COMBINING MACHINE LEARNING AND SEMANTIC ANALYSIS FOR EFFICIENT MISINFORMATION DETECTION OF ARABIC COVID-19 TWEETS

International Journal of Computer Science & Information Technology (IJCSIT) Vol 14, No 3, June 2022
DOI: 10.5121/ijcsit.2022.14303 31
COMBINING MACHINE LEARNING AND SEMANTIC
ANALYSIS FOR EFFICIENT MISINFORMATION
DETECTION OF ARABIC COVID-19 TWEETS
Abdulrahim Alhaizaey and Jawad Berri
Department of Information Systems, King Saud University, Riyadh, Saudi Arabia
ABSTRACT
With the spread of social media platforms and the proliferation of misleading news, misinformation
detection within microblogging platforms has become a real challenge. During the Covid-19 pandemic,
many fake news and rumors were broadcasted and shared daily on social media. In order to filter out these
fake news, many works have been done on misinformation detection using machine learning and sentiment
analysis in the English language. However, misinformation detection research in the Arabic language on
social media is limited. This paper introduces a misinformation verification system for Arabic COVID-19
related news using an Arabic rumors dataset on Twitter. We explored the dataset and prepared it using
multiple phases of preprocessing techniques before applying different machine learning classification
algorithms combined with a semantic analysis method. The model was applied on 3.6k annotated tweets
achieving 93% best overall accuracy of the model in detecting misinformation. We further build another
dataset of Covid-19 related claims in Arabic to examine how our model performs with this new set of
claims. Results show that the combination of machine learning techniques and linguistic analysis achieves
the best scores reaching 92% best accuracy in detecting the veracity of sentences of the new dataset.
KEYWORDS
Misinformation, machine learning, Arabic NLP, contextual exploration, rumor detection.
1. INTRODUCTION
In recent years, the use of social media platforms has increased considerably, imposing itself as
one of the most important sources for spreading news and broadcasting opinions. Social media
platforms have become a rich source of information and have provided many researchers with an
opportunity to analyze the vast amount of data that users of these platforms post on a daily basis.
Within the Arab world, a recent study indicated that social media platforms had become the
dominant source of news for Arab youth [1]. Twitter1
is one of the most important of these
platforms, as it is the second most used platform after Facebook2
by Arabic speakers [2]. The
Arabic language is the sixth largest language present on Twitter [3], making it an important
source for data mining research and analysis.
With the emergence of the Covid-19 pandemic, Twitter, like many other social media platforms,
has witnessed an increased interaction by people in expressing opinions and discussions related to
the Coronavirus. Regardless of the different measures enforced by governments to contain the
spread of this virus, the spread of misleading news about the Coronavirus represents a unique
1
https://twitter.com/home
2
https://www.facebook.com/

