Fake news detection using discourse segment

structure analysis
Anmol Uppal Vipul Sachdeva Seema Sharma
Amity School of Engineering & Amity School of Engineering & Amity School of Engineering &
Technology, Amity University Technology, Amity University Technology, Amity University
Noida, Uttar Pradesh, INDIA Noida, Uttar Pradesh, INDIA Noida, Uttar Pradesh, INDIA
[email protected] [email protected] [email protected]

Abstract— Online news platforms greatly influence our society The rest of the paper is structured as follows: in section II,
and culture in both positive and negative ways. As online media background study is carried out on the broad spectrum of fake
becomes more dependent for source of information, a lot of fake news detection approaches. In section III, literature survey is
news is posted online, that widespread with people following it done on most of the latest work done for detecting fake news.
without any prior or complete information of event authenticity. Section IV gives a formal definition of the problem statement
Such misinformation has the potential to manipulate public for proposed methodology and deep leaning approach for
opinions. The exponential growth of fake news propagation have structure parsing. The results obtained from the experimental
become a great threat to public for news trustworthiness. It has setup is explained in section V and section VI gives future
become a compelling issue for which discovering, examining and
insights to this research.
dealing with fake news has increased in demand. However, with
the limited availability of literature on the issue of uncovering fake II. BACKGROUND STUDY
news, a number of potential methodologies and techniques
remains unexplored. The primary aim of this paper is to review There has not been a specific definition of fake news since a
existing methodologies, to propose and implement a method for news is considered fake depending on the way it is perceived.
automated deception detection. The proposed methodology uses There are different categories into which a news article may fall:
deep learning in discourse-level structure analysis to formulate the
Table 1: Categories of Fake news [2]
structure that differentiates fake and real news. The baseline
model achieved 74% accuracy.
Authenticity Intention News?
Keywords— fake news detection, news content model; social context Maliciously false
model, discouse parsing fake news detection News FALSE Bad Yes
Fake News FALSE Unknown Yes
Satire News Unknown Not Bad Yes
I. INTRODUCTION Disinformation FALSE Bad Unknown
Social media platforms greatly influence our society and culture Misinformation FALSE Unknown Unknown
in both positive and negative ways [1]. As online media becomes Rumor Unknown Unknown Unknown
more dependent for source of information, we tend to get a lot
of fake news that widespread with people following it without Fake news can be simply termed as false news. It can be
any prior or complete information of the event authenticity. The considered as deliberate misinformation that spreads through the
increasing growth of fake news propagation has become a great traditional or online news media in order to mislead public,
threat to public trust for news trustworthiness. So the need of cause damage of any kind or for financial gains. Fake news
discovering, examining and dealing with fake news has generally has two main features: intent and authenticity that can
increased in demand. However, it faces many challenges such as classify categories of fake news as shown in table 1.
limited literature and dataset availability. This paper proposes to
inculcate a method for detecting fake news centered on the idea Fake news detection is done based on the perspective of fake
of a news articles’ structural representation. The main idea is to news via main models:
detect fake news by justifying some structure-related properties 1) News Content Model
that explains the structure formulated for different type of news. a) Knowledge based detection - The informative content
Most of the existing methods rely heavily upon existence of a
presented in the news article
detailed corpus to segregate fake and real news, which is very
difficult to define for fake news. b) Style based detection - Writing style and linguistic
pattern in the content
Existing methodologies might be biased to underestimate the
2) Social Context Model
characteristics of the way in which content is presented in a news
article. Use of discourse-level structure analysis for style-based a) The propagation pattern of the news
detection would help in automatic fake news detection. b) The credibility of the creators and the propagators of
the news

