**Table 1.** Summary of attempts to process COVID-19 social data.

*ISPRS Int. J. Geo-Inf.* **2022**, *11*, 476

**Table 1.** *Cont.*


To summarize, researchers used analysis of social media content to learn more about how people react to the COVID-19 pandemic. The exiting work has been conducted with English data, while Arabic data receives fewer contributions. Moreover, the existing research works on Arabic social data do not consider the spatial-temporal aspect of the COVID-19-related content. Furthermore, the correlation between official health data and social media material has not been fully studied.

#### **3. Research Methods**

In this paper, we propose a method to automatically detect and process social datasets containing Arabic tweets related to the COVID-19 pandemic using ML models and topic detection and tracking techniques. Figure 1 shows the workflow of our methodology, while the below subsections describe each phase in more detail. Because we focused on analyzing the spatio-temporal social data in Arabic tweets related to the COVID-19 pandemic, accordingly our methodology started with data collection of Arabic tweets. We found two publicly shared Arabic datasets for COVID-19 tweets [11] (3,314,859 tweets) and [12] (2,111,650 tweets). Unfortunately, both have a low number of geo-tweets, which is about 2% of the total tweets as discussed in Section 2. Thus, we decided to merge both datasets (about 5.5M tweets) to feed our experiments. Then, we processed the merged dataset to infer locations from non-geotagged tweets, thus generating a new dataset of location-enabled tweets which contains about 46% (about 2.5M tweets) of the original combined dataset.

**Figure 1.** The Workflow of our methodology to analyze the social data.

Subsequently, useful features are extracted from COVID-19-related tweets. Then, classification techniques are applied to the extracted feature vectors to detect topics from the tweet's text and generate the hot topics. Then again, the sentiment analysis technique is applied with multiple perspectives to infer the sentiment and opinion polarity at many spatial (city, country, region) and temporal (day and month) levels. Next, correlation analysis between the collected tweets' data and the official health records is applied at different spatial granularities, such as country and region. Finally, we demonstrate the visual analytics based on a comprehensive set of experiments use the new Arabic COVID-19 dataset. The following subsections describe each step in our methodology.

#### *3.1. Dataset Collection*

The Twitter platform is well suited for studying the sentiment of users during COVID-19, with over 353 million monthly active users. Therefore we used Twitter because it focuses on becoming a broadcast platform similar to a real-time news station, with posts by default being global-readable to link people to the rest of the world. Most of the public Arabic datasets related to COVID-19 contain a low number of location-enabled tweets, such as Alanazi et al. [11] and Haouari et al. [12]. Both datasets are collected using Twitter Streaming API that matched Arabic tweets with the set of COVID-19-related keywords widely used by people, news media, and official organizations such as Coronavirus, Corona, and Pandemic. We collected tweets from both datasets and from which we built a new geo-tagged dataset. In addition, we defined the period from 1 January 2020, to 30 November 2020, to filter tweets on the merged dataset. Unfortunately, based on the privacy policy of Twitter, both shared datasets contain only tweet IDs. Therefore, we hydrated (recollected) the dataset to get the full tweet objects from Twitter using the TWARC and Hydrator Python libraries developed for this purpose, which helps to generate JSON files for each day's tweets. The new dataset contains more than 5.5M tweets (5,054,141 tweets are unique and about 2.5M geo-tweets) with an average of Words per Tweet equal to 21. Table 2 presents the statistics of the monthly distribution of unique tweets, unique words, unique hashtags, and unique user IDs.


**Table 2.** Statistics of the monthly distribution of dataset contents.

Figure A1 illustrates the monthly distribution of tweets and hashtags during the period from February to April 2020 in which most countries around the world were in COVID-19 Lockdown. One can notice the general trend of tweeting about COVID-19 was at its peak in March and then gradually reduced to normal about normal levels. It shows that the general trends of the number of tweets and hashtags were similar.