32
phenomenon, requiring many decision-makers to hold daily conferences to answer and refute
many widespread fake news and rumors. Moreover, with millions of daily posted tweets and
replays, it becomes nearly impossible for reliable fact-checking services to keep up checking
claims with such a massive amount of data. We argue that data mining and machine learning
tools can take some of the burden off fact-checking platforms that fundamentally rely on the role
of experts in checking and refuting rumors and misleading news. A broad range of literature has
studied claim verification using machine learning tools in the English language online content.
Many new recent studies address the problem of Covid-19 rumors detection in English online
content such as [4] and [5]. However, very few studies have addressed this problem in Arabic
online content. For example, [6] suggested a model for rumor detection in Arabic tweets using
semi-supervised and unsupervised expectation-maximization. Another study [7] introduced an
Arabic corpus of fake news targeting YouTube. However, both of these works did not address the
misinformation detection related to the Coronavirus pandemic in Arabic language. In general,
studies on the verification of Arabic language claims and rumors, particularly those related to the
pandemics, are rarely explored. To the best of our knowledge, this is one of the very few studies
that address this topic.
This paper provides a model that uses machine learning tools combined with a semantic analysis
method to detect misleading news related to the Coronavirus pandemic. We used the ArCOV19-
Rumors dataset [8] in our study, which is an Arabic COVID-19 related dataset retrieved from
Twitter. The dataset ArCOV19-Rumors is the only available Arabic dataset to support claims
verification studies. In addition, the authors who released this dataset have intensively annotated
the veracity of the tweets. This intensive annotation makes it a good resource to support research
on misinformation detection as one of the major problems encountered during the pandemic.
Furthermore, the dataset is balanced in terms of the distribution of true and false tweets, making
it a sound resource for building and training a verification system. The dataset we used was
analyzed using Sentiment Analysis approaches to provide instant detection of fake news. Despite
that this model could arguably be a more affordable alternative than relying on expert editors or
journalists, this model could also help in immediately notifying those who are interested in fact-
checking such as professional journalists and editors running fact-checking services. We used the
Ar-Covid-19 dataset [8] in our study. This dataset contains a total of 1831 Arabic tweets with a
valid claim and 1660 Arabic tweets with a false claim related to the Coronavirus during the
period from 27𝑡ℎ January till the end of April 2020. Table 1. shows a statistical summary of the
labeled tweets in ArCOV19-Rumors dataset. Based on this labeled dataset, we developed a claim
verification system taking advantage of the capabilities provided by Python programming
language with a suite of libraries for natural language processing and machine learning. Our
model consists of multiple components: data collection, data preprocessing and cleansing, and
machine learning classification. In the next section, we provide a background of the related
works. Then, in Section 3, we present the methodology. Section 4 presents the results and
discussion. Section 5 shows how the accuracy of the classifiers can be enhanced on our testing set
of claims after applying some sematic based rules. The conclusion of the work is then provided.
Table 1. Statistics of Claims in ArCOV19-Rumors [8].
Label Number of tweets Percentage
False Tweets 1753 (48.9%)
True Tweets 1831 (51.1%)
Total 3,584 (100%)

33
2. LITERATURE REVIEW
In the internet age, the rise of fake news highlights the urgent need for organizational solutions
for such a challenging problem. Fake news and manipulating people for different reasons have a
long history and used to be addressed by journalistic norms of objectivity; however, social
networks have created a context in which fake news can attract a mass audience [9]. With the
emerging of Covid-19, fake news and claims spreading on social media platforms are immensely
remarkable. With the increased transformation to social media as the new news sources, the
traditional ones that had enjoyed good levels of public trust and reliability have lost their
popularity. Hence, a new way of verifying news on social media is needed. Fact-checking
websites can be seen as an attempt to alleviate this complex problem in the internet world. Very
few Arabic platforms provide Fact-Checking services to evaluate factual claims of news stories in
Arabic. For example, Misbar3
, AFP4
, and fatabyyano5
are digital services that have been launched
for the purpose of news verification. These fact-checking services rely heavily on their networks
of expert editors and journalists to run their fact-check services. However, despite that the use of
these fact-checking services is still not measured, the effectiveness of such services in the Arabic
world needs investigation.
There is a broad body of literature that would suggest that humans are not good at predicting and
decision making. For that reason, machine learning has been widely used in many real-life
applications to improve and enhance human decision-making. For example, in August 2016, the
Allegheny County Department of Human Services implemented a machine learning-based
predictive risk modeling tool called AFST [10]. This tool was explicitly designed to improve
child welfare call screening decisions. The tool was the result of a two-year process of
exploration about how existing administrative data could be used more efficiently to improve
decision-making at the time of a child welfare referral. Facebook also uses machine learning
algorithms to identify and remove sensitive content such as violence and sexual contents
automatically. In addition, realizing that it is impossible for human fact-checkers to review stories
on an individual basis with around two billion Facebook registered users, Facebook recently
started to use similar technologies to recognize false news and take action on a bigger scale [11].
Likewise, we believe that using the state of the art machine learning and artificial intelligence
techniques could be an effective tool in combating misinformation in the near future.
Instantaneous insights into fake news and rumors in social media could be done using sentimental
analysis (SA) and machine learning (ML) techniques. SA is a methodology used for natural
language processing. It is used to analyze human opinions, emotions, sentiments, and attitudes.
The noticeable increase in employing SA techniques in opinion mining has been co-occurred
with the increased use of social media, benefiting from the huge amounts of available data in
social media platforms. SA can be implemented using machine learning models. So, could we
take advantage of such techniques and methods to detect misinformation in social media? This is
what we are trying to do in this research. We apply SA using ML to Ar-Rumor dataset [8] to see
to what extent these techniques can be beneficial in detecting Arabic fake news. The main
contribution of this work is, in one aspect, that it shows the feasibility of automating the detection
of misinformation on the microblogging platforms such as Twitter. There must be automated
systems for such purposes, given the uncontrollable vast amount of content and users. The second
aspect is that misinformation detection in Arabic is rarely explored, which this paper has touched
upon. The third aspect is the combination between machine learning techniques and semantic
analysis which turned out to produce better results in detecting misinformation. The following
section presents our methodology for tackling such a problem.
3
https://misbar.com/
4
https://factuel.afp.com/ar/list
5
https://fatabyyano.net