c
978-1-7281-2791-0/20/$31.00 2020 IEEE 751
 B. Style based detection
In this technique, a news is denoted as a set of machine learning
features that are able to identify fake news and separate it from
real news. It aims to focus on the issue that a fake news article is
deceptive and this deception needs to be detected in order to
separate it from real news. It captures the disorientations in the
writing style in the content of the news article. This form of
detection methodology can be categorized into objectivity
oriented and deception oriented.
Approaches related to capture the objective orientation of a news
article indicate the non-neutrality and partial opinion of an event
to mislead the readers such as biased writings and yellow
journalism. Biased writings present a greatly favored opinion
towards a particular group, which can be detected using
linguistic features that are even able to differentiate satire from
mainstream as well as from hyperpartisan [5]. Yellow
journalism highlights display of not so well examined news but
acts as clickbait with eye catchy headlines to raise curiosity and
Figure 1: Fake news detection perspectives in its news life cycle [2] hostility, a sudden urge to make sensitive readers anxious and
magnify their fear of danger to their beliefs. Content style that
aims to deceive the readers with exaggerated expressions and
Fake news can be detected via the content presented within the strong annotations is different from that of a truly factual news.
article or by the social media metadata it carries when it is
published and propagated online. Multiple techniques can be
applied to any of the above-mentioned models for detection.
These models are further explained below. C. Social Context Model for fake news detection
The networking and sharing features of social media platforms
give extra information like social engagements in analysis thus
A. Knowledge based detection making fake news detection perceivable a multiple different
angles. Again, social engagements that can deliver the metadata
This style of fake news detection sounds mostly manual and about reception of the news by the readers such as comments,
deals with true fact checking of the news conveyed in the article. can be identified by majorly two methods propagation based and
The primary purpose of knowledge based detection is fact- credibility based.
proving in two ways – traditional fact checking and automated
fact checking. This aims for assessing the authenticity of news Propagation based techniques aim to find out the relations
by comparing the knowledge extracted from news article to be between several news propagated on the social media and track
verified with actual events. it to the original news. Credibility based techniques use
relationships between news articles and other components such
Manual or traditional fact checking can be done via two sources as users, publishers and posts. For example, news published by
by expert checking and via crowd sourced fact checking. Expert unreliable websites or forwarded by unreliable users has a very
checks are mostly done via small group of trustworthy fact high chance to be fake than the news originating from respected
checkers such as PolitiFact [3], Factcheck [4] and many others and credible users. Fake news can be identified using credibility
that give accurate results. These are costly and non -scalable for for the following components of a propagated news: headlines,
handling great volume of incoming news traffic. On the other comments, source and propagators.
hand, crowd sourced fact checking relies on participation of
large number of individuals acting as fact checkers. Crowd
sourced fact checking happens to be more scalable than expert
fact checking but it is less credible and accurate due to bias of III. LITERATURE SURVEY
fact checkers to one side or another. Table 2 below highlights the recent journal articles, conference
Automated fact checking has the ability to deal with huge proceedings and research papers published relating to various
volume of to be verified news and relies heavily on information methodologies for fake news detection. The search results are
retrieval, natural language processing and to some extent on restricted to be published from or after the year 2016. The table
network theory as well. Automated fact checking is divided into summarizes the technique adopted and its performance measure
two steps knowledge base construction and knowledge in success of distinguishing fake news from real news for every
comparison. Here, knowledge means representation of paper. Most of these methods are related to use of sentiment
information in a set of attributes extracted from the knowledge. analysis and NLP while others consider some other features like
discourse and pragmatic analysis of the text. It was observed that
the success of the classification was also dependent on the
machine learning model chosen/formulated and trained for the
purpose.

