Impact of Unreliable Content on Social Media Users during COVID-19 and Stance Detection System

: The abundant dissemination of misinformation regarding coronavirus disease 2019 (COVID-19) presents another unprecedented issue to the world, along with the health crisis. Online social network (OSN) platforms intensify this problem by allowing their users to easily distort and fabricate the information and disseminate it farther and rapidly. In this paper, we study the impact of misinformation associated with a religious inﬂection on the psychology and behavior of the OSN users. The article presents a detailed study to understand the reaction of social media users when exposed to unveriﬁed content related to the Islamic community during the COVID-19 lockdown period in India. The analysis was carried out on Twitter users where the data were collected using three scraping packages, Tweepy , Selenium, and Beautiful Soup , to cover more users affected by this misinformation. A labeled dataset is prepared where each tweet is assigned one of the four reaction polarities, namely, E (endorse), D (deny), Q (question) , and N (neutral). Analysis of collected data was carried out in ﬁve phases where we investigate the engagement of E, D, Q, and N users, tone of the tweets, and the consequence upon repeated exposure of such information. The evidence demonstrates that the circulation of such content during the pandemic and lockdown phase had made people more vulnerable in perceiving the unreliable tweets as fact. It was also observed that people absorbed the negativity of the online content, which induced a feeling of hatred, anger, distress, and fear among them. People with similar mindset form online groups and express their negative attitude to other groups based on their opinions, indicating the strong signals of social unrest and public tensions in society. The paper also presents a deep learning-based stance detection model as one of the automated mechanisms for tracking the news on Twitter as being potentially false. Stance classiﬁer aims to predict the attitude of a tweet towards a news headline and thereby assists in determining the veracity of news by monitoring the distribution of different reactions of the users towards it. The proposed model, employing deep learning ( convolutional neural network (CNN) ) and sentence embedding ( bidirectional encoder representations from transformers (BERT) ) techniques, outperforms the existing systems. The performance is evaluated on the benchmark SemEval stance dataset. Furthermore, a newly annotated dataset is prepared and released with this study to help the research of this domain.


Introduction
The year 2020 is marked by one of the major public health threats and emergency which the whole world witnessed due to outbreak of the coronavirus disease 2019 (COVID- 19) infection [1]. In the face of coronavirus pandemic, social lockdown is imposed across the globe where people are strictly asked to stay home for more than a month to slow down the spread of infection. The isolation aspects of the lockdown period have led to overwhelming growth in the user activities on social media platforms such as Twitter and Facebook, as they have allowed people to remain connected and updated about the current status of the disease [2]. However, heavy usage of these networking sites has also caused the massive circulation of COVID-19 related misinformation such as false cures, bogus medical advice, the origin of the virus, 5G conspiracy, etc. [3,4] at an alarming rate.
Misinformation can be of several types: (1) it can be completely inaccurate, (2) it can be a belief disseminated without any authentic proof, or (3) it can carry biased information where selective facts are shared to achieve some mischief propaganda [5][6][7]. The study [8] pointed out that unreliable content on social media spread way faster than fact-based content. It has been noticed that the truthfulness of content is not a motivational aspect for forwarding the information; rather, people tend to share the news based on their favoritism, biasedness, or gaining attention in the community. Earlier studies have observed that in real life, the dissemination of unreliable content on social media impacts significantly on political campaigning [9,10], terrorism [11], natural disaster management [12]. This article conducts the study to explore the influence of the propagation of religion-based toxic misinformation on the psychology and behavior of social media users in causing unrest and communal violence. It is seen that the rumors having some religious inflection are potentially strong in creating severe chaos and dividing society [13]. Social media like Twitter can ignite its consequences on the psychology of the society as these platforms are well known for rapidly communicating fabricated and distorted information, which usually carries the news based on the views and emotions of the users rather than the facts. This paper analyses the diffusion of religion-based unverified news on Twitter during the COVID-19 panic to investigate how it affected the mental state and behavior of the users in triggering communal unrest. A similar study is conducted by the authors [5] that conducts an offline survey in politically unstable and conflicting areas to show how easily a rumor can amplify the instability and violence in the society. In this paper, we conducted a digital survey to investigate the behavior of people regarding a religious-based event, i.e., Tablighi Jamaat (an Islamic sect) that took the turn of communal violence mainly because of the COVID-19 crisis and the spreading of unverified facts. COVID-19 pandemic has been observed as the source of the surge of various fake and unverified content, mostly regarding the cause of the outbreak of the disease, and ironically, people are highly noticed as believing such information [14]. This paper explores the reactions of Twitter users towards the Tablighi Jamaat (TJ) event that happened in mid-March 2020 in Delhi (the capital of India country) and became the most debatable topic if it is responsible for the outbreak of coronavirus infection in the country [15].
Around 8000 people, including foreigners, are claimed as the attendees of the TJ event. The event caught attention when later on, many attendees were tested as coronaviruspositive, and various non-attendees who picked up the COVID virus were somehow found in link with attendees. Many people associated this event with the whole Muslim community, calling it a conspiracy and blaming the community for the spread of the disease [15]. In a verdict, the Bombay High Court of India slammed the electronic media for being part of big political propaganda that attempts to create the image that TJ members are the cause of the outbreak of coronavirus infection in India [16]. However, although the claim that TJ (i.e., Tablighi Jamaat) event is the main cause for the spread of the virus has not been proven yet [17], it had sufficient potential to contribute to the instability and communal violence in the society.
In this paper, we observed the usage of the Twitter networking site by its users towards the mentioned event and see if it has contributed to the instigation of violence, hate, anger, and division among people. To obtain the TJ event-based tweets, data were collected in three iterations by using three packages, namely, Tweepy (https://www.tweepy.org/), Selenium (https://selenium-python.readthedocs.io/), and Beautiful Sou (https://pypi.org/ project/beautifulsoup4/). Afterward, an annotated dataset was created by determining the stance of the tweets with respect to the claim that the TJ event was responsible for the spread of coronavirus in the country. The stance represents the opinion or a stand of a person about something. For news, there could be four types of reaction/stance, namely, endorse (E), deny (D), question (Q), or neutral (N). Therefore, each tweet in the Electronics 2021, 10, 5 3 of 21 dataset was labeled into one of the four reaction polarities, E (tweet that supports the claim that the event was responsible for causing coronavirus outbreak in India), D (tweet that outrightly denies the claim), Q (tweet that questions for figures-based facts) and N (tweet that shares the facts regarding the claim). Afterward, a deep analysis was conducted on the corpus to gain a thorough understanding of the attitude of different user groups regarding the misinformation. The article also presents the hashtags-based clustering of tweets for detecting echo chambers in the network. Earlier studies have mentioned the phenomenon of the formation of echo chambers on social media where people tend to communicate mostly with similar-minded users [18]. The problem with these echo chambers is that they can enhance the trust factor of a user on an unauthenticated piece of information by getting repeated affirmations from people with a similar belief system and can make a discussion more radicalized and polarized.
Furthermore, the article also presents a stance detection system to combat the problem of fake news on the social network. Stance classification aims to determine whether the author of the text stands in favor or against a news title [7]. Stance detection has been considered a vital task for solving many problems, including fake news detection [19][20][21].
The study [22] shows that the tweets spreading rumors propagate differently than the news spreading truth as the former content is mostly questioned or denied. Therefore, monitoring the orientation of the users to the news is one of the crucial steps for determining the veracity of the news. At present, there are numerous fact-checking organizations where people are continuously investigating the veracity of news articles to make an informed judgment of their potential of being fake. A reasonable approach for checking the truthfulness of the claim is performed by retrieving the related content of the news under scrutiny from different sources such as social media, determine the stance of the retrieved documents, perform the stance analysis, and finally, make a prediction. However, the manual process is very time-consuming, complex, and does not scale well. Therefore, automatic detection of the stance of the posts/tweets shared with respect to the news would assist the fact checker in effectively evaluating the distribution of different opinions and flagging the news article as suspicious upon encountering a large number of countered reactions.
In this paper, we present a deep learning-based classifier that identifies the stance of the tweet of a claim. The work utilizes the dataset provided by Sem-Eval 2016 competition for the stance detection task [23]. Earlier works have majorly focused on word embedding techniques such as word2vec [24] for representing the text for the deep learning classifier [25]. On the contrary, the proposed model employs sentence embedding to convert the tweet text into a vector. Moreover, instead of designing the classifier from scratch, transfer learning has been adopted where we use the pre-trained bidirectional encoder representations from transformers (BERT) [26] to obtain the sentence embeddings vector. BERT is a powerful deep learning-based language representation model proposed by Google in late 2018 for natural language processing (NLP). On top of BERT-based features, four supervised learning algorithms, namely, support vector machine (SVM), random forest, neural network, and convolutional neural network (CNN), are employed for designing the respective four stance detectors bert-svm, bert-rf, bert-nn, and bert-cnn. The performances of the models are compared with the score of top-ranked teams that participated in the Sem-Eval 2016 competition. It is observed that the bert-cnn outperformed the baseline systems. Furthermore, we demonstrate that the annotated Tablighi Jamaat dataset prepared in this work has the potential to enhance the efficiency of stance classification, and thereby, can be utilized by the concerned researchers in their studies.
The main contributions of the paper are as follows: • A three-step data collection procedure is presented using three APIs, Tweepy, Selenium, and Beautiful Soup. Each API is associated with certain limitations in scrapping, which are discussed in the paper. The cooperative use of these APIs in data collection helped to overcome the limitations of one another, and therefore, assisted in extracting a wide range of TJ event-based tweets effectively; • Designed a stance-based annotated dataset where the reactions of users towards the unverified claim are classified into four classes, i.e., E, D, Q, and N. The dataset can be made available to the researchers for further study; • Polarity-based text filtering is implemented for extracting the desired text from the collected tweets; • There are a total of 5 phases of analysis that attempt to explore the tweeting pattern of E, D, Q, and N user classes, the kind of language used by them, and the change in their language when they are repeatedly exposed to the harmful unauthenticated content; • Hashtags-based clustering is performed to identify the echo chambers in the network; • A deep learning-based stance classifier has been proposed that would assist in tracking the misinformation on social media. The results show that the proposed model outperforms the existing systems; • Demonstrate the usefulness of the prepared annotated dataset in stance classification.
Furthermore, the key findings of the analysis are as follows: • Circulation of information initially started with neutral news containing facts and figures, which later on turned into incredible information due to the misinterpretation by the users; • The sharing of distorted and incomplete content regarding the TJ event led to the constant arguments between the two user groups (E and Q), reflecting the signs of social unrest. It is to be noted that both the groups are opposing each other on unverified information; • The language of E users was more of an attacking mode treating the TJ attendees as criminals and super-spreaders of infection, whereas D users were found using more defensive terms, mostly asking the people to stop spreading hate in the community; • Interestingly, the terms which were found common in the language of E and D users were primarily educative, providing the information of the event itself, such as its timing and location; • User language was getting more violent and communal over time. The evidence indicates that the repeated exposure of the rumor over a period can worsen the effect on the cognitive ability of the individuals by imbibing continuous distress and anger in them; • The analysis also revealed the presence of echo chambers where similar mindset users were observed forming clusters based on their language.
The paper is structured as follows: Section 2 provides the background of the TJ event. Section 3 highlights the related existing studies. The three-step data collection procedure is described in Section 4. Section 5 provides the details of designing the reaction-based labeled dataset. The five phases of data analysis are given in Section 6. Section 7 discusses the findings of the study. In Section 8, the methodology for designing the stance classifier has been presented. Section 9 describes the experiments and results of stance detection. Finally, Section 10 concludes the overall work around the spread of misinformation and stance detection system to combat the problem of fake news on social networks.