34
3. METHODOLOGY
Once we retrieved and collected the dataset, we build our claim verification system, which
contains multiple components. The first part is data preprocessing and normalization using text
preprocessing techniques followed by an additional layer of data filtering and feature extraction.
The next component is the classification using different machine learning classifiers. Our model
is depicted in Figure 1. The details of the model components are described in the following
paragraphs.
3.1. Retrieving Dataset
We retrieved the dataset using Hydrator [12], a free tool for hydrating Twitter ID datasets. We
collected 9.4k labeled relevant tweets. A total of 3.6k of the collected tweets are labeled either
true or false. A total of 1,753 tweets were labeled false, and a total of 1,831 tweets were labeled
true. The distribution of true tweets versus false tweets is balanced, making it a good resource for
training verification systems.
3.2. Data Preprocessing and Normalization
After collecting the Twitter dataset, the data were prepared for analysis. Due to the noisy nature
of the microblog data such as Twitter tweets, text preprocessing and cleaning is an essential step
to eliminate noise and unnecessary data that do not represent any value about the sentiment of the
text. Data preprocessing plays a significant role in improving the accuracy of classification
algorithms and speeding up the training process to enhance the models' learning efficiency [13].
For this purpose, we used a natural language toolkit (NLTK) [14] of Python, which has a suite of
text processing libraries.
Preprocessing and cleansing include the following steps: removing repeated letters, removing
stop words, which are provided in [14], removing words that have less than two letters, removing
numbers and additional spaces, removing any non-Arabic letters, removing punctuation marks
and emojis, removing Arabic diacritics, and normalizing the data by converting multiple forms of
a letter into one uniform letter since some Arabic letters could be represented in different forms.
3.3. Additional Filtering
Besides eliminating noise data, and since the dataset might contain duplicate tweets and what
could be considered spam tweets, we performed an additional filtering step. In this step, tweets
with URLs, phone numbers, more than four hashtags, and duplicated tweets are classified as
spam, so we deleted them. Then, we removed hashtags, URLs, and mentions. After that, the
duplicated tweets are dropped to keep only one representation of each tweet. Finally, after
preprocessing and cleaning, some tweets resulted in a very short text of only a few words, so we
removed any tweet with less than two words because we believe such tweets do not contribute
towards the classification and need to be removed.
3.4. Feature Extraction
Predictive modeling requires text data to be customized and prepared before using it for
prediction purposes. The textual data have to be parsed and tokenized such that the words are
converted into numbers or to be used as input for the machine, which is referred to as
vectorization. For this purpose, the scikit-learn library [15] provides convenient tools to perform
feature extraction of text data. With this tool, we used a popular technique called "Term