#### *3.2. Features Extraction*

We developed several processes for the Feature Extraction module, as shown in Figure 2. The first process "Prepare Dataset" contains several sub-processes: "Clean Dataset", "Filter Fields", and "Prepare Arabic Text" The "Clean Dataset" process removes null values from the tweet's object. On the other hand, the "Filter Fields" process removes the unnecessary fields from the tweet's metadata. Subsequently, to prepare the Arabic text in each tweet's object for further text analysis, the process "Prepare Arabic Text" has been applied to perform the following processes. Figure A2 provides a sample of Arabic tweets with their English translation after applying all of the following:


**Figure 2.** Reprocess COVID-19 tweets dataset.

Each tweet object contains several metadata fields that hold much information such as tweet text, hashtag, and user ID. The available features can be incorporated into our ML model with relative ease. The geo-location information, such as "Place" and "User Information", are important pieces of information that define the originating location of the tweet. Unfortunately, it needs further work because it is based on the user's choice to enable or disable the "Place" option in the Twitter settings. Hence, the number of geotagged tweets is so small compared to the total number of tweets. In the COVID-19 Arabic dataset [11] the number of geotagged tweets was 64,705 tweets (2%). Similarly, the ArCov-19 [12] dataset has only 50,856 tweets (2.4%). Therefore, we develop an algorithm to "generate Location-Enabled Tweets" from the non-geotagged tweets, as illustrated in Algorithm 1. The objective of the algorithm as shown in Algorithm 1 is to extract the location-enabled information. It required using the GeoDB, a manually developed Geo-location database that contains Bilingual names (English and Arabic) of world country names, capital cities, and famous towns in the Arab region. This approach analyzes all geo fields found with the tweet's object such as "Place name", "Country", and "User location" to extract the tweet source. Unfortunately, the user location information is an optional field that may is manually entered by the users. Therefore, it may be written in many languages or contained misinformation.

Our approach worked on two levels to extract chrononyms and astionyms (chrononyms: proper names of regions or countries; astionyms: proper names of towns and cities [33]) from user location. The first level tries to infer the country name from the information written in the user location metadata ([user][location]). It queries the GeoDB to extract the country name matching the ['user'] ['location']. Meanwhile, the second level works only if the first level failed, and it tries to infer the country name based on detecting the city name from the information in user location metadata and then retrieves the country name containing that city.

As a result of applying the Location Extraction algorithm, we increased the size of the experiments' dataset from about 115K (2%) up to 2.5M geotagged tweets (46%). Figure 3 illustrates the percentage of the geotagged tweets before and after applying our approach. Meanwhile, Figure A3 presents the monthly distribution of geo/non-geo tweets in our new Geo-Tweets dataset. Finally, using the occurrence-based approach, we run a process to numerically analyze the location-enabled Arabic tweets in the new geo-tweets' dataset (2.5M tweets). Most of the Arabic tweets are coming from Arab users living around the world, and most of them live in the Arab region. As a result, we compared the two regions (Arab and non-Arab) for originating the Arabic tweets that feed our experiments. This process extracted all information, which presents the proportion between Arabic tweets and top Hashtags tweeted by users in Arab and non-Arab regions. Figure 4 illustrates the comparison between Tweets and Hashtags (Top percentages only, more than 1%) in Arab and non-Arab countries during the time frame from January to November of 2020. Note that the top two countries for both tweet generation and hashtags in the Arab regions are Saudi Arabia and Kuwait. On the other hand, UK and France were the top two countries in the non-Arab region.


**Figure 3.** Percentage of the Geo-tagged tweets in the new COVID-19 Tweets Dataset after applying the Location Extraction algorithm.

**Figure 4.** Tweets and Hashtags in Arab and non-Arab Countries.

#### *3.3. Topic Clustering*

In this section, we conducted some experiments to present an initial outcome of dataset distribution analysis. Our Tweets dataset was formed from two datasets originally collected using a list of keywords (about 45 unique keywords). We applied the AraVec tool to extend the list of keywords based on semantic matching. AraVec is a word embedding open-source project that aims to provide free and efficient word embedding models to the Arabic NLP research community [34]. The new list contains more than 1600 keywords. Then, we filtered the list (160 unique topics only) by removing topics non-related to COVID-19 hot topics such as names of countries and cities. Then, we applied an occurrence-based technique to each tweet text to find the best-matched keywords that represent the topics illustrated by tweets.

We used the FuzzyWuzzy Python library (an open-source string similarity matching library) to determine string similarity as a degree out of 100. It uses the Levenshtein Distance that calculates the differences between sentences. It also uses the Fuzzy C-Means (FCM) clustering method to generate a recommender system predicated on homogeneous attribute measures [35]. Figure 5 presents the monthly distribution of the top topics during the time frame from January to November 2020. Meanwhile, Figure 6 illustrates the comparison between the top 20 topics tweeted by users in Arab and non-Arab countries. Figure A4 displays the word cloud of the top topics. Figure A5 shows the comparison between top five topics in Arab and non-Arab countries. While, Figure A6 illustrates the monthly top topics distribution. The top five topics in the top 10 countries is illustrated in Figure A7. Figures A8 and A9 present the distribution of some main topics (Vaccine and Treatment) in Arab and non-Arab countries.

