Sieving Fake News From Genuine: A Synopsis

Shahid Alam Abdulaziz Ravshanbekov

Department of Computer Engineering Department of Computer Engineering
Adana Alparsalan Turkes Science and Technology Adana Alparsalan Turkes Science and Technology
University, Adana, Turkey University, Adana, Turkey
[email protected] [email protected]

Abstract—With the rise of social media, it has become easier II. FAKE N EWS
to disseminate fake news faster and cheaper, compared to tradi-
arXiv:1911.08516v1 [cs.CR] 19 Nov 2019

tional news media, such as television and newspapers. Recently Before the invention of paper in China, rumors (fake news)
this phenomenon has attracted lot of public attention, because used to spread by word of mouth, and were mostly confined
it is causing significant social and financial impacts on their to a local community. The invention of paper made the
lives and businesses. Fake news are responsible for creating
false, deceptive, misleading, and suspicious information that can dissemination of news easier. But still it used to take time
greatly effect the outcome of an event. This paper presents a for a news to reach/spread the far corners of a state/country.
synopsis that explains what are fake news with examples and The emergence of Internet, social media and smart phones
also discusses some of the current machine learning techniques, have revolutionized this process, and now almost everyone can
specifically natural language processing (NLP) and deep learning, claim that I have news in my pockets. This phenomenon is very
for automatically predicting and detecting fake news. Based on
this synopsis, we recommend that there is a potential of using useful in making people more aware and knowledgeable of the
NLP and deep learning to improve automatic detection of fake events and their surroundings, but presents new challenges of
news, but with the right set of data and features. sieving fake news from genuine out of this large amount of
Index Terms—Fake news, Automatic fake news detection, information/data in our pockets.
Machine learning, Natural language processing, Deep learning.
Fake news spread mostly through online media, in the form
of text, images and videos. Fake news are false, and there
I. I NTRODUCTION purpose is to create deceptions and mislead people, so that it
Fake news is not a new phenomenon, but recently it has changes the outcome of an event. The events can be sending
attracted more public attention. The rise of social media, someone to jail, or manipulating the result of elections, etc.
easier access, and faster and cheaper online dissemination of In this paper we only focus on fake news that are in the form
fake news compared to other traditional news media, such as of text. We formally define fake news as follows.
newspapers and television, makes it particularly relevant in
this new age of information. DEFINITION: Fake news is any text data disseminated
Conroy et al. [1] divides the approaches to detecting de- through online media, such as newspapers, websites, social
ception into two major categories, linguistic and network networks, etc., that creates false, deceptive, misleading, and
approach. In the linguistic approach researchers have used: suspicious information which can significantly effect the
analysis of n-grams of words (i.e., bag of words); syntax and outcome of an event.
semantic analysis of the text; and structural analysis of text
to find incoherency between deceptive and truthful messages. Some recent examples of fake news:
In the network approach researchers have used: knowledge
networks for fact-checking; and social network behavior, such (1) The government of Japan announced that it was banning
as using the metadata and telltale behavior of questionable the use of microwave ovens in the country by 2020 [2]:
sources. This fake news was originated on a Russian website about
There are basically two major techniques manual and auto- Japan abandoning the use of microwave ovens by 2020.
matic detection of fake news. Manual detection is carried out The complete news was in Russian and when translated
by experts in the field or by crowds (a large number of regular via Google it stated about banning microwave ovens and
people acting as fact checkers). Automatic detection relies on prisoning people not fulfilling the requirement. It caused panic
a combination of information retrieval and machine learning in the mind of people. This news was verified to be fake by
techniques. Manual detection is more reliable but not scalable. several fact checking web sites, such as by chek4spam.com
Automatic detection is less reliable but scalable. In this paper shown in Figure 1, as it was published by a satire website
we only discuss automatic detection of fake news. panorama.pub. Fact checking web sites either use experts in
In the next Sections we explain what are fake news with the specific domain or use crowd sourcing for detecting fake
examples and present and discuss the current machine learning news. This is an example, showing manual detection of fake
techniques for automatically detecting fake news. news.
 • The ability to learn feature representations rather than
(2) President Trump to ban schools in order to stop school manually specifying and extracting features from the
shootings [3]: After a mass shooting at a school in Florida this natural language, and can continually improve.
fake news became viral on social media in February 2018. • Easily adapted to a new problem.