Background of Tablighi Jamaat (TJ) Event
Almost every country implemented the lockdown to control the COVID-19 pandemic. The Government of Delhi state of India also announced the ban on the gathering of more than 200 people on 13 March 2020, which was further strengthened on 16 March 2020 with the limit of 50 people. The Tablighi Jamaat (TJ) event was held on March 13-15 in Delhi (capital state of India). It was claimed that over 8000 people were present in the event, majorly belonging to the Islamic community, from several states of India along with over 1000 foreigners. The event came under focus after a week later when many attendees or non-attendees somehow linked with the TJ event were tested COVID-19 positive across the different states of the country. On 31 March, the Union health ministry of India claimed that 30% of COVID positive cases in Delhi were found in linking with the Tablighi Jamaat incident. On 2 April, another news surfaced which said that this incident caused a doubling of coronavirus-infected cases. This news took a significant turn on social media, where users were increasingly calling its members responsible for the spreading of coronavirus across India. However, later on, 13 April, a piece of news [17] published which reported that according to a group of Indian scientists, there is no evidential data that concludes that the Tablighi Jamaat convention is the primary reason for the continuous growth of the COVID-19-infected Indian citizens. Wikipedia (a great source of information) also witnessed the confusion regarding the claim when the article portraying Tablighi Jamaat as a hotspot of spreading coronavirus was deleted on account of lack of numeral census [27]. However, the page was restored after a couple of days with some modifications [28]. Finally, on 22 August, the Bombay High Court of India slams both media and electronic media for being part of political propaganda against Tablighi foreigners that aimed to manipulate the society into believing that they have spread the coronavirus in India and portray their picture as criminal [16].

Related Work
In this section, we briefly discuss the contribution of social media data analysis in solving various research problems and state-of-the-art stance detection.

Social Media Data Analysis
The extensive usage of social media platforms has produced a massive amount of data that contains information about different dimensions of people's lives, be it professional, personal, and social. This enormous data offers the researchers a valuable resource for understanding the various aspects of human behavior such as mental state [29], insecurities [30], opinions or sentiments [31], the kind of response to be expected in the wake of panic events [32], criminal psychology [33], etc. In the study [30], an investigation is carried out on Instagram media to observe the reaction of men towards body satisfaction and physical appearance comparison. In [34], the authors explored the posting behavior of Twitter users to understand the factors that motivate a user to publish content on social media. The study [29] show that the people living in a politically non-stable environment are emotionally degraded and represent a different mental state than the people living in stable areas in terms of expressing more fear, anxiety, and negativity in their Facebook posts. Since understanding and modeling human response behavior can assist the administration in dealing with adverse conditions such as natural disasters, in the study [32], the authors analyzed the tweets during the 2010-2011 floods crisis in Australia to identify the active user communities and their role in disseminating critical information. Social media data can also be utilized by researchers to study the psychology of criminals and curb cybercrimes. For instance, the study [33] explored the emotions expressed in the content published by real and fake profiles to design the fake profile detection system.