35
Frequency and Inverse Document Frequency" (TF-IDF) to represent the data as a feature where
words are given frequency scores. This helps in determining the words that are more interesting
for classification. Also, Arabic stop words imported from NLTK [14] were used to drop specific
words from data. These words do not contribute towards classification.
3.5. Machine Learning Classification
Many supervised ML classifiers have been widely used in the literature for text analysis. We
have used four different common ML algorithms to analyze and compare the results, namely:
Logistic Regression (LR) [16], Random Forest (RF) [17], Multinomial Naive Bayes (MNB) [18],
and Support Vector Machine (SVM) [19]. Although many ML classifiers are available in text
classification problems, we adopted the four mentioned classifiers because these classifiers have
been widely applied and used in this context. The ones involved here are just chosen as a proof
of concept, although other classifiers could be chosen. These four ML algorithms have been
popularly used for text and sentiment analysis, so we found them proper to use in our study
context. After preprocessing data, we supplied the dataset to different ML algorithms to build a
classifier that learns from training labeled tweets. In order to build our classifier, we used a
Scikit-learn library, an open-source machine learning package in Python that provides various
classification algorithms. We have randomly set our ML algorithms to choose 20% of the dataset
as test sets. Then we have used the remaining 80% of the dataset to train our machine learning
models. The trained models can then predict whether another claim from the test set is true or
false claim. In our analysis, the dataset we used was classified into two classes, true and false.
Figure 1. Verification Model Architecture

36
4. EXPERIMENT AND RESULTS ANALYSIS
After building our verification prediction model, we have carried out our experiments to test and
evaluate the accuracy of this model. In order to evaluate our model, we used the accuracy metric
(1), which is a common metric to measure performance in machine learning classification and
pattern recognition. A true positive (TP) is an outcome where the machine learning model
correctly predicts the true claim. Likewise, a true negative (TN) is an outcome where the model
correctly predicts the false claim. Similarly, a false positive (FP) is a result where the model
incorrectly predicts the true class. And a false negative (FN) is a result where the model
incorrectly predicts the false claim. We further build another dataset of Covid-19 related claims
in Arabic to examine how our model performs with this new set of claims. Figure 2 shows the
analysis of the results of our model for the two datasets and the machine learning models chosen
in the experiments.
 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
𝑇𝑃+𝑇𝑁
𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁
 
Our first observation on the accuracy of our model is that the SVM model achieved 93%
accuracy compared to the remaining machine learning approaches, where the accuracy of LR,
MNB, RF are 91%, 91%, and 89%, respectively. On the other hand, when testing our model
against our dataset to measure its performance, we see that our current system performs well for
the new dataset that we built for the purpose of measuring the performance of the model. Our
accuracy analysis shows that our dataset results in a bit lower accuracy. The accuracy of the
machine learning algorithms in detecting veracity in our dataset is 79%, 75%, 85%, and 87% for
LR, RF, MNB, and SVM approaches, respectively. It can be observed that SVM and MNB
significantly outperform LR and RF in our dataset. This is because our dataset is relatively small.
In addition, the proper learning of our system with more data can increase the accuracy of the
system as a whole.
Figure 2. Results Analysis of the two datasets