At the time of shooting, the U.S president Trump offered • Has shown great strength in processing text, speech and
number of ideas, such as arming teachers, and raising the image, all of them are excessively used in fake news.
minimum age to purchase gun, but never suggested banning Besides some of the advantages listed above, to produce
schools. This news was published on several satire websites good results deep learning requires massive data and time for
and as a meme 1 on me.me [4] as shown in Figure 2. training.
Due to the natural relation between NLP and fake news, and
(3) Two altar boys put marijuana in the censer-burner of strengths of deep learning listed above, recently [7], [8], [9],
a Cathedral in Spain [5]: This fake news went viral on [10], [11], [12], [13], [14], [15], [16], [17], [18], [19] there is
social media in January 2018 in Spain. More than 2 million an interest in applying these two techniques for detecting fake
people read the news. As a consequence, they were detained news.
overnight and released the other day without charge, but were
fired from the Cathedral as altar boys. III. S TATE OF THE A RT
Here we discuss some of the previous recent works on
(4) Barack Obama attended Columbia University as a predicting and detecting fake news using machine learning,
foreign student [6]: This fake news was created in 2012 U.S specifically NLP and deep learning.
presidential elections about Barrack Obama, that he attended Kai et al. [20] proposed a technique that exploits rela-
a college in U.S as a foreign student, to demonstrate that tionship among publishers, news pieces and users to predict
at some point in his life he was not a U.S citizen and is fake news. For modeling this tri-relationship they presented a
therefore ineligible to become the president of the United new framework TriFN that employs a linear classifier. They
States. assign each user a credibility score based on the user’s online
behavior. A user with a low credibility score is more likely
From the above examples, we can see that these fake news to share fake news. Their classifier outperforms the baseline
have three things in common: (1) they became viral on social classifiers, such as Decision Tree, NaiveBayes, and Random
media; (2) they were deceptive and misleading; (3) they had Forest etc.
the potential, and to some extent were able, to effect an event, Yang et al. [7] proposed a model for detection of fake
such as causing panic and outcome of an election, etc. news on social media through classifying news propagation
paths. They capture the global and local variations of user
A. Natural Language Processing and Deep Learning
characteristics to build a time series classifier. They compare
Natural language processing (NLP) is a branch of machine their classifier with a series of baseline fake news detection
learning that deals with processing a natural (human) language, classifiers, such as SVM, Decision Tree, and Random Forest
such as English, Spanish, French, and Chinese etc. NLP can etc. Their classifier combines both recurrent and convolutional
be used for extracting important part of a text and providing neural networks in one, and therefore shows better results than
an interpretation for that text. NLP is a hard problem, because the other compared classifiers in the paper [7].
unlike computer languages, natural languages are inherently Volkova et al. [8] presented a technique that classify suspi-
ambiguous. There is a strong and natural relation between NLP cious posts using linguistic and network features. They build
and fake news (as defined in Section II) processing. a neural network model based on state of the art techniques,
Deep learning is one of the fields of machine learning. such as recurrent and convolutional neural networks. They also
It involves learning from several layers. These layers consist build a logistic regression model for comparison. They got
of increasingly meaningful representations of input data. The better results with neural network model compared to logistic
depth of the model is the number of layers in the model. regression.
These layers of representations are learned through neural Siering et al. [21] addressed detecting deception on crowd-
networks. These networks map the input (such as an image) to funding platforms. They extract linguistic and content based
the target (such as cat) during a deep sequence of simple data features related to the different types of communication on
transformations (layers). A neural network is parametrized by these platforms. They examine both static and dynamic com-
its weights. A loss score is used as a feedback signal to adjust munications. Dynamic communication is carried out in real-
the weights. After several iterations the loss is minimized time, i.e., it is analogous to face-to-face communication, and
and the output is produced which is as close to the target static communication is carried out in delayed mode. They also
as possible. used different classifiers for testing their detection approach
There are some advantages of applying deep learning to and achieve good results with Support Vector Machine (SVM)
fake news detection such as: classifier.
1 Ideas, stories, phrases, etc., typically funny in nature, spread by Internet Afroz et al. [22] used stylometry to detect deception in
users, often with slight variations. online writing. They selected a large number of (736) different
 Fig. 1. Example showing manual detection of a fake news about microwaves banned in Japan.

techniques compared in the paper.