Social Media and Misinformation Analysis
In this paper, we explore the reaction of Twitter users towards the circulation of unverified information that can amplify communal violence in society. Some existing research studies tend to investigate the impact of the spreading of misinformation on the people. For example, the study [35] shows that once a person has drawn the wrong image due to the consumption of false information, its effect cannot be easily undone even after presenting the facts. The results point out that initial social impressions have a long-lasting effect even if these impressions are based on some false news. The study also shows that individuals with relatively lower cognitive skills are more prone to the effects of false content. The study [36] discussed a case study where a false rumor regarding the kidnapping of children from a school was disseminated like a wildfire on Twitter. In addition, it is observed that the news created massive chaos and traffic jams in a couple Electronics 2021, 10, 5 6 of 21 of minutes out of panic as the parents rushed to collect their children from the school. The proliferation of fake news has been seen as strongly correlated with the occurrence of real-time events such as natural disasters, election campaigning, etc. For example, the paper [11] analyzed the quality of the tweets shared during the social crisis and revealed various factors like anxiety and source ambiguity, which can lead to the massive overload of rumors on social media during the period. The paper [37] demonstrates that user interaction with unreliable content on social media increases sharply during the election. The spike in rumors has also been related to the areas which are considered to be conflicted and politically unstable. The authors [5] surveyed two insurgency-affected areas and claimed that distrust and threat perceptions in these areas increase the odds of unverified information being perceived as true among people. The studies [9,10] show how the spreader of fake news on social media influenced the outcome of the US 2016 presidential election. In this paper, we seek to determine the impact of religion-based misinformation on the behavior of twitters during the COVID-19 pandemic.

Stance Detection
Stance detection has many application areas in the domain of NLP [25]. For instance, automatic stance detection can replace the traditional opinion surveys that are carried out to determine if a community stands in favor or against an ideology [38]. Stance detection is also proven useful in designing recommendation and market forecasting system [25]. Most importantly, it has gained significant importance in the field of rumor and fake news detection. The authors have shown the stance detector as the core component for monitoring the rumors on social media [19]. Similarly, fake news challenge 2017 [20] has also included stance detection as a vital subtask in determining whether a given piece of the news article is fake or real. Existing studies have employed feature-based machine learning (e.g., SVM [39], logistic regression [39], etc.), and deep learning techniques (e.g., Recurrent neural network [40,41], convolutional neural network [42], etc.) for predicting the stance of a text. The machine learning models are mostly trained on features based on bag-of-words, Tf-IDF, POS tags, sentiments, etc. There are a few studies that utilized word embedding vectors also with the help of pre-trained word2vec and GloVe models [43]. However, deep learning techniques have majorly relied on word embedding representations of a text for designing the stance detector. In this study, we employ sentence embedding techniques for text representation by utilizing the pre-trained BERT model. Furthermore, both machine and deep learning techniques have been experimented with for designing the stance classifier.

Data Collection
The study focuses on the examination of the attitude of Twitter users towards the unverified claim that held the TJ event responsible for the outbreak of COVID-19 in India and determining whether the speculation of such misinformed content catalyzed the communal violence. The data collection is directed towards obtaining only event-based tweets shared by users. The whole data were collected in three iterations using the three Python libraries, namely, Tweepy, Selenium, and BeautifulSoup, to gather more tweets and cover more users affected by the claim. Figure 1 provides the pseudocode of the three iterations for data collection. The input to the algorithm is the list of query strings such as "jamaat covid case increase" and "jamaat responsible corona," which were intelligently designed so as to locate the concerned tweets on the network. The search functionality of the Tweepy library has been utilized to crawl the TJ event-based tweets via query strings. Four attributes of the tweet, namely, tweetscreen_name, tweet id, tweet text, and tweet date, were scrapped and saved in the file, tweetFile. This is the first iteration of data collection. the next task is to visit the Twitter profile of the retweeter and search for the particular tweet, which is the retweet of the original tweet. Tweepy library is utilized to execute this functionality. Since the retweets we are looking for on the retweeter profile contains the text of the original tweet, the value of tweet text saved in the retweetersFile is used to identify them. Once the retweet is found on the profile, all the four attributes, tweet-screen_name, tweet id, tweet text, and tweet date, are scrapped and stored in the tweet file. Figure 1. Pseudocode for three-step data collection.
As already discussed, while retweeting, there are two possibilities, (1) the users may share it as it is (simple-retweet), or (2) the users may add some of their comments to the original content (quoted-retweet). Therefore, a delimiter "$$$$$$##$$" is used in storing the retweeted text. Text before the delimiter represents the content of the tweet that was retweeted, and text after the delimiter contains the added comment if any.
In order to further extend the coverage of data collection, in the third iteration, we extracted the replied-tweets of the tweets saved in tweetFile. The BeautifulSoup library was used for collecting these tweets due to the following reasons: The second iteration aims to scrap the re-posting (or retweet) details of the tweets collected in the first iteration in order to gather more reaction towards the rumor. A tweet can be shared as it is, called simple-retweet, or by adding comments, referred to as quotedretweet. Since the Tweepy library does not give the details of quoted-retweet, the Selenium package of Python has been used to scrap the retweets record as it has the capability of imitating human action and extract the content from the browser itself. In this iteration, the Selenium script is designed to obtain the list of retweeters who further shared the tweets saved in tweetFile. The task of the Selenium script is to navigate to the URL of each original tweet (https://twitter.com/tweetscreen_name/status/tweetid), open the dialog box which displays the screen names of the retweeters, and scraps their IDs. The scrapped re-tweeters IDs, along with the text content of the original tweet, i.e., tweet text, are saved in the retweetersFile. It is to be noted that the dialog box did not show the IDs of all the retweeters, and so, Selenium could not provide the record of all retweeters. Afterward, the next task is to visit the Twitter profile of the retweeter and search for the particular tweet, which is the retweet of the original tweet. Tweepy library is utilized to execute this functionality. Since the retweets we are looking for on the retweeter profile contains the text of the original tweet, the value of tweet text saved in the retweetersFile is used to identify them. Once the retweet is found on the profile, all the four attributes, tweetscreen_name, tweet id, tweet text, and tweet date, are scrapped and stored in the tweet file.
As already discussed, while retweeting, there are two possibilities, (1) the users may share it as it is (simple-retweet), or (2) the users may add some of their comments to the original content (quoted-retweet). Therefore, a delimiter "$$$$$$##$$" is used in storing the retweeted text. Text before the delimiter represents the content of the tweet that was retweeted, and text after the delimiter contains the added comment if any.
In order to further extend the coverage of data collection, in the third iteration, we extracted the replied-tweets of the tweets saved in tweetFile. The BeautifulSoup library was used for collecting these tweets due to the following reasons: • Tweepy library does not provide the functionality to extract replied-tweets unless we have a premium account • Designing a beautiful soup-based script is way easier than the Selenium script • Selenium increases the chances of suspension or blocking of the Twitter account being used for collection.
The text of replied-tweet is stored similarly as retweeted-tweet, i.e., separating the content via "$$$$$$##$$" delimiter. Here, the former part of the delimiter represents the text of the tweet to which the user is responding, and the latter part holds the responded text.
In the end, around 6000 tweets were extracted associated with the claim. However, we also observed various factors that limited the data collection process. These are discussed as follows: (1) Tweepy does not provide data older than 2 weeks, and the data collection was started in the first week of April, (2) The dialog box of retweeters was not giving more than 80 users ids even if a tweet has 10k retweets, (3) It was not possible to locate the concerned tweets on the walls of every retweeter since the Tweepy can extract a maximum of 40,000 tweets of a user and if the concerned tweet is older than that, we could not reach to it, and (4), Beautiful soup cannot extract dynamically loaded content, and so, can return only those replies which appear on the source page of the tweet URL.