37
5. SEMANTIC TWEETS ANALYSIS
In order to improve the accuracy of the results obtained in the previous step with machine
learning algorithms, we used a linguistic analysis method which aim to analyze the results of the
learning algorithms semantically. Machine learning algorithms are efficient in dealing with huge
data and can learn and improve. However, these algorithms may not classify the tweets in the
dataset correctly when it comes to very fine semantic analysis that is necessary to define the
meaning of sentences. On the other hand, semantic analysis lacks performance when dealing with
big datasets but is efficient in discovering fine semantic nuances in the meaning of tweets which
can falsify the results. A combination of both methods is crucial to improve the accuracy of
misinformation detection in the dataset used in the experiment. Semantic analysis should explore
the tweet linguistic context to be able to detect the right semantic value of the tweet. We used a
linguistic method, namely the Contextual Exploration (CE) method [22], which can analyze the
context and take the right decision for labeling sentences semantically. CE is used here to
compute the semantic value representing the meaning of a sentence. CE does not require huge
linguistic resources and fine semantic descriptions of language units. Instead, it requires only the
linguistic grammatical and lexical know-how used by a decision maker when assigning the
semantic value to a sentence [23]. A close look at the dataset shows that many linguistic markers
can falsely classify tweets' semantic value. We focused in this research on four semantic labels,
namely: the interrogation, the negation, the claim and the reported statement. Table 2 provides
illustrative examples of tweets from the dataset. Using interrogation sentences is common in
news headlines to grab the audience's attention. On Twitter, it is usually posted with an attached
URL to the news webpage. Interrogation is expressed in Arabic by using the interrogation mark
"‫؟‬" such as in the tweet example S1. This sentence questions the veracity of the Hanta virus that
appeared in China without stating any truth value. CE detects these sentences by the interrogation
mark and accordingly labels them as an interrogation sentence and reclassify them as Neutral.
Negation in standard Arabic can be expressed in different ways depending on the verb tense and
using the five negation forms: ‫ليس‬ ،‫لن‬ ،‫ما‬ ،‫لم‬ ،‫ال‬. For instance, sentence S2 in Table 2 is labeled as
a negation sentence since it includes the word "‫ال‬". These sentences are reclassified to the
opposite class assigned by the classifier.
Table 2. Examples of Tweet sentences.
Text Sentence (Tweet)
S1
‫الحقيقة؟‬ ‫ما‬ ...‫الصين‬ ‫في‬ ‫هانتا‬ ‫فيروس‬ ‫ظهور‬ ‫كورونا‬ ‫بعد‬
After Corona, the emergence of the Hanta virus in China ... What is the truth?
S2
"S Gene not detected"
‫عن‬ ‫بمنأى‬ ‫أنك‬ ‫يعني‬ ‫ال‬
‫كوفيد‬
-
19
!
"S Gene not detected" doesn't mean you are immune to COVID-19!
S3
‫كورونا‬ ‫جائحة‬ ‫بدأ‬ ‫قبل‬ ‫حتى‬ ‫موجود‬ ‫كورونا‬ ‫لقاح‬ :‫أن‬ ‫يدعي‬ ‫فيديو‬
)...(
Video claims that: Corona vaccine existed even before the Corona pandemic began (…)
S4
:‫البريطانية‬ ‫ميل‬ ‫ديلي‬ ‫صحيفة‬
‫الموسم‬ ‫هذا‬ ‫أخرى‬ ‫مرة‬ ‫اإلنجليزي‬ ‫الدوري‬ ‫يعود‬ ‫لن‬
)...(
The British Daily Mail: The English Premier League will not resume this season (...)
A claim is a statement conveyed to assert something either by the speaker or it is reported as a
claim of another party such in sentence S3, where the claim is reported from a video (third
person) using the verb "‫يدعي‬" (claims) and the colon mark (:) which is explicitly used in the
sentence. In general, when the speaker claims something (by using the pronoun I, or we) this
means that the claim is most likely definite and hence it is labeled as a positive sentence. On the
other hand, when the claim is stated by another party (not the speaker) the sentence is classified
as neutral. A reported statement is frequently used by news reporters who convey news
statements from other news agencies or information sources such in sentence S4. In such case the
sentence is labeled as neutral since the veracity of the reported sentence is referred to the claimer

38
without any responsibility on the speaker. It is worth mentioning that some sentences can be
associated with more than one semantic label. For example, sentence S4 in Table 3 is labeled as a
negation claim and a conveyed claim. For such sentences another semantic analysis stage is
necessary to assign dominant semantic label and then reclassify it. Sentences falling in these four
linguistic categories in the test dataset are handled properly and reclassified during the post-
semantic analysis. Then the accuracy of the dataset is recalculated.
Figure 3. Results with Post Semantic Analysis
Figure 3 shows the results after the post-semantic analysis. The post semantic analysis has clearly
improved accuracy for all algorithms for the test dataset. The proposed approach, which
combines machine learning algorithms with post semantic analysis has better accuracy then the
machine learning approach alone. Although the linguistic approach currently uses only four
semantic labels in the classification process, it may be further enhanced by including additional
semantic labels to capture the fine semantic meanings carried by natural language sentences.
6. DISCUSSION
Social media networks have introduced a new era for the public to share ideas, opinions, and
debates. These shared data establish a significant source to detect opinion trends regarding
different topics. In Parallel, different methods are becoming efficient in analyzing attitudes of
complex data in social media. For example, artificial intelligence, sentiment analysis, data
mining, and machine learning would help by providing different means of information extraction
and knowledge acquisition regarding different topics. With people's tendency to exchange and
broadcast their opinions on social media platforms, and with the emergence of the COVID-19
pandemic, many legislative and health authorities have found themselves in the face of
pandemic-related rumors. These rumors affect and slow down the efforts to combat the
pandemic, as it consumes much time and effort to refute them. For that scenario, researchers and
decision-makers can benefit from this data in analyzing and studying several social, behavioral,
or organizational phenomena. Unlike the traditional knowledge management practices that rely
on human experts as a primary source of knowledge acquisition, harvesting and analyzing social
media data with the help of machine learning could represent the ultimate source from which the
knowledge is extracted. Harvesting and mining the COVID-19 Twitter data could provide