**Figure 5.** Monthly distribution of Top 10 Topics in the Tweets Dataset.

Subsequently, to have a higher-level vision of topics found in Arabic tweets, we have used clustering techniques to cluster the 140 top topics extracted from our experiments into main topics. We applied the FCM clustering method that extracts 14 clusters to represent main topics. Table 3 represents the main high-level topics and their coverage. The coverage of a topic is calculated from the distribution of top topics in Arabic tweets in our Geo-dataset.

Figure 7 shows the percentage of main topics distribution in Arab and Non-countries. It illustrates that the top topics in Arab countries are Corona, Prayer, Lockdown, Hospital, and vaccines. Meanwhile, in non-Arab countries, the top topics are Sterilization, Organization, Feeling, Disease Symptoms, and Treatment. Figures A8 show Treatment topics tweeted in the Arab countries more than Vaccine topics during the first half of the year 2020 and vice versa. Meanwhile, Figure A9 displays the Vaccine topics appeared on social media in non-Arab countries from the second quarter of 2020.


#### **Table 3.** Main topics clusters.

#### *3.4. Sentiment Analysis*

Social data mining has recently become an attractive field of research that relates to the method of automatically extracting valuable information from computerized textual data [36]. It includes various associated fields such as question answering systems, text summarization, and SA [2]. Generally, research on social data mining and SA can be classified into supervised and unsupervised approaches [37]. Supervised (or corpus-based) approaches require labeled datasets, usually expensive to collect and label. On the other hand, the unsupervised approach (recognized as the lexicon-based approach) depends on

extracting features for classification and clustering purposes. For instance, unsupervised sentiment analysis employs sentiment lexicons built with the assumption that words have prior sentiments. Consequently, because of the variety and wide distribution of informal writing in languages, it is a time-consuming and challenging task to create such lexicons. Many ML classifiers can be applied to improve the performance of SA on both approaches such as Decision Tree (DT), Support Vector Machine (SVM), Multinomial Naïve Bayes (NB), or deep neural networks, among others.

We used the lexicon-based approach in the SA processes with the social network data (tweets from Twitter) as an opinion resource. We built our Arabic sentiment lexicon by merging nine Arabic lexicons (Bing Liu Lexicon; NRC Emotion Lexicon; MPQA Subjectivity Lexicon; SemEval-2016 Arabic Lexicon; AEWNA Lexicon; NileULex Lexicon [8,38–42] previously tested by the research community. The Arabic sentiment lexicon contains annotation for Arabic words such as "positive", "negative", or "neutral". We used this lexicon to extract the feature of the tweet's polarity as the numbers of positive, negative, and neutral words. Then, we used polarized bag-of-words features, representing a tweet as a set of words without order, each word being a feature. Then combined these features with the number of words by polarity. Principally, the Bag of Words (BoW) technique is a word-frequency approach that counts positive, negative, and neutral words in each tweet to assign the tweet polarity as "positive" (if positive count—negative count > zero), "negative" (if positive count—negative count < zero), otherwise it is "neutral". Subsequently, we applied four ML classifiers—Decision Tree (DT), Multinomial Naïve Bayes (MNB), Linear Support Vector Machine (SVM), and Random Forest (RF)—to check the classification performance of the sentiment analysis.

To apply the sentimental analysis processes, we developed a Sentiment Extraction algorithm for the SA processes to run in our experiments. Algorithm 2 starts with loading the Arabic corpus of the text collected from Arabic tweets, which are extracted from the geo-tweets' dataset. Subsequently, the approach prepares the Arabic corpus by applying many sub-processes: (1) filtering tweets to throw away the unnecessary fields and use the necessary fields (such as: "id", "text", "country code"). (2) Remove duplicated tweets. (3) Clean tweets' text by removing special characters such as (#, &, %), hyperlinks, punctuation, numbers, Unicode characters, non-ASCII characters, and unnecessary white spaces. (4) Normalize Arabic characters (Alef, Yaa, . . . ) in tweets' text. (5) Remove Arabic stop words such as the Arabic prepositions (Men, Ela, Fe, . . . ). (6) Remove duplicated words and single-character words as well. (7) Tokenize Tweets' text. (8) Stem tweets' words to rooted words.