Data PreProcessing
The preprocessing is conducted to prepare the tweets' data for analysis. It is implemented in two stages, Tweets Labeling and Polarity-based Text Extraction, which are discussed in the following subsections.

Tweets Labeling
In this stage, the task is to prepare a labeled dataset by classifying the tweets into one of the following polarities, i.e., E endorse), D (deny), Q (question), and N (neutral) to categorize the different kinds of the reaction of the user towards the claim. E tag denotes the tweets which supported the claim and agreed that the event is the cause of the coronavirus outbreak in the country. Similarly, D and Q tags denote those tweets who either disagree outright or raise the doubt regarding the claim, respectively. In the N tag, we mark those tweets that just share the facts and figures related to the information. Table 1 shows some tweets samples that were categorized under these labels. Since the collected dataset also includes the tweets which were retweeted as it is, there is a high possibility of having duplicated tweet text in the corpus. Therefore, before classifying the tweets into four classes, we filter the unique tweets. Furthermore, sarcasm is found in the tweets; for example, "TJ members are way too good", "Tablighi jamaat people are very funny", etc. This made it difficult to label the tweets with a program. Therefore, the labeling of the selected tweets was done manually with the mutual consent of the authors of the paper. Once the tweets with unique text were assigned classes, the rest of the tweets were labeled by the program.

Polarity-Based Text Extraction
As we discussed in the section, the text of retweeted and replied-tweets contains two parts of information, i.e., the text of the original tweet and appended/responded Electronics 2021, 10, 5 9 of 21 content, which is separated by a delimiter. It is to be noted that when a user retweets the content as it is, it shows user approval regarding the text. On the contrary, when the user adds his/her views while retweeting, there are possibilities that the user may support, question, or deny the text of the tweet to be retweeted. Similarly, the replied-tweets or comments can also show either approval or disapproval of the original tweet. Since one of the objectives is to analyze the language used in the four classes of tweets (E, D, Q, and N), text filtering is done for extracting the part of the information that projects the intention of the replied/retweeted tweet.
The task of the text extraction is to first identify whether the appended/responded content shows approval or disapproval of the original tweet. In order to detect the agreement or disagreement, the polarities of the original tweets and retweets/replied-tweets were compared if they are the same or change. The same polarity indicates approval, and therefore, in this case, both the parts of the information, original and added/responded, were considered as the views of the user. On the other hand, if the change in polarity is detected, only the added/replied text was considered for language analysis. Table 2 depicts two examples to demonstrate the text filtration in two scenarios, disapproval and approval. Case 1 shows the text extraction when a disagreement is detected between the reaction polarities (P) of replied or retweet (RT) and original tweet (OT). Let "The TJ is responsible" be the original tweet (OT), and the user further shares this tweet by adding the content "You are lying #islamophobia" to the content of OT. Since the polarities of OT and RT are different, while analyzing the language of the RT, only the appended part will be considered. Case 2 explains the scenario when RT content shows approval for the OT content. In such a condition, we kept both the part for language analysis of RT.

Data Analysis
The data were analyzed in five phases to understand the behavior and reaction of users from various aspects. The following five subsections discuss the findings of these phases, respectively.

First Phase
To study the behavior of Twitter users towards the claim, we first look at the tweeting patterns of the four classes of users, i.e., E, D, Q, N. We evaluate the frequency of the tweets for E-D-Q-N users on each day, which is shown in Figure 2. It can be seen that the timeline of tweets collected by the program is from 31 March 2020 to 2 May 2020. The initial thing we noticed, on the first day of the timeline (i.e., 31 March 2020), the number of tweets sharing neutral information (#41) about the event are exceedingly greater than the tweets endorsing (#11) or denying the claim (#2). On this date, news citing "Union Health Ministry of India" as reference was published, which reported 175 fresh COVID infected cases in Delhi, doubting Tablighi gathering as the prime reason [44]. It can be seen that although the news shares some credible facts, the claim towards TJ is still unverified. Following the news, a peak can be seen on 1 April for endorsed (E) users with #340 number of tweets. A trend was noticed on Twitter with the headline "Tablighi Jamaat is now responsible for the largest single-day increase in Indian COVID-19 cases". On the same date, the number of tweets (#115) from the users denying the information also increases relatively, which were citing other events and gatherings which happened in India during the COVID pandemic to be the potential reasons.
Electronics 2021, 10, x FOR PEER REVIEW 10 of 21 Health Ministry of India" as reference was published, which reported 175 fresh COVID infected cases in Delhi, doubting Tablighi gathering as the prime reason [44]. It can be seen that although the news shares some credible facts, the claim towards TJ is still unverified. Following the news, a peak can be seen on 1 April for endorsed (E) users with #340 number of tweets. A trend was noticed on Twitter with the headline "Tablighi Jamaat is now responsible for the largest single-day increase in Indian COVID-19 cases". On the same date, the number of tweets (#115) from the users denying the information also increases relatively, which were citing other events and gatherings which happened in India during the COVID pandemic to be the potential reasons. Next wave of tweets is seen on 5 April with an increase in the number of tweets from neutral (#90) and questioning (#144) users. On this date, another news was surfaced from the media that cited the statement of the Joint Secretary of India claiming that COVID-19 cases in India have doubled in 4.1 days, which would be 7.4 days without TJ event [45]. This news led to the flooding of the tweets that were raising the distrust regarding the claim by mostly asking the statistics of relatively increment in testing counts or arrival of the total number of tourists during that period. The main arguments that were put by Q users are that the doubling of figures could be due to the increase in the number of testing or the non-TJ tourists who have not been monitored.
11 April marks the highest peak for all types of user tweets, E-D-Q-N. This was the period when the news was published, claiming that according to a group of Indian scientists, the available data does not support the blame that the TJ event is mainly responsible for the continuous growth of COVID cases [17,46]. They also mentioned that the skewed counts of COVID cases linked with the TJ event are due to the low number of testing being performed across India.
Surprisingly, although the end of the timeline of the tweets collected by our program is 2 May, it can be seen that after 26 April, no neutral tweets were observed. It indicates that users at a certain point stopped sharing credible facts, and only the tweets containing stories based on the views and assumptions of the users were doing rounds on social media. Next wave of tweets is seen on 5 April with an increase in the number of tweets from neutral (#90) and questioning (#144) users. On this date, another news was surfaced from the media that cited the statement of the Joint Secretary of India claiming that COVID-19 cases in India have doubled in 4.1 days, which would be 7.4 days without TJ event [45]. This news led to the flooding of the tweets that were raising the distrust regarding the claim by mostly asking the statistics of relatively increment in testing counts or arrival of the total number of tourists during that period. The main arguments that were put by Q users are that the doubling of figures could be due to the increase in the number of testing or the non-TJ tourists who have not been monitored.
11 April marks the highest peak for all types of user tweets, E-D-Q-N. This was the period when the news was published, claiming that according to a group of Indian scientists, the available data does not support the blame that the TJ event is mainly responsible for the continuous growth of COVID cases [17,46]. They also mentioned that the skewed counts of COVID cases linked with the TJ event are due to the low number of testing being performed across India.
Surprisingly, although the end of the timeline of the tweets collected by our program is 2 May, it can be seen that after 26 April, no neutral tweets were observed. It indicates that users at a certain point stopped sharing credible facts, and only the tweets containing stories based on the views and assumptions of the users were doing rounds on social media.