752 10th International Conference on Cloud Computing, Data Science & Engineering (Confluence)
 Table 2: Literature Survey, papers taken after 2016 Agarwalla Accuracy in Naïve An algorithm have been
et al. ,2019 Bayes classifier explored that can
CONCLUSION/ [15] with lid stone distinguish the difference
AUTHOR OUTCOME
TECHNIQUE USED
smoothing is 83% between the fake and true
Monu GRU-2 can Rational investigation of and in Naïve news
Waskale, identify bits of large portion of existing
Bayes (without lid
Prof. gossip with techniques identified
exactness 83.9% from the earlier bits of stone smoothing)
Pritesh is 74%.
for twitter within gossip using GRU-2 and
Jain, 2019 12-hours. Tanh-RNN classifier for Zellers et Grover-Large Investigated the threats
[6] comparison. al. ,2019 yields 78% posed by adversaries
Cody 66.93% and Automated system for [16] accuracy, 92% seeking to spread
Buntain, 70.28% in distinguishing fake news when dataset is disinformation and the
Jennifer PHEME and in famous Twitter increased. possibilities of machine
Golbeck, CREDBANK threads. generated fake news.
2018 [7] Sivasangar Precision: 0.86 Rumor Detection by lever
Kursuncu Analysis resulting Discussed Sentiment i V et al. F1-score: 0.86 maturing the setting
et al. , 2018 in 73% F score. predictive analysis by ,2018 [17] going before a tweet with
[8] Supervised time series in different a consecutive classifier.
learning F score of domains. O'Brien et Accuracy: 93.5% Deep neural networks to
88.2 and ROC of al. ,2018 ± 0.2. capture steady
94.3. [18] differences in the
Jiawei Comparing with A deep diffusive network language of fake and real
Zhang, existing systems, model to learn the news: signatures of
Bowen Bi-Class Inference depictions of news exaggeration and other
Dong, Results in 14.5% articles, subjects and forms of rhetoric.
Philip S. higher and 40% creators. Shu et al. Studying existing Datasets, evaluation
Yu, 2019 higher in multi- ,2017 [19] literature in two metrics, and promising
[9] Class Inference. segments: future ways in fake news
Ruchansky Positive Propose the CSI model detection and detection discussed.
et al. , 2017 correlation of which consists of three characterization.
[10] 0.867 for Weibo, modules. Silva et al. ANN achieving An amalgamation of
0.631 for Twitter. ,2019 [20] 75% accuracy. classic techniques with
Monther 99.4% accuracy Clickbait Detection using neural network.
Aldwairi, Logistic Classifier
Ali Dong et al. Detect fake news Deep semi-supervised
Alwahedi ,2019 [21] from PHEME learning model by
,2018 [11] datasets using constructing two-path
Traylor et Classifier ML-classification for labeled data and CNN
al. ,2019 Accuracy: 0.69 large fake news unlabeled data.
[12] Overall Classifier documents with one Yang et al. 88.063% accuracy Soft labels are used to
Error:0.31. extraction feature. ,2019 [22] fine-tune NLI models,
Oshikawa Maximum Comparison of existing BERT, and the
et al. ,2018 accuracy in 94.4% fake news detection Decomposable Attention
[13] using GCN model. models model. NLI models are
Gurav et Series of methods Pure NLP perspective trained independently
al. ,2019 accomplished by towards false news and ensemble with a
[14] Machine Leaning detection BERT model to define
the soft labels.

10th International Conference on Cloud Computing, Data Science & Engineering (Confluence) 753
 Monti et al. More than 93% .The underlying core of the user. 75.9% from social media users..
,2019 [23] accuracy achieved algorithms allow for a max. Accuracy Reality and credibility is
by automatic fake fusion of dissimilar data accomplished on considered by an efficient
news detection such as content, profile, LIAR dataset. Gibbs-sampling method.
model built on activity of a user, social Zhou et al. Models so far Simply looking into
geometric deep graph and propagation ,2018 [30] possess a greater Linguistic aspects is not
learning. pattern of news, which is possibility to enough for fake news
achieved by generalizing misclassify fake detection.
CNN to graphs. news that tampers
Ajao et al. 82% accuracy via A hybrid deep learning with facts as well
,2018 [24] both text and model of LSTM and as under-written
images by CNN models is used. real news articles.
automatic Álvaro and 93% accuracy BERT, LSTM and
identification of Lara,2019 with superior Convolutional Neural
features within [31] metrics compared Network models are
Twitter posts. to other deep trained based merely on
Thota et al. 94.21% accuracy. A finely tuned TF-IDF learning models textual features.
,2018 [25] The Dense Neural Dense neural network
Network beats architecture to predict the
existing model stance between a given IV. PROPOSED METHODOLOGY
architectures by pair of headline and Discourse structure is the connection between surface features
2.5%. article body. and document-level properties, which remains unseen and
underestimated most of the times [32].This paper proposes a
Helmstette F1 score of 0.9 Trustworthy or plan to make use of discourse level analysis for deception
r, Heiko achieved untrustworthy source are detection of news documents to make a hierarchical structure
Paulheim used to automatically that will have distinguishable properties associated with fake
,2018 [26] label the data during and real news articles. These properties can be used to
differentiate between the two.
collection, and train a
classifier on this dataset. Given a news article β consisting of various creator features that
Atodiresei Tweet score can Credibility score from describe the content features containing a set of attributes that
represent the news item, the job of fake news detector is to
et al. ,2018 be 1000, -500 or in hashtag sentiments, emoji
predict whether the news article β is fake or not, i.e.,
[27] [-50,100] sentiments, text
User score can be sentiments and named- τ(β) = {1, if β is fake news article,
0, elsewise} … (1)
in [0,12] entity recognition. Higher
the credibility score, Where τ is the representation of the prediction function in
higher the trust Eq.(1). The model would take news event by means of an
Hamid 82% accuracy Hierarchical Discourse input and based on the sequence of information provided in
Karimi, level structural data the news article predicts news being fake or real.
Jiliang analysis for fake news
Tang ,2019 detection. A structure is
[28] trained on the dataset,
quantifiable properties of
which are used for
classification process in
the model.
Shuo Yang Graphical model Unsupervised method is
et al., 2019 is built taking into investigated. Opinion is
[29] account reliability extracted from hierarchy
of the news and social engagement
credibility score information acquired Figure 2: Framework for Fake news detection using document structural
analysis