Subsequently, the algorithm will run on each tweet's words to get the tweet's polarity by using our Arabic sentiment lexicon to get polarity for each word then run the BoW technique to set the polarity for each tweet. Finally, the algorithm runs four ML classifiers to improve the accuracy of the SA process. The algorithm extract features and labels from the Arabic polarity corpus. Then, it divides the corpus into train and test datasets by splitting with ratios 80% and 20% consequently. Next, the algorithm vectorizes the training dataset using the TF-IDF Vectorizer. Following this, the algorithm predicts the features using the classifier model applied to the test dataset. Finally, a confusion matrix with the classifier's scores is outputted such as the precision, recall, f1-score, and accuracy values.


**Algorithm 2:** Sentiment Extraction

#### **4. Results and Discussion**

Figure 8 illustrates the monthly results from the sentiment analysis of COVID-19 Arabic tweets. Meanwhile, Figure A10 displays the top scores evaluating sentiment analysis in Arab countries. It shows Arab countries have the most positive/negative/neutral feelings based on users' opinions.

**Figure 8.** SA of COVID-19 Arabic Tweets.

Moreover, we trained four classic ML algorithms (DT, MNB, SVC, and RF) using our Geo COVID-19 dataset. Next, we evaluated the results with four sentiment-labeled Arabic datasets collected from Twitter, as shown in Table 4. Furthermore, we employed an NLP pre-training model dubbed Bidirectional Encoder Representations from Transformers (BERT), which was developed to pre-train deep bidirectional representations from the unlabeled text by conditioning on both left and right contexts in all layers [43]. We used the Arabic-BERT model (developed by Safaya et al. [44]) trained on our Geo COVID-19 dataset and evaluated with the same four datasets. We run certain model modification iterations, assess using the other datasets, and then examine how it affects the overall performance.

**Table 4.** Sentiment-labelled Arabic satasets.


Figure 9 presents the comparison between results collected after applying ML algorithms trained and evaluated with five Arabic datasets. We used both Unigrams and Bigrams features with TF-IDF Vectorization. We found that scores are close for both, so we present unigrams only. It shows the Accuracy measure results of the traditional ML classifiers for unigram and TF-IDF feature representation. It shows that applying the SVM and RF classifiers achieved a high score on our Geo-COVID-19 dataset than using other datasets. The results achieved by some ML classifiers are also higher in performance. We applied the Arabic-BERT model and compared our findings using the five datasets that accomplished a higher accuracy than the others, especially with the BERT-mini model.

**Figure 9.** SA Classifiers' performance applied on Arabic Datasets.

#### *4.1. Correlation Analysis*

Daily, the official health records related to COVID-19 statistics are published by many official organizations such World Health Organization (WHO), European Centre for Disease Prevention and Control (ECDC), and Johns Hopkins University (JHU). We collected official COVID-19 records from ECDC and JHU for world countries during the time frame from January to November 2020. We implemented many mechanisms for the correlation-based analysis using occurrence-based to find the correlation between sentiment analysis, official health providers' data, lockdown information, and topic frequencies. Figure 10 illustrates the correlation between sentiment analysis of Arabic tweets related to COVID-19 and the COVID-19 new cases over the world. It shows that there were some high negative feelings at the beginning of the COVID-19 pandemic until the second quarter of the year 2020, subsequent it was in a steady state while cases increased. In addition, Figure 11 shows separately the correlation between feelings of the Arab tweets in the Arab countries and non-Arab countries and the COVID-19 new cases with some differences in both, especially in the last quarter of 2020.

**Figure 10.** Correlation between SA and the Official Health Records.

**Figure 11.** Correlation between SA and the Official Health Records in Arab and Non-Arab Countries.

Similarly, Figure A11 illustrates a considerable difference in the correlation between feelings shown in Arabic tweets and the COVID-19 new cases announced in some countries, around the world. On average most feelings are equalized between positive and negative while the number of official new cases of COVID-19 are increased. In response to the COVID-19 outbreak, governments are imposing several measures against the COVID-19 pandemic such as {School closing}, {Workplace closing}, {Cancel public events}, and {travel restrictions}. On 23 January 2020, the first COVID-19 pandemic lockdown was initiated in Wuhan [49]. Since January 2020, most governments throughout the world enacted full or partial lockdowns to stop the virus from spreading, leaving millions stranded. Moreover, a third of the world's population is restricted in some way [50]. From the middle of March 2020, several Arab countries began implementing various types of lockdowns.