Second Phase
In the first phase, we observed the tweeting pattern of E-D-Q-N users in association with the real-time event. Here, we seek to determine if there exists any correlation between the tweeting pattern of different classes of users. In order to obtain the pattern, tweets of four classes of users are first grouped by their date such that each date holds the number of tweets shared by E, D, Q, and N users, respectively. The constructed table has five columns, namely, date, #tweets by E, #tweets by D, #tweets by Q, and #tweets by N. Spearman rank correlation [47] was used to determine the association between the tweeting patterns of the user classes as the data do not follow a gaussian distribution. Table 3 displays the values of the correlation between the four classes. It can be seen that the correlation value obtained by endorsing and deny users is significantly higher than the rest of the classes indicating that as the number of tweets from the E group rises, the D group also becomes active and post counter-tweets. These are the only two groups that stand extremely opposite to each other. The second and third highest correlation values are seen between E-N and D-N user groups, respectively, which further indicates that the most conflicting groups, i.e., E and D tend to follow each other in terms of the number of tweets. These findings highlight the environment of social unrest created among Twitter regarding the claim. Table 3. Correlation between the tweeting pattern of four kinds of reaction towards unverified news.

Third Phase
As the second phase analysis indicated the sign of social conflict between the endorsing (E) and denying (D) users, in this phase, we are interested in exploring the language used by these two communities. To serve the purpose, we extracted the list of popular hashtags in the tweet content shared by them. The re library of Python was used to fetch the hashtags in the text. After extracting the hashtags of E and D users, we analyze their usage in three aspects, namely, (1) hashtags (h) used by both E and D (i.e., h(E)∩h(D)), (2) hashtags used exclusively by E (i.e., h(E)-h(D)), and, (3) hashtags used exclusively by D (i.e., h(D)-h(E)). Table 4

shows the example of hashtags found in three lists, h(E)∩h(D), h(E)-h(D), and h(D)-h(E)
. It can be seen that the topmost hashtags which were used commonly by both groups, i.e., h(E)∩h(D), include terms that mainly provide the information of the event that took place. For example, #covid19 and #coronavirus indicate the period of the event, #tablighijamaat mentions the name of the event, and #nizamuddinmarkaz tells the location of the event. These sets of tags do not include any racial, hatred, or communal term, and so can be considered as neutral and informative. If we look at the tags used by E users only, i.e., h(E)-h(D), a variety of tags from no violence (neutral) to violence (hatred, communal, and personal-attack) can be found. Neutral tags, as usual, include those words which provide the information about the event. We found various novel neutral terms in this list such as #Indiafightscorona (specifying the country, India where the event took place), #lockdown (specifying that event occurred in the lockdown period), and #maulanasaad (indicating the name of the chief of the event). The term # thoorahulkanwalthoo is considered as personal-attack hashtags since this tag is directed against a well-known personality of India who denied the claim that the TJ event is responsible for the COVID outbreak. The tag has three parts, thoo (English meaning is spit), rahulkanwal (a renowned Indian TV journalist), and again thoo, i.e., spit.
The list of hatred and communal hashtags comprises those terms that attempt to spread hate and encourage violence among the people based on their religion. For example, #Nizamuddinidiots consists of two parts, Nizamuddin (location of the event) and idiots (refers to the people who gathered in the event as idiots). The hashtag #tabligijamaatvirus treats the COVID-19 as a Tablighi Jamaat scourge. The list also includes severely violent terms such as terrorist in #tablighijamaatterrorist and jihad in #coronajihad. We also observed the hashtags which directly express the anger towards the Islamic community, such as #covid786 (786 is the holy number in Islam).
Finally, we explore the hashtags exclusively used by the denying user group, i.e., h(D)-h(E). Similar to the previous two hashtags lists, h(D)-h(E) also contains neutral terms such as #covid_19 and #coronavirus. However, it is surprising to see that the list does not include any communal or hate tags but fear. For example, the term islamophobia in the two hashtags, #stopcovidislamophobia and #islamophobia_in_india, is indicating the atmosphere of fear created among the people due to the diffusion of communal and hatred tweets against the Islamic religion. Moreover, #stopcovidislamophobia urges the users to stop spread the tweets that relate coronavirus with Islam, and #hatecoronanotcommunity appeals to the users to hate coronavirus, but not the Islamic community.

Fourth Phase
In the previous analysis, we explored the kind of language used by endorsing and denying communities in terms of the hashtags being used in their tweets. We observed various categories of hashtags in the content, from no violence (neutral) to severe violence (hate and communal). In this stage, we are keen to conduct the timeline analysis of these hashtags to understand what is the impact on the language of the tweeter users over a period of time due to the repeated circulation of the unverified claim that has the potential to create communal violence.
Here, we first determined the date of hashtags when they were used for the first time by the twitters. Figure 3 shows the plot between the hashtags and the dates they were used at first. It can be seen that at the beginning of the timeline (31 March 2020), neutral hashtags appeared, such as #covid19, #tablighijamaat, etc. The date 1 April 2020 also witnessed informative tags such as #indiafightscorona and #coronavirus. After this date, the data started showing hate or communal terms like #coronajihad (3 April), #nizamuddinidiots (4 April). Later on the timeline, personal-attack tags such as #thoorahulkanwalthoo is also noticed (9 April). The rear end of the timeline experiences more violent tags such as #tablighijamaatterrorist (11 April) and distress tags such as #islamophobia_in_india (21 April treat the information as true and make the language of the twitters more harmful, conflicting, violent, and terrible over time.

