dataset is publicly available. The collection has three columns: the unique identification (ID), the tweet content, and the label, which distinguishes authentic from fraudulent tweets, like in the example shown in Table 2 which is discussed briefly below:

• ID: This column contains unique identification numbers assigned to each entry in the dataset, allowing for easy referencing and identification of individual tweets.
• Tweet: This column comprises text data representing specific tweets related to the COVID-19 pandemic. Each row encapsulates the textual content of a particular tweet, ranging from updates, statistics, precautions, to opinions on the pandemic.
• Label: This column serves as a classifier, indicating the authenticity of the information contained in each tweet regarding COVID-19. Tweets are categorized as "real" when containing factual or authentic information and labeled as "fake" when potentially conveying misleading or false information pertaining to the pandemic.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Sample data having columns with the information like ID, Tweet, and Label.

Adapted from: [https://www.kaggle.com/datasets/csmalarkodi/covid-fake-news-dataset].

https://doi.org/10.1371/journal.pone.0315407.t002

In summary, Table 2 consists of columns representing unique identification numbers, the textual content of COVID-19-related tweets, and a classification label distinguishing between authentic ("real") and potentially misleading ("fake") information shared within the dataset.The dataset is labeled when its obtained from the recourse and no require any human annotation

3.2. Pre-processing

Once the data has been acquired, it is crucial to clean and preprocess it before proceeding with further processing. Utilizing thoroughly cleaned and preprocessed data is essential for obtaining accurate and favorable results [62]. In this specific procedure, various features such as punctuation marks, excessive spaces, and URLs are removed. Tokenization, the process of splitting a statement into smaller parts, either at the sentence or word level, is employed. Stop words, like "the," "for," "is," and "and" are excluded from the document as they do not contribute significantly to its meaning. The transformation of a term derived from its root word is known as "stemming." This involves converting terms like "needed" to "need," "eating" to "eat," and "cutting" to "cut." The training data has been condensed to a higher degree. Automated classifiers, which typically do not directly analyze text, require the conversion of text into a digital representation in the form of a vector. This digital vector is commonly referred to as a feature vector, and the conversion process is known as feature extraction. The TF-IDF (Term Frequency—Inverse Document Frequency) method [63] is widely used in the realm of feature extraction.

The pre-processed data is presented visually using both a histogram and a word cloud, as illustrated in Fig 2. In the context of sentiment analysis on Twitter (now X) data during the COVID-19 pandemic, two visualizations—a bar chart and a word cloud—offer valuable insights into the most frequently occurring words after pre-processing. The bar chart illustrates the frequency distribution of selected keywords. Notably, the term "COVID" dominates with a frequency of approximately 3500, followed by "cases" and "coronavirus," both hovering around 1800. Additionally, words such as "new," "people," "test," "deaths," "states," "total," "number," and "India" exhibit frequencies ranging from 500 to 1000. This quantitative representation underscores the prominence of specific terms in the dataset. Complementing the bar chart, the word cloud visually emphasizes the significance of certain terms. Words like "COVID," "new," "cases," "confirmed," "India," "coronavirus," "deaths," "vaccine," and "quarantine" stand out prominently. The varying font sizes in the word cloud directly correlate with the frequency of each term, providing an immediate visual cue to their importance. The discussion stemming from these visualizations could delve into the prevalent sentiments expressed on Twitter (now X) during the pandemic. For instance, the high frequency of terms like "cases" and "deaths" may indicate a heightened awareness and discussion around the impact of the virus. Additionally, the prominence of "vaccine" suggests a significant focus on vaccination efforts and related sentiments. These visualizations offer a quantitative and qualitative understanding of the sentiments encapsulated in the Twitter (now X) data, aiding in a more nuanced interpretation of public discourse during the COVID-19 pandemic.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2.

(a) Most frequently occurring words, and (b) word cloud, at post-processing.

https://doi.org/10.1371/journal.pone.0315407.g002

3.3. Data vectorization

The data vectorization step involves converting text which is input data into numerical vectors as most machine learning algorithms cannot process text directly. The transformation is achieved using the bag-of-words technique, treating text as a collection of words without considering grammar or order. TF-IDF is then used to assign scores to words, measuring their importance in documents. This process creates digital representations essential for algorithms to analyze textual data.

3.4. Classification

The dataset has two parts: training and testing. Training takes up 70% of the dataset, while testing and assessment take up 30%. The models used in this study undergo training with a specific dataset, constituting 70% of the total dataset. Testing is then conducted using a subset of data, specifically 30% of the entire dataset. Model evaluation relies on performance metrics, including the confusion matrix and classification rate. The analysis outcomes predict and classify the results as either "Real" or "Fake."

The following is a list of machine learning and deep learning algorithms, accompanied by an explanation of their implementation procedures.

4.1. Comparison with previous studies

In the rapidly evolving landscape of the COVID-19 pandemic, the detection and mitigation of misinformation have This comparison aims to highlight the strengths of Our proposed study in achieving superior results in the classification of COVID-19-related news compared to [70, 71] leverages a dataset amalgamated from various social media and news sources, focusing on the preprocessing steps, tokenization, and feature selection to enhance the efficiency of the classification process. The utilization of several state-of-the-art machine learning algorithms demonstrates a meticulous approach to model selection. Notably, the random forest classifier attains an impressive accuracy of 88.50%. While employing a diverse set of classifiers, including both machine learning and deep learning models, the BiGRU deep learning classifier is identified as the most effective, achieving an accuracy of 0.91. The comparison between these studies reveals that random forest classifier outperforms the classifiers in in terms of accuracy. While these studies address the critical issue of misinformation related to COVID-19, the distinctive strength of studies lies in its meticulous feature selection process and the careful consideration of various machine learning algorithms. The 88.50% accuracy achieved by study [70] random forest classifier signifies a robust performance in distinguishing between authentic and misleading COVID-19-related news.

4.2. Limitation and challenges

Detecting COVID-19 misinformation is challenged by the complexities inherent in natural language processing, requiring navigation through intricate nuances, varied contextual elements, and subjective aspects for accurate identification. The temporal dimension is crucial; conducting analyses within specific lockdown phases risks producing irrelevant findings due to evolving public sentiments. Neglecting emoticons and hashtags poses a significant challenge, compromising sentiment interpretation and classifier efficacy, especially in contexts involving sarcasm or contradictory expressions. Further, limited reliance on lexicons restricts the depth needed to identify evolving narratives of COVID-19-related fake news, potentially hindering the understanding of nuanced misinformation during the pandemic. Addressing these challenges mandates a comprehensive strategy encompassing language intricacies, adapting to temporal changes, integrating emoticons and hashtags, and broadening lexicon use. To achieve heightened accuracy in detecting COVID-19-related fake news, an approach encompassing the multifaceted information landscape during this unprecedented global crisis is imperative.