Some organizations, such as Oxford University, collect data on 20 indicators to inform a Risk of Openness Index, which aims to aid countries to understand whether it is safe to "open up" or "shut down" in their fight against the coronavirus [51]. We got data from the Oxford dataset related to the lockdown status worldwide. We used data of the "School closing", "Workplace closing" and "Cancel public events" indicators as the most important indicators to feed our mechanism applied in an experiment aimed to find the correlation between lockdown and the spatial-temporal data found in our Arabic tweets' dataset. Figures 12 and 13 displays the comparison between the correlation of the lockdown with COVID-19 new cases in Arab and non-Arab countries, which present the impact of lockdown

on the spread of COVID-19 new cases. It shows a negative correlation between the number of COVID-19 new cases and the number of lockdown days in the Arab region— especially for the indicator "Cancel Public events". While, in the non-Arab region, there is a negative correlation between the number of COVID-19 new cases and the number of lockdown days—especially for the indicators "Cancel Public events" and "School Closing".

**Figure 12.** Correlation between lockdown and Official COVID-19 New Cases in Arab Countries.

**Figure 13.** Correlation between lockdown and Official COVID-19 New Cases in Non-Arab Countries.

Figures 14 and 15 clarify the correlation between lockdown and the sentiment analysis of Arabic tweets tweeted by users during the time frame from January to November 2020, although the lockdown started late in Arab countries. The relationship shows different perspectives regarding the lockdown indicators and the sentiment of Arabic tweets related to COVID-19 in Arab and non-Arab countries. For example, they clarify a positive correlation between the positive feelings shown in Arabic tweets and the number of closing days of a school lockdown.

Back to the top-topics issue illustrated early in the above sections, we implemented another mechanism to define the correlation between sentiment analysis and topic-frequencies in a spatial-temporal manner. Subsequently, to have a wide vision of topics found in Arabic tweets, we clustered them into main top topics as discussed early in Table 3. Figures 16 and 17 illustrate the relationship between main topics and the sentiment analysis of Arabic tweets presented those topics in Arab and non-Arab countries. It shows the positive/negative feelings associated with each topic in that region.

**Figure 14.** Correlation between Lockdown and SA in Arab Countries.

**Figure 15.** Correlation between Lockdown and SA in Non-Arab Countries.

**Figure 16.** Correlation between main topics and SA in Arab countries.

**Figure 17.** Correlation between main topics and SA in non-Arab countries.

We ran multiple experiments on our Arabic tweets COVID-19 dataset and the other data collected from previous experiments to explain the insight obtained from the analysis of increasingly large datasets. The results could play an essential role in representing our large-scale dataset with official data resulting early. We provided figures for visualizing data and findings in an approachable and stimulating way to enable us to extract information, superior to understand the data, and make more effective decisions. We tried to visualize the dataset components on world maps to learn COVID-19 pandemic distribution affects social media users globally. Figures A12–A14 illustrates the distribution of Arabic tweets, Arabic Hashtags, and Twitter users around the world. They clarify that most of them originated from Saudi Arabia and Egypt because they have the highest population as well the highest number of Internet users [1]. Meanwhile, Figure A15 displays the SA of Arabic tweets from our Geo-dataset over the world. It shows the average sentiment analysis of Arabic tweets is negative in most countries. Furthermore, Figure A16 illustrates the Word cloud of Sentiment Analysis over the world.

In addition, Figure A17 shows the distribution of COVID-19 confirmed cases globally, which emphasizes the known fact that the USA, India, and Brazil had the highest total number of COVID-19 confirmed cases. Finally, Figure A18 shows the distribution of the total number of lockdown days around the world based on the indicators used such as "Cancel Public Events" and "School closure". Specifically, it shows that most Arabic countries were in Lockdown for almost the year.

#### *4.2. Discussion*

In this research, we developed a novel location inference technique from non-geotagged tweets based on user profiles and textual content, which resulted in increasing the total percentage of geotagged tweets from 2% to 46% (about 2.5M tweets). Currently, we are conducting further research to improve the proposed location inference technique by using location inference from tweet textual content, such as the technique presented in the Arabic NER model using Flair embeddings [52]. This aims at increasing the percentage of produced geotagged tweets by utilizing an NLP process to manipulate the tweet's text. Moreover, evaluating our geo-dataset compared with four sentiment-labeled Arabic datasets (using four ML algorithms and one DL algorithm) shows that it accomplished a higher accuracy than the others, specifically with the DL model (BERT-mini).