Fifth Phase
Hence, far, we have noticed two conflicting user communities, E and D, which appeared to be continuously fighting and arguing on Twitter over an unverified claim. We also observed the kind of hashtags used by them and how the intensity of negativity of these tags grow over time. All these results point out that the unverified claim had sufficient potential to polarize society and cause the situation of communal violence and social unrest. In this stage, we are interested to know if the four classes of users are making separate groups or echo chambers in the network. Their presence in our dataset will further provide the existence of polarization, groupism, and division in society.
In order to detect the echo chambers, we perform clustering on the tweets of the users to investigate whether the tweets with the same reaction polarity are using similar language. Each tweet is mathematically represented in vector form by employing the bag-ofhashtags (BoH) technique [48], similar to the traditional bag-of-words (BoW) text representation. In BoH vectorization, at first, a dictionary of hashtags (DH) is designed by extracting all of the hashtags that appeared in the tweet content of the dataset. The size of the dictionary is 175. Afterward, each tweet is assigned a vector, say, t(BoH) = (h1,h2,… hm), where m denotes the size of the vector. Since each hashtag in the dictionary, DH represents a feature, the length of the BoH vector of a tweet is the same as the dictionary, i.e., m = 175. The value of hi in the vector represents the number of times the respective hashtag occurred in the tweet.
K-means clustering was applied to group the BoH vectors of tweets. The similarity between the vectors is determined by evaluating the Euclidean distance. The K-means algorithm needs the value of the number of clusters in advance, which is calculated using the elbow method. The method gives nine as the optimal number of clusters. Table 5 shows the membership of E, D, Q, and N users in each cluster. It can be seen that out of 9 clusters, 2 clusters (C2 and C3) have the majority of E users, two clusters (C4 and C6) have the majority of D users, one cluster (C8) comprises Q users more in number, and one cluster (C9) has N users more in number. Only 3 clusters (C1, C5, and C7) were observed with a mixed population. These results show that the language used by the E, D, Q, and N is similar within themselves and differ reasonably greater from each other, thereby forming separate groups while clustering. Therefore, the findings reveal the presence of echo chambers as people with similar mindsets or reactions are forming groups based on their The timeline analysis of the birth of the hashtags indicates that repeated exposure to the claim that TJ causes coronavirus outbreak in the country has severely affected the psychology of twitters. According to a study [5], the dissemination of a rumor, time and again, can affect the cognitive and perceiving skills of the people into believing the information as credible. Therefore, continuous circulation of unverified claims persuaded the user to treat the information as true and make the language of the twitters more harmful, conflicting, violent, and terrible over time.

Fifth Phase
Hence, far, we have noticed two conflicting user communities, E and D, which appeared to be continuously fighting and arguing on Twitter over an unverified claim. We also observed the kind of hashtags used by them and how the intensity of negativity of these tags grow over time. All these results point out that the unverified claim had sufficient potential to polarize society and cause the situation of communal violence and social unrest. In this stage, we are interested to know if the four classes of users are making separate groups or echo chambers in the network. Their presence in our dataset will further provide the existence of polarization, groupism, and division in society.
In order to detect the echo chambers, we perform clustering on the tweets of the users to investigate whether the tweets with the same reaction polarity are using similar language. Each tweet is mathematically represented in vector form by employing the bag-of-hashtags (BoH) technique [48], similar to the traditional bag-of-words (BoW) text representation. In BoH vectorization, at first, a dictionary of hashtags (DH) is designed by extracting all of the hashtags that appeared in the tweet content of the dataset. The size of the dictionary is 175. Afterward, each tweet is assigned a vector, say, t(BoH) = (h1,h2, . . . hm), where m denotes the size of the vector. Since each hashtag in the dictionary, DH represents a feature, the length of the BoH vector of a tweet is the same as the dictionary, i.e., m = 175. The value of hi in the vector represents the number of times the respective hashtag occurred in the tweet. K-means clustering was applied to group the BoH vectors of tweets. The similarity between the vectors is determined by evaluating the Euclidean distance. The K-means algorithm needs the value of the number of clusters in advance, which is calculated using the elbow method. The method gives nine as the optimal number of clusters. Table 5 shows the membership of E, D, Q, and N users in each cluster. It can be seen that out of 9 clusters, 2 clusters (C2 and C3) have the majority of E users, two clusters (C4 and C6) have the majority of D users, one cluster (C8) comprises Q users more in number, and one cluster (C9) has N users more in number. Only 3 clusters (C1, C5, and C7) were observed with a mixed population. These results show that the language used by the E, D, Q, and N is similar within themselves and differ reasonably greater from each other, thereby forming separate groups while clustering. Therefore, the findings reveal the presence of echo chambers as people with similar mindsets or reactions are forming groups based on their language usage.

Discussion
The main objective of the analysis carried out in this study is to determine whether the misinformation disseminated on Twitter during COVID-19 has the potential to catalyze the instability and communal violence in society. The paper analyzed the tweets generated related to a specific religion-based rumor that claims the members of the TJ event are the spreaders of coronavirus infection in the country. The analysis was carried out in 5 phases. In the first phase, we look at the tweeting pattern of E, D, Q, and N users where a form of synchronization was observed between the number of tweets of the user-classes and the news that were published in real life regarding the claim. For example, we observed that the timeline started with neutral tweets in correspondence to the news, which reported the total count of COVID positive cases found as linked with the TJ event. Afterward, we see several small peaks of the users who were either supporting (E), refuting (D), or interrogating (Q) the claim. The highest peak was recorded on 11 April for all user groups when for the first time, an official statement was released, which refuted the allegation that the TJ event was responsible for the outbreak. The findings also highlighted the phenomenon when at a certain point in time, people started getting influenced more towards the fabricated version of the information that mostly contains the views of a similar mindset community rather than credible facts.
In the second stage of analysis, we determine whether there exists any correlation between the tweeting pattern of E, D, Q, N user classes. E and D classes were observed as the most strongly correlated groups in terms of the number of tweets shared by their users. It points to the presence of social unrest among Twitter users since these two are the two groups that stand on opposite sides of the claim, one endorsing the claim and another denying the claim.
As the third step of the analysis, we studied the language used by E and D user classes in terms of hashtags that appeared in their content. We analyzed three aspects, hashtags used by both E and D, hashtags used by E only, and hashtags used by D only. Here, we noticed a very interesting trend. The hashtags that appear in both classes were primarily neutral and educational, carrying the information of the event itself in terms of time, location, type, etc. However, the hashtags that were exclusively used by E class were mostly related to communal and hated showing their anger towards the participants who attended the event, and in severe cases, relating the event with the whole Muslim community. Surprisingly, the hashtags observed in the content of D class were mostly containing fear elements. The analysis of the language of E and D users further showed the sign of communal violence on Twitter as well.
In the fourth step, we observed the timeline of the birth of hashtags of E and D user classes in order to get insight into the impact on their language with time. We noticed that over time, the language was getting more violent and communal. For example, the timeline starts with the neutral and informative tags (e.g., #covid19, #tablighijamaat), get violent in middle (e.g., #tablighijamaatterrorist) and ended with extreme distress (e.g., #islam-ophobia_in_india). The findings reveal the effect of prolonged exposure to the toxic rumor on the psychology of the users. In the last section of the analysis, we performed clustering based on the language of all four user classes in terms of hashtags. Hashtag tweet vector was designed by applying the bag-of-words approach. In total, 9 clusters were formed where we noticed that E, D, Q, N do form separate groups and share similar language skills within themselves. The clusters also indicated the presence of echo chambers where users with similar mindsets seem to form groups.
This paper referred to several news articles as support. Hence, it is necessary to look at the credibility of news-sites whose articles have been cited. We used the media bias/fact check (MBFC) website to determine the factual accuracy of referred news media. The website scores the news media on six categories, i.e., very high, high, mostly factual, mixed, low, and very low, based on their credibility of publishing stories. The very high category includes those news-sites which get a score of 0, implying these sites are always observed as sharing credible news and never failed a fact check. Similarly, very-low-labeled news-sites are scored as 10, implying these sources can never be considered trustworthy. The news sources referred to in this article received three types of ratings, namely, high (1-3 score), mostly factual (3-5 score), and mixed (5-6 score). Those websites are categorized as high and mostly factual, which have failed just one or two fact checks, respectively. Mixed-labeled news sources are those websites which have not always used authentic source and often fail in fast checks. Table 6 shows the scores of the news media assigned by the media bias/fact check.