Hamid et al. [11] proposed a technique that use Convolu-
tional Neural Network (CNN) to extract features from a text
and then apply LSTM on these features to capture the temporal
dependency in the entire text. Accuracy achieved during the
experiments carried out in the paper is low.
Borges et al. [19] proposed a technique based on bi-
directional RNN for stance detection. First, they combine
representations inferred from three different inputs from the
text, including the headline, the first two sentences of the news
article, and the entire document. Then, a final layer processes
these results and predicts the stance of the news article.
Fig. 2. Example of a fake news about Donald Trump banning schools. psh
is an expression used when the writer disagrees with (urbandictionary.com).
A. Discussion
Table I gives a synopsis of the current works discussed in
features. The major type of features included lexical, syn- Section III that use machine learning techniques for automat-
tactic, content specific, grammar and vocabulary complexity, ically detecting fake news.
uncertainty, author’s attributes etc. For classification they used Out of the nine works examined, four combine NLP with
different classifiers and achieved good results with the SVM deep learning, two use only deep learning, and three use other
classifier. The results presented in the paper show that they classifiers for automatic detection of fake news. Accuracy of
were able to detect imitation attacks with 85% accuracy and detection varies from 38.8% – 96.6%. The most important
obfuscation attacks with 89.5% accuracy. factor among these research works that effects their accuracy
Hannah et al. [9] used NLP to compare language of real is the extraction and selection of features. Other factors are the
news with fake news, such as hoaxes and propaganda, to find type of dataset used for training and the ML method used to
linguistic characteristics of untrustworthy text. A ratio was build the model. Less than half (four) of these research works
found for each linguistic feature, and refers to how frequently achieve an accuracy > 90%. This shows that there is a need
it appeared in the fake news compared to the real news. to further this research for improving the automatic detection
They build a Long Short-Term Memory (LSTM) model and of fake news.
compared it with other Maximum Entropy and Naive Bayes
models. LSTM model achieved better results than the other IV. C ONCLUSION
two. The phenomenon news in my pockets is very useful in
Ruchansky et al. [10] proposed a model named CSI which making people aware and knowledgeable of the events and
classify an article based on temporal pattern and behavior their surroundings, but also presents new challenges of sieving
of a user activity on the given article. Recurrence Neural fake news from genuine out of this large amount of data
Network (RNN) was used to build the CSI model. Results in our pockets. In this paper, we have made an effort to
presented show CSI achieving a higher accuracy than the other present a synopsis that first explains what are fake news
 TABLE I
A SYNOPSIS OF THE CURRENT WORKS DISCUSSED IN S ECTION III THAT USE MACHINE LEARNING (ML) TECHNIQUES FOR AUTOMATIC DETECTION OF
FAKE NEWS .

Research work Features extraction ML methods used Dataset Accuracy

Tri-relationship among publishers, news pieces and users.
Kai et al. [20] Each user is assigned a credibility score based on the Linear classifier FakeNewsNet [23], [24] 87.8%
user’s online behavior.

Yang et al. [7] News propagation path through social media and captur- Neural networks Weibo [25] and Twitter 92%
ing the global and local variations of user characteristics. [26]
Volkova et al. [8] Linguistic and network features. Neural networks Twitter corpus [27] 95%

Siering et al. [21] Linguistic and content based features related to different SVM classifier 80%
Kickstarter [28]
types of communication (dynamic and static).
Brennan-Greenstadt [29],
Afroz et al. [22] Stylometry (lexical, syntactic, grammar, and author’s SVM classifier 96.6%
Hemingway-Faulkner [30]
attributes etc) to detect deception in online writing.
and Thomas-Amina [31]
Hannah et al. [9] Linguistic characteristics of untrustworthy text. NLP and Neural networks PolitiFact [32] 56% (F-score)

Ruchansky et al. [10] Temporal pattern and behavior of a user activity on the NLP and Neural networks Weibo [25] 95.3%
given article.

Hamid et al. [11] Local (similar to n-grams) and global features (capturing NLP and Neural networks 38.8%
LIAR [33]
temporal dependencies in the entire text).

Borges et al. [19] Combine representations inferred from the headline, first NLP and Neural networks NLIs [34], [35] and 83.38%
two sentences of news article, and entire document. FNC-1 [36]