754 10th International Conference on Cloud Computing, Data Science & Engineering (Confluence)
Evaluation of properties learnt from classification and analysis, fake news articles. The training error and step wise accuracy is
will lead to describing news existing to be fake or real. Discourse shown in Figure 3(a) and 3(b) respectively.
level structure parsing of the news article is preferred for better
understanding of how fake news articles are structures since the 1
meaning of a sentence depends on the sentence that precedes it 0.8
and may influence the meaning of the sentence succeeding it.
0.6

Loss
For this, discourse dependency parsing approach is used to
create a dependency tree structure similar to (Liu and Lapata, 0.4
2018 [33]). 0.2
Forward and backward Gated Recurrent Unit (GRU) 0
(Kyunghyun Cho et al, 2014 [34]) is used to create Bidirectional 0 50 100 150 200 250 300
GRU and get the representation of a sentential clause in a news Steps
document. Next, making use of inter-sentential matrix to get
discourse level relationships in automated manner, a Figure 3(a): Training error
dependency tree is completed (Karimi and Tang, 2019 [28]).
This dependency tree is going to be the structural representation 80
of the news document (Liu and Lapata, 2018 [33]) It is done by
computing parent child properties of each sentence using 70

Accuracy (%)
activation function.
60
The classifier model attained after achieving document level
representation of the news document can be trained for binary 50
classification using Eq.(2) and Eq.(3):
40
χi =í (yilog(pif) + (1íyi)log(pir)) … (2) 0 100 200 300
Steps
Λ(Θ) =™‫׊‬ni‫א‬N (χi) …(3) Figure 3(b): Accuracy with step learning

χi is the cross entropy loss, yi represents the ground label for a The structural analysis on the three general attributes obtained
text, pif and pir are the probabilities of a text to be fake or real from particular tree data structure supports our argument on the
obtained from structural level document representation.ni associated properties with the contextual analysis of news
represent the news article in the corpus N. Λ is the total loss documents. (Karimi and Tang, 2019 [28]) have also observed
some of these properties in their model. The below-mentioned
properties have been attained as a normalized sum over the
V. EXPERIMENT length of inter-sentential matrix (number of sentences in an
article):-
In the experiment, the aim is to find out the performance of
proposed methodology on fake and real news classification, Preorder traversal difference: - The difference between
and the degree of association of the structure related properties preorder traversal of a discourse dependency tree and its normal
to their respective fake or real class news documents. sequential order.
Dataset: Thee following five datasets combined is collected and
combined in order to avoid any biased results towards any 50 400
400 41.57 330.91
particular kind of news limited to a certain climate. The datasets 330.72
40 300
are collected by kaggle.com, (Khan et al.,2019 [35]) and (Shu et 300
al., 2017 [36]) which includes articles from Buzzfeed and 30 27.62
200 200 159.54
politifact [3]. 153.09 20
Simulation: The dataset gathered was already preprocessed 100 10 100
with all non-English characters and stop words removed and
converted to lowercase. With 6398 data documents for training, 0 0 0
188 were reserved for testing on unseen data. We used Google’s Fake Real Fake Real Fake Real
pre trained word2vec [37] embedding and ELU as the nonlinear
activation function. The resulting and final simulation is run for Figure 4: Avg. preorder positional diff, Avg. postorder positional diff
300 steps with mini-batch size of 50 documents with initial and Avg. parent-child distance respectively.
learning rate of 0.01, dropout rate to 0.0 and L2 regularization
set to 0.01 .The hidden unit size was set to 100. ADAM Postorder traversal difference: - The difference between
optimizer [38] was used for hyperparameter tuning to implement postorder traversal of a discourse dependency tree and its normal
our method. We performed our simulations on Google Colab sequential order.
GPU hardware accelerator. Distance between a parent and a child node: - The distance
Results: The proposed model was able to attain 74.62% difference between child and parent unit in a discourse
accuracy on unseen data with F1-score of 0.76 on detecting dependency tree in their original sequential order.