In addition, some of the conducted correlation analysis shows that negative feelings of Arab Twitter users were raised during this pandemic. Moreover, we illustrated that there is a correlation between the discussed topics in the Arabic social media, such as lockdown and travel restrictions, which were enforced by governments, and the number of COVID-19 new cases. Furthermore, the analysis showed that a positive correlation exists between the negative feeling of Arab users and the number of daily confirmed cases of COVID-19. Moreover, evaluating sentiment analysis in Arab countries shows that Saudi Arabia, Kuwait, and Egypt have the most positive/negative/neutral feelings based on users' opinions.

The correlation analysis in related work was mostly focused on time series analysis over health data, where we demonstrated in this work that spatio-temporal correlation with health and social data, can provide us with deeper insights on how people's sentiments differ over different topics and subtopics, and with different spatial scales, such as cities, countries, and regions. Finally, we visualized data and findings to understand the data components and distribute the effects of using social media related to COVID-19 globally.

#### **5. Conclusions**

This paper introduces a comprehensive social data mining approach for deriving COVID-19-related insights in the Arabic language, with a focus on the correlation between spatio-temporal social data and health data. In addition, it presented a sentiment analysis mechanism at multiple levels of spatial granularities and several topic scales. In addition, a technique to infer geo-information from non-geotagged tweets was developed, which increased the total percentage of location-enabled tweets from 2% to 46%, superior to most previous related works. To verify sentiment analysis performance, we applied sentimentbased classifications at many location resolutions (regions/countries) and some topic abstraction levels (subtopics and main topics) to derive people's opinions. Finally, we conducted many experiments and visualized our results based on the generated geosocial dataset, sentiment analysis, official health records, and lockdown data worldwide. Our findings show the great potential of integrating social data mining with other data sources, such as health data, to predict the evolution of such phenomena. In addition, such correlation can later be applied to other types of data such as contact tracing and GPS data, to provide an in-depth understanding of human behavior and the correlation between social and physical user interactions.

We intend to expand the dataset in the future to include more Arabic social contents to analyze the most recent periods when the social media users' focus has changed from COVID-19 in general to vaccinations. In addition, we plan to conduct further research to improve the proposed location inference technique by utilizing an NLP process to manipulate the tweet's text to increase the percentage of produced geotagged tweets.

**Author Contributions:** "Conceptualization", Imad Afyouni, Ibrahim Hashem and Zaher Al Aghbari; "Data curation", Tarek Elsaka; "Formal analysis", Tarek Elsaka and Imad Afyouni; "Investigation", Zaher Al Aghbari; "Methodology", Tarek Elsaka, Imad Afyouni and Ibrahim Hashem; "Software", Tarek Elsaka; "Supervision", Zaher Al Aghbari; "Validation", Ibrahim Hashem; "Visualization", Tarek Elsaka and Imad Afyouni; "Writing—original draft", Tarek Elsaka, Imad Afyouni, Ibrahim Hashem and Zaher Al Aghbari; "Writing—review & editing", Tarek Elsaka and Imad Afyouni All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **Abbreviations**

The following abbreviations are used in this manuscript:



#### **Appendix A. More Detailed Figures**

**Figure A1.** Monthly distribution of tweets and Hashtags.


**Figure A2.** Sample of Arabic tweets with English translation.

**Figure A3.** Monthly distribution of Geo and Non-geo tweets in Tweets Dataset.

**Figure A4.** Word cloud of the top topics in the COVID-19 Arabic Tweets Dataset.

**Figure A5.** Top 5 Topics in Arab and Non-Arab Countries.

**Figure A6.** Monthly Top Topics in Arab and Non-Arab Countries.

**Figure A7.** Top 5 Topics in Top 10 Countries.

**Figure A8.** Frequency-Occurrence of Some Main Topics in Arab Countries.

**Figure A9.** Frequency-Occurrence of Some Main Topics in Non-Arab Countries.

**Figure A10.** Top Sentiment (Positive/Negative/Neutral) in Arab Countries.

Feb

Jan

SA vs. COVID-19 New Cases in UK

0

JunJul

Aug

Sep

Oct

Nov

50,000

100,000

150,000

Apr

May

Mar

Iran

0% 10% 20% 30% 40%

**Figure A11.** Correlation between SA and Official Health Records in some countries.

**Figure A12.** Distribution of Arabic Tweets.

**Figure A13.** Distribution of Arabic Hashtags.

**Figure A14.** Distribution of Users.

**Figure A15.** Sentiment Analysis of Arabic Tweets.


**Figure A17.** Distribution of COVID-19 Cases.

**Figure A18.** Covid-19 Lockdown Days.

#### **References**