Stance Detection
The paper emphasis the analysis of the behavior of Twitter users towards religionbased misinformation disseminated during the COVID-19 pandemic. The main objective of the study is to determine whether Twitter plays a role in instigating communal violence and social instability due to spreading.
In this section, we present a stance detection model to identify the opinion of a tweet towards a specific news headline. Using this model, we can get the distribution of different reactions from the people. The proposed model will help fact-checking organizations with automatic monitoring of reactions from the people in estimating the reputation of the news. The main task of the stance detector is to determine the attitude (favor or against) expressed in the text of the author towards a piece of news. In our context, we address the stance detection problem as a multiclass classification where the input to the model is a pair of text fragments (tweet and claim), and the job of the model is to predict whether the tweet favors, denies, or is unrelated with the claim.
The model is initially trained and tested on the benchmark dataset provided by SemEval-2016 [23]. The training and testing data were provided separately by organizers, consisting of around 3000 and 1500 tweets, respectively. The tweets were covering a total of five domains as targets, namely, "atheism", "climate change is a real concern", "feminist movement", "Hilary Clinton," and "legalization of abortion". Each tweet is labeled into one of the three following classes: Favor (if the tweet expresses the opinion in favor of the target), Against (if the tweet is against the target), or None (if the tweet belongs to a different news domain). The text representation (input features) and the architecture of the proposed stance detection system are described in the following subsections:

Input Features
Our stance detector model takes two parameters, tweet, and target as input. Since both the inputs are in text form, they need to be converted into mathematical representation to make the model learn on the data. One hot encoding and sentence embedding were employed as text representation techniques.
Since the domain of the target has a limited vocabulary of words, it is represented with a one-hot vector [49]. However, before applying the one-hot encoding, we need a dictionary that is designed by aggregating the content of the five targets, converting it into lower case, tokenizing it, and removing punctuations. In the end, the vocabulary contains a total of 11 words, atheism, climate, change, real, concern, feminist, movement, hillary, clinton, legalization, and abortion. Once the vocabulary is designed, the target of each sample of the dataset is assigned one hot vector, which is a sparse binary vector of length 11. In the vector, all the indices which correspond to the words present in the target are marked 1, and the remaining indices are assigned a 0 value.
Since the dictionary of tweets would be exceedingly large, therefore, the sentence embedding technique was considered a good choice for the representation task. Here, the entire sentence is converted into a fixed-length vector where each value holds its semantic information. A pre-trained BERT [26] model is employed to obtain the sentence embedding for the tweet. At present, there are several varieties of pre-trained BERT; in this work, we have used bert_base_case(https://huggingface.co/bert-base-uncased) pre-trained model that is trained on the huge corpora of English sentences. This model is uncased in a way that it does not differentiate between the two words, english and English. The size of the embedded vector in the case of bert_base_case is 768. BERT returns two vectors for a tweet, the first vector holds the embedded representation for each word in the tweet, and therefore, it has a size equal to max_length_o f _sequence * 768. On the other hand, the second vector is one dimensional of length 768, containing aggregated embedded representation for the whole tweet.

Model Architecture
On top of one-hot vector representation of the target, and BERT-based representation of tweet, we experimented with four learning techniques, namely, support vector machine (SVM), random forest (RF), neural network (NN), and convolution neural network (CNN) to develop the stance classifier. Figure 4 depicts the methodology for designing the architecture of the proposed models.  For designing the bert-cnn, we utilize the first output vector of the BERT, which is two-dimensional with size _ ℎ_ _ * 768. We run 2-D convolutions on this input with the kernel size 5 × 768 to select the significant words in the filter. The weights of the kernels are randomly initialized. The number of filters in the Conv2d layer is 100. The output of this layer is followed by ReLu (rectified linear) layer and max-pooling with size (4 × 1). Finally, the resulting vector with dimensions 100 * 8 * 1 is flattened and combined with the target vector, which yields a one-dimensional vector with size 800 + 11 = 811. The resultant vector is then passed through the fully connected classification layer of CNN.
The classification layers of both NN and CNN have three nodes corresponding to three labels, Against, Favor, and None, with a softmax activation function. The activation function computes the probability distribution over the labels where the class with maximum probability is taken as the predicted value for the type of stance. The weights are trained using the cross-entropy loss function and adam weight optimizer. The used batch size is 16. Both CNN and NN models are implemented using PyTorch framework.

Experiments and Results
During the experiments, we trained and tested our four models (bert-svm, bert-rf, bertnn and bert-cnn) on the respective training and testing sets provided in the SemEval-2016 [23] stance detection task. The solutionss submitted to the competition were evaluated using the macro-average of the F1 scores of Favor and Against classes. The F1 score for None class was not considered while evaluating the model performance; however, the class was considered to be important during prediction to deal with noise. Table 7 below shows the results obtained by our four designed models, along with the results obtained by the top four ranked teams in the competition. It can be seen that among three machine learning models (bert-svm, bert-rf and bert-nn), bert-nn achieves the highest value (67.17%) for Faverage. However, looking at the overall, our deep learning model, bert-cnn, outperforms all the approaches, including the top scorers of the competition, by obtaining 68.24% as an average F1 score. It is to be noted that top scorer teams also utilized deep learning techniques for designing the stance classifier. For example, the first ranked submission employed the long-term short memory network (LSTM) for building the classifier, and the team at the second position used the CNN approach on the input representation of the word2vec pre-trained model.
However, if we compare the individual F1 scores of favor class, although none of the bert-based classifiers could achieve the first rank, the bert-cnn gets the third-highest value (60.22) for favor class. Furthermore, in the case of the F1 score of against category, our machine learning model, bert-rf, performs the best with a 76.52% score. In the case of SVM, RF, and NN, we combine the target vector and the second output vector of the BERT model into one to get an integrated representation with size 779 (11 + 768). In the case of SVM and RF, this integrated vector is passed as input features to the models to develop the bert-svm and bert-rf stance detector, respectively. However, for NN, the integrated vector is feed into a fully connected layer working as a classification layer to design the bert-nn classifier.
For designing the bert-cnn, we utilize the first output vector of the BERT, which is twodimensional with size max_length_o f _sequence * 768. We run 2-D convolutions on this input with the kernel size 5 × 768 to select the significant words in the filter. The weights of the kernels are randomly initialized. The number of filters in the Conv2d layer is 100. The output of this layer is followed by ReLu (rectified linear) layer and max-pooling with size (4 × 1). Finally, the resulting vector with dimensions 100 * 8 * 1 is flattened and combined with the target vector, which yields a one-dimensional vector with size 800 + 11 = 811. The resultant vector is then passed through the fully connected classification layer of CNN.
The classification layers of both NN and CNN have three nodes corresponding to three labels, Against, Favor, and None, with a softmax activation function. The activation function computes the probability distribution over the labels where the class with maximum probability is taken as the predicted value for the type of stance. The weights are trained using the cross-entropy loss function and adam weight optimizer. The used batch size is 16. Both CNN and NN models are implemented using PyTorch framework.