10th International Conference on Cloud Computing, Data Science & Engineering (Confluence) 755
 VI. CONCLUSION AND FUTURE WORK [16] A. H. H. R. Y. B. F. F. R. Y. C. Rowan Zellers, "Defending Against
Neural Fake News," arXiv, 2019.
Automatic fake news detection is a thought-provoking problem
[17] V. A. P. S. R. Sivasangari V, "A Modern approach to identify the fake
in deception detection, and it has incredible real-world political news using machine learning," International Journal of Pure and
& social influences. In this paper, Deep Learning (GRU) was Applied Mathematics, vol. 118, no. 20, 2018.
used for discourse segment analysis and constructing the [18] N. L. S. E. G. B. X. O'Brien, "The Language of Fake News: Opening the
dependency tree that offered distinguishable features for real and Black-Box of Deep Learning Based Detectors," in "AI for Social Good",
fake news. The model achieved 0.76 F1 score and 74.62% NIPS 2018, Montreal, Canada, 2018.
accuracy. This paper can be carried out further to various other [19] A. S. S. W. J. T. H. L. Kai Shu, "Fake News Detection on Social Media:
directions. There can be many other structures like dependency- A Data Mining Perspective," ACM SIGKDD Explorations Newsletter,
trees with more advanced structure related properties that can be pp. 22-36, 2017.
considered. In addition, using ensemble learning and other [20] R. V. d. C. A. A. C. B. G. Fernando Cardoso Durier da Silva, "Can
structure breeding techniques is also an area of research in order Machines Learn to Detect Fake News? A Survey Focused on Social
Media," in 52nd Hawaii International Conference on System Sciences,
to improve the accuracy of our presented model. This model can
Grand Wailea, Maui, 2019.
be further developed and applied for real time detection using
[21] U. V. S. C. L. Q. Xishuang Dong, "Deep Two-path Semi-supervised
Intra document and Inter document linking such as relation Learning for Fake News Detection," arXiv, 2019.
between events, participants, credibility of sources and positive
[22] T. N. H.-Y. K. Kai-Chou Yang, "Fake News Detection as Natural
responders. At last, testing the proposed model on datasets Language Inference," arXiv, 2019.
limited to certain period of news climate may also offer some
[23] F. F. D. E. D. M. M. M. B. Federico Monti, "Fake News Detection on
interesting insights. This refers to training and testing Social Media using Geometric Deep learning," arXiv, 2019.
political/economic/miscellaneous news only considered over
[24] D. B. S. Z. Oluwaseun Ajao, "Fake News Identification on Twitter with
restricted period and region. Hybrid CNN and RNN Models," in SMSociety '18 Proceedings of the 9th
International Conference on Social Media and Society , Copenhagen,
REFERENCES Denmark, 2018.
[1] M. G. Hunt Allcott, "Social media and Fake News in the 2016 Election," [25] P. T. S. A. N. L. Aswini Thota, "Fake News Detection: A Deep Learning
Journal of Economic Perspectives, vol. 31, no. 2, pp. 211-236, 2017. Approach," SMU Data Science Review, 2018.
[2] R. Z. Xinyi Zhou, "Fake News: A Survey of Research, Detection [26] H. P. Stefan Helmstetter, "Weakly Supervised Learning for Fake News,"
Methods, and Opportunities," arXiv:1812.00315v1. in IEEE/ACM International Conference on Advances in Social Networks
Analysis and Mining (ASONAM), Barcelona, Spain, 2018.
[3] "Fact-checking U.S. politics | PolitiFact," Poynter Institute, [Online].
Available: https://www.politifact.com/. [Accessed 2019]. [27] A. A. I. Costel-Sergiu Atodiresei, "Identifying Fake News and Fake
Users on Twitter," Procedia Computer Science, pp. 451-461, 2018.
[4] "FACTCHECK.ORG," The Annenberg Public Policy Center of the