39
innovative insights and ideas on how to better manage the implications and apprehensions
associated with pandemic-related rumors. Such studies could provide some roadmap for potential
future alternative sources of knowledge and suggest future machine learning and system
involvement. In addition, media practitioners can benefit from such affordable technical solutions
to automate the process of refuting rumors and misinformation.
Figure 4. False tweets word clouds Figure 5. True tweets word clouds
In terms of results accuracy, the experiment shows that combining machine learning techniques
with semantic analysis can increase accuracy in detecting rumors in a dataset and avoid claims
that may false classify tweets. Text analytics could be enhanced using data visualization
techniques such as words cloud. Though a simple concept, word clouds are a good visualization
technique to communicate an overall picture of text contents [20]. A study [21] concluded that
word clouds are certainly an effective tool for text analysis if enhanced with further information
and powerful interaction techniques. Figure 4 and Figure 5 display word clouds generated from
the dataset [8] based on its class label. Yet, there are some limitations of this work. First, the
focus of this work was on Twitter content only due to the availability of the annotated dataset.
However, Twitter is not the only source that misinformation can spread. Second, the dataset is
relatively small, which impacts the splitting decision of the training and testing sets. Thus, the
proper and sufficient learning dataset of the system with more data can increase the system's
accuracy as a whole.
7. CONCLUSION
Analyzing Twitter news data using natural language processing, sentimental analysis, machine
learning, and artificial intelligence techniques could serve as critical tools to help protect people
from misinformation. With millions of contents posted every day on social media platforms,
relying on human fact-checkers to review stories on an individual basis is almost impossible.
Alternatively, there is an urgent need for a new way to quickly detect posts that may contain false
claims and automatically send them to fact-checkers to focus their time and expertise on fact-
checking new content. English text analysis had had a great deal of work on detecting rumors.
However, in Arabic, this field is still in an early stage where the lack of resources and language
nature confront huge challenges to the research. This study presents an Arabic misinformation
detection model that mines related Arabic texts from Twitter to detect fake news. For this study,
we preprocessed a dataset of 3.6k tweets related to Covid-19 claims. This dataset was then fed
into different machine learning classifiers to measure classifier accuracy in detecting
misinformation. The results showed that the best overall accuracy of the model in detecting
misinformation was 93%. Another dataset of Covid-19 related claims in Arabic was built to
examine the performance of our model. The results showed that the model's overall accuracy in
detecting the veracity of the new dataset was 87% at best. The results provided by the classifiers