Experiments and Results
During the experiments, we trained and tested our four models (bert-svm, bert-rf, bert-nn and bert-cnn) on the respective training and testing sets provided in the SemEval-2016 [23] stance detection task. The solutionss submitted to the competition were evaluated using the macro-average of the F1 scores of Favor and Against classes. The F1 score for None class was not considered while evaluating the model performance; however, the class was considered to be important during prediction to deal with noise. Table 7 below shows the results obtained by our four designed models, along with the results obtained by the top four ranked teams in the competition. It can be seen that among three machine learning models (bert-svm, bert-rf and bert-nn), bert-nn achieves the highest value (67.17%) for F average . However, looking at the overall, our deep learning model, bert-cnn, outperforms all the approaches, including the top scorers of the competition, by obtaining 68.24% as an average F1 score. It is to be noted that top scorer teams also utilized deep learning techniques for designing the stance classifier. For example, the first ranked submission employed the long-term short memory network (LSTM) for building the classifier, and the team at the second position used the CNN approach on the input representation of the word2vec pre-trained model. However, if we compare the individual F1 scores of favor class, although none of the bert-based classifiers could achieve the first rank, the bert-cnn gets the third-highest value (60.22) for favor class. Furthermore, in the case of the F1 score of against category, our machine learning model, bert-rf, performs the best with a 76.52% score.
In the present study, an annotated stance-based dataset was prepared regarding the claim 'TJ is responsible for coronavirus', we named this dataset as TabJam dataset. This dataset will be made available for the other researchers working in this area. In the following section, we are further keen to investigate the following research question with respect to the best performing bert-cnn stance classifier: Research question: can we further enhance the performance of the proposed bert-cnn model by utilizing the TabJam dataset?
For obtaining the answer to the research question, we integrated the TabJam-based tweets with the tweets of the SemEval benchmark dataset. However, certain filtering was needed to be done on the new stance-based dataset before combining it with the benchmark. The SemEval dataset has three labels (favor, against and none), whereas, TabJam dataset has four labels (endorse, deny, question, and neutral). In order to maintain the symmetry with the benchmark dataset, we removed all the instances (tweets) from the new dataset that were labeled with "question" and "neutral". In addition, since the TabJam dataset contains duplicated tweets, we prepared the dataset with unique tweets only by removing duplicate ones. Afterward, endorse and deny labels of the filtered dataset are replaced with favor and against labels, respectively. Finally, the new dataset is split in the ratio of 8:2 to produce train (TabJam) and test (TabJam) datasets, respectively. The instances of the train (TabJamJ) and test (TabJam) are then combined with train (SemEval) and test (SemEval) to obtain train (SemEval + TabJam) and test (SemEval + TabJam) datasets, respectively.
To answer the above research question, the bert-cnn model is again trained and tested on the train (SemEval + TabJam) and test(SemEval + TabJam), respectively, named this model as BERT-CNN (combined) . The model achieved 65.06 and 75.00 as F1 scores of favor and against classes, thereby giving 70.03 as F average. Upon comparing these scores with the results of models given in Table 7, it can be realized that the BERT-CNN (combined) receives the highest score for F1 score for a favor (65.06) with the considerable margin of 4.13% (the highest shown in the table for favor class is 61.98). In addition, in the case of the average F1 score, BERT-CNN (combined) (70.03) performs better than bert-cnn (68. 24).
Furthermore, we also tested our BERT-CNN (combined) model on the original SemEval benchmark test dataset where we achieved 61.59 and 75.96 as F1 score of favor and against classes, respectively, thereby giving 68.78 as F average. It can be seen that BERT-CNN (combined) (68.78) performs better than BERT-CNN (68.24). Thus, from these results, we observed that the newly introduced TabJam dataset has the potential to enhance the performance of stance classification, and therefore will be highly beneficial to the researchers working in these domains.

Conclusions
The paper analyzes the behavior of Twitter users towards religion-based misinformation disseminated during the COVID-19 pandemic. The main objective of the study is to determine the role of social media in instigating social instability during the spreading of unverified news. The analysis is conducted on local news of India, which claimed that the Tablighi Jamaat (an Islamic gathering) held in mid-March 2020, during the COVID-19 crisis, is responsible for a disease outbreak in the country. To conduct the analysis and perform the experiments, an annotated tweet dataset is prepared with four reaction-polarities as classes, namely, endorse (D), deny (D), question (Q), and neutral (N).
The analysis is conducted in 5 phases. In the first phase, the tweeting pattern of the users was evaluated, where we observed the synchronization between the four user classes (E, D, Q, and N) and news articles published by the media. In the second stage, we observed the sign of social unrest on Twitter regarding the claim as the two extremely opposite classes of users, E, and D were constantly clashing with each other. The analysis in the third stage showed the sign of communal violence, hate, and fear in tweets shared by E and D user communities. The fourth stage of analysis presents the impact of time on the user language as the hashtags used in the content tend to move towards more negative and violent over the period. In the last part of the analysis, we notice the birth of echo chambers as the four classes, E, D, Q, and N, were using similar language within themselves. It further indicated the sign of social unrest as E and D users were sharing only tweets that contradict each other. Therefore, the five staged analysis conducted in this paper supports the hypothesis that the unverified claim having a religious touch have the potential to catalyze communal violence and instability in society.
Based on the analysis, the paper also proposes a deep learning classifier for stance detection to assist fact-checking organizations to curb the spread of fake news articles. The stance classifier would help the fact-checker in automatizing the process of monitoring the attitude of the people towards news headlines to determine their veracity and identify suspicious claims. The work utilizes a pre-trained BERT model for representing the text in mathematical vectors. The experiments are conducted on the tweets dataset provided by the SemEval-2016 competition for stance detection. The results show that the CNN model trained on BERT representation (bert-cnn) outperforms the state-of-the-art models. In addition, we show that the annotated reaction-based polarity dataset (TabJam) prepared in this study further enhances the performance of the stance classification model, highly beneficial to the researchers working in these domains.
In the future, an analysis can be conducted on worldwide news to have a deeper exploration of the impact of religion-based unverified content in persuading society. We can also study the difference in reactions of the users (endorsement, refutation, and counternarrative) when they are exposed to credible vs. incredible news. The study would assist in designing an efficient system for detecting and controlling the diffusion of unverified and unreliable content on social media. Furthermore, the dataset prepared in this work can be used to design a stance classifier for automatically assigning E, D, Q, and N labels to the new posts or tweets.

Data Availability Statement:
The data presented in this study are available on request from the corresponding authors. The data are not publicly available due to privacy and ethical concerns.