University of Pennsylvania, [Online]. Available: [28] J. T. Hamid Karimi, "Learning Hierarchical Discourse-level Structure for
https://www.factcheck.org/. Fake News Detection," arXiv, 2019.
[5] J. K. K. R. J. B. B. S. Martin Potthast, "A Stylometric Inquiry into [29] K. S. S. W. R. G. F. W. H. L. Shuo Yang, "Unsupervised Fake News
Hyperpartisan and Fake News," arXiv, 2017. Detection on Social Media: A Generative Approach," in AAAI
Conference on Artificial Intelligence, Honolulu, Hawaii, 2019.
[6] P. P. J. Monu Waskale, "A Review Rumors Detection on Twitter Using
Machine Learning Techniques," International Journal of Scientific [30] H. G. M. M. B. J. H. Zhixuan Zhou, "Fake News Detection via NLP is
Research & Engineering Trends, vol. 05, no. 03, 2019. Vulnerable to Adversarial Attacks," arXiv, 2019.
[7] J. G. Cody Buntain, "Automatically Identifying Fake News in Popular [31] L. L. I. Álvaro Ibrain Rodríguez, "Fake news detection using Deep
Twitter Threads," arXiv, 2018. Learning," arXiv:1910.03496v2, 2019.
[8] M. G. U. L. K. T. Ugur Kursuncu, "Predictive Analysis on Twitter: [32] E. F. M. S. Joan Bachenko, "Verification and implementation of
Techniques and Applications," arXiv, 2018. language-based deception indicators in civil and criminal narratives," in
COLING '08 Proceedings of the 22nd International Conference on
[9] B. D. P. S. Y. Jiawei Zhang, "FAKEDETECTOR: Effective Fake News
Computational Linguistics, Manchester, UK, 2008.
Detection," arXiv, 2019.
[33] M. L. Yang Liu, "Learning Structured Text Representations,"
[10] S. S. Y. L. Natali Ruchansky, "CSI: A Hybrid Deep Model for Fake
Transactions of the Association for Computational Linguistics, vol. 6,
News Detection," in CIKM 2017 Internationall Conference on
pp. 63-75, 2018.
Information and Knowledge mangaement, Singapore, 2017.
[34] B. v. M. C. G. D. B. F. B. H. S. Y. B. Kyunghyun Cho, "Learning Phrase
[11] A. A. Monther Aldwairi, "Detecting Fake News in Social Media
Representations using RNN Encoder-Decoder for Statistical Machine
Networks," in The 9th International Conference on Emerging Ubiquitous
Translation," in Proceedings of the 2014 Conference on Empirical
Systems and Pervasive Networks, Leuven, Belgium, 2018.
Methods in Natural Language Processing (EMNL, Doha, Qatar, 2014.
[12] J. S. G. N. S. Terry Traylor, "Classifying Fake News Articles Using
[35] M. T. I. K. A. I. S. A. Junaed Younus Khan, "A Benchmark Study on
Natural Language Processing to Identify In-Article Attribution as a
Machine Learning Methods for Fake News Detection,"
Supervised Learning Estimator," in IEEE 13th International Conference
arXiv:1905.04749v1, 2019.
on Semantic Computing, Newport Beach, California, 2019.
[36] S. W. a. H. L. Kai Shu, "Ex-ploiting tri-relationship for fake news
[13] J. Q. W. Y. Ray Oshikawa, "A Survey on Natural Language Processing
detection.," arXiv preprint arXiv:1712.07709, 2017.
for Fake News Detection," arXiv, 2018.
[37] K. C. G. C. J. D. Tomas Mikolov, "Efficient Estimation of Word
[14] S. S. S. S. P. W. S. H. Subhadra Gurav, "Survey on Automated System
Representations in Vector Space," arXiv preprintarXiv:1301.3781, 2013.
for Fake News Detection using NLP & Machine Learning Approach,"
International Research Journal of Engineering and Technology, vol. 06 [38] J. B. Diederik P Kingma, " Adam: A method for stochastic optimization,"
Issue, no. 01, 2019. arXivpreprint arXiv:1412.6980, 2014.
[15] S. N. V. A. N. D. D. H. Kushal Agarwalla, "Fake News Detection using
Machine Learning and Natural Language Processing," International
Journal of Recent Technology and Engineering, vol. 07, no. 06, 2019.

756 10th International Conference on Cloud Computing, Data Science & Engineering (Confluence)