40
were analyzed further by a semantic analysis method to detect neutral sentences which can false
classification of tweets. The results show that this post semantic analysis phase has improved
accuracy of the results of the top result in the first phase to 92%.
REFERENCES
[1] Radcliffe and H. Abuhmaid, “Social media in the middle east: 2019 in review,” SSRN Electron. J.,
2020.
[2] ExtraDigital Ltd, www. extradigital. co. uk, “Prominent Arabic Social Media,” Extradigital.co.uk.
[Online]. Available: https://www.extradigital.co.uk/articles/arabic/social-media.html. [Accessed: 02-
May-2021].
[3] “Twitter: most-used languages 2013,” Statista.com. [Online]. Available:
https://www.statista.com/statistics/267129/most-used-languages-on-twitter/. [Accessed: 02-May-
2021].
[4] Alam et al., “Fighting the COVID-19 infodemic: Modeling the perspective of journalists, fact-
checkers, social media platforms, policy makers, and the society,” arXiv [cs.CL], 2020.
[5] P. Patwa et al., “Fighting an Infodemic: COVID-19 Fake News Dataset,” arXiv [cs.CL], 2020.
[6] S. M. Alzanin and A. M. Azmi, “Rumor detection in Arabic tweets using semi-supervised and
unsupervised expectation–maximization,” Knowl. Based Syst., vol. 185, no. 104945, p. 104945, 2019.
[7] M. Alkhair, K. Meftouh, K. Smaïli, and N. Othman, “An Arabic corpus of fake news: Collection,
analysis and classification,” in Communications in Computer and Information Science, Cham:
Springer International Publishing, 2019, pp. 292–302.
[8] Haouari, M. Hasanain, R. Suwaileh, and T. Elsayed, “ArCOV19-Rumors: Arabic COVID-19 Twitter
dataset for misinformation detection,” arXiv [cs.CL], 2020.
[9] M. J. Lazer et al., “The science of fake news,” Science, vol. 359, no. 6380, pp. 1094–1096, 2018.
[10] R. Vaithianathan, N. Jiang, T. Maloney, P. Nand, and E. Putnam-Hornstein, Developing predictive
risk models to support child maltreatment hotline screening decisions: Allegheny County
methodology and implementation [PDF. Auckland: Centre for Social Data Analytics, 2017.
[11] “Working to stop misinformation and false news,” Facebook.com. [Online]. Available:
https://www.facebook.com/formedia/blog/working-to-stop-misinformation-and-false-news.
[Accessed: 02-May-2021].
[12] Documenting the Now, Hydrator [Computer Software]. Retrieved from
https://github.com/docnow/hydrator. Accessed March, 2021.
[13] S. Larabi Marie-Sainte, N. Alalyani, S. Alotaibi, S. Ghouzali, and I. Abunadi, “Arabic natural
language processing and machine learning-based systems,” IEEE Access, vol. 7, pp. 7011–7020,
2019.
[14] S. Bird, E. Klein, and E. Loper, Natural language processing with python: Analyzing text with the
natural language toolkit. O’Reilly Media, 2009.
[15] Pedregosa et al., “Scikit-learn: Machine Learning in Python,” arXiv [cs.LG], 2012.
[16] I. Webb et al., “Logistic Regression,” in Encyclopedia of Machine Learning, Boston, MA: Springer
US, 2011, pp. 631–631.
[17] M. D. Buhmann et al., “Random Forests,” in Encyclopedia of Machine Learning, Boston, MA:
Springer US, 2011, pp. 828–828.
[18] I. Webb, E. Keogh, R. Miikkulainen, R. Miikkulainen, and M. Sebag, “Naïve Bayes,” in
Encyclopedia of Machine Learning, Boston, MA: Springer US, 2011, pp. 713–714.
[19] Cortes and V. Vapnik, “Support-vector networks,” Mach. Learn., vol. 20, no. 3, pp. 273–297, 1995.
[20] B. Y.-L. Kuo, T. Hentrich, B. M. Good, and M. D. Wilkinson, “Tag clouds for summarizing web
search results,” in Proceedings of the 16th international conference on World Wide Web - WWW ’07,
2007.
[21] F. Heimerl, S. Lohmann, S. Lange, and T. Ertl, “Word cloud explorer: Text analytics based on word
clouds,” in 47th Hawaii International Conference on System Sciences, 2014.
[22] J. Berri, M. Al-Khamis, Information Exploration Using Mobile Agents, WSEAS Transactions on
Computers, vol. 3, no. 3, 706-712, 2004.
[23] J. Berri, R. Benlamri, Y Atif, H. Khallouki, Web Hypermedia Resources Reuse and Integration for
On-Demand M-Learning, International Journal of Computer Science and Network Security, vol. 21,
no. 1, pp. 125-136, 2021

COMBINING MACHINE LEARNING AND SEMANTIC ANALYSIS FOR EFFICIENT MISINFORMATION DETECTION OF ARABIC COVID-19 TWEETS

More Related Content

Similar to COMBINING MACHINE LEARNING AND SEMANTIC ANALYSIS FOR EFFICIENT MISINFORMATION DETECTION OF ARABIC COVID-19 TWEETS (20)

Recently uploaded (20)

COMBINING MACHINE LEARNING AND SEMANTIC ANALYSIS FOR EFFICIENT MISINFORMATION DETECTION OF ARABIC COVID-19 TWEETS