with examples, and then discuss some of the current machine [9] H. Rashkin, E. Choi, J. Y. Jang, S. Volkova, and Y. Choi, “Truth of
learning techniques, such as NLP and deep learning, for varying shades: Analyzing language in fake news and political fact-
checking,” in Proceedings of the 2017 Conference on Empirical Methods
automatic detection of fake news. Based on this synopsis, in Natural Language Processing, 2017, pp. 2931–2937.
we recommend that there is a potential of using NLP and [10] N. Ruchansky, S. Seo, and Y. Liu, “Csi: A hybrid deep model for fake
deep learning to improve automatic detection of fake news, news detection,” in Proceedings of the 2017 ACM on Conference on
Information and Knowledge Management. ACM, 2017, pp. 797–806.
but with the right set of data and features. In future work, [11] H. Karimi, P. Roy, S. Saba-Sadiya, and J. Tang, “Multi-source multi-
we will explore the use of NLP with other machine learning class fake news detection,” in Proceedings of Data Science, Journalism
techniques, such as neural networks and other classifiers, to and Media workshop at KDD, 2018, pp. 1546–1557.
[12] A. Kirilin and M. Strube, “Exploiting a speakers credibility to detect
automate and improve fake news detection. fake news,” in Proceedings of the 27th International Conference on
Computational Linguistics, 2018.
R EFERENCES [13] T. T. Pham, “A study on deep learning for fake news detection,” Master’s
thesis, Japan Advanced Institute of Science and Technology, 2018.
[1] N. J. Conroy, V. L. Rubin, and Y. Chen, “Automatic deception detection: [14] M. L. Della Vedova, E. Tacchini, S. Moret, G. Ballarin, M. DiPierro, and
Methods for finding fake news,” Proceedings of the Association for L. de Alfaro, “Automatic online fake news detection combining content
Information Science and Technology, vol. 52, no. 1, pp. 1–4, 2015. and social signals,” in 2018 22nd Conference of Open Innovations
[2] “The government of japan announced that it was Association (FRUCT). IEEE, 2018, pp. 272–279.
banning the use of microwave ovens in the country [15] E. Kochkina, M. Liakata, and A. Zubiaga, “All-in-one: Multi-task learn-
by 2020,” 2019. [Online]. Available: https://check4spam.com/ ing for rumour verification,” arXiv preprint arXiv:1806.03713, 2018.
japanese-government-dispose-microwave-ovens-spam/ [16] A. Hanselowski, H. Zhang, Z. Li, D. Sorokin, B. Schiller, C. Schulz, and
[3] “President trump to ban schools in order to stop school I. Gurevych, “Ukp-athene: Multi-sentence textual entailment for claim
shootings,” 2018. [Online]. Available: https://www.snopes.com/ verification,” arXiv preprint arXiv:1809.01479, 2018.
fact-check/trump-end-shootings-ban-schools/ [17] S. D. Bhattacharjee, A. Talukder, and B. V. Balantrapu, “Active learning
[4] “President trump to ban schools in order to stop based news veracity detection with feature weighting and deep-shallow
school shootings,” 2018. [Online]. Available: https://me.me/i/ fusion,” in 2017 IEEE International Conference on Big Data (Big Data).
donald-j-trump-realdonaldtrump-im-putting-a-stop-to-school-20806779 IEEE, 2017, pp. 556–565.
[5] “Two altar boys arrested for putting marijuana in the censer- [18] F. Torabi Asr and M. Taboada, “Big data and quality data for fake news
burner,” 2018. [Online]. Available: https://www.snopes.com/fact-check/ and misinformation detection,” Big Data & Society, vol. 6, no. 1, p.
marijuana-censer-arrest 2053951719843310, 2019.
[6] “Barack obama attended columbia university as a foreign [19] L. Borges, B. Martins, and P. Calado, “Combining similarity features
student,” 2012. [Online]. Available: https://www.snopes.com/fact-check/ and deep representation learning for stance detection in the context of
obama-student-id checking fake news,” Journal of Data and Information Quality (JDIQ),
[7] Y. Liu and Y.-F. B. Wu, “Early detection of fake news on social media vol. 11, no. 3, p. 14, 2019.
through propagation path classification with recurrent and convolutional [20] K. Shu, S. Wang, and H. Liu, “Beyond news contents: The role of
networks,” in Thirty-Second AAAI Conference on Artificial Intelligence, social context for fake news detection,” in Proceedings of the Twelfth
2018. ACM International Conference on Web Search and Data Mining. ACM,
[8] S. Volkova, K. Shaffer, J. Y. Jang, and N. Hodas, “Separating facts from 2019, pp. 312–320.
fiction: Linguistic models to classify suspicious and trusted news posts [21] M. Siering, J.-A. Koch, and A. V. Deokar, “Detecting fraudulent
on twitter,” in Proceedings of the 55th Annual Meeting of the Association behavior on crowdfunding platforms: The role of linguistic and content-
for Computational Linguistics (Volume 2: Short Papers), 2017, pp. 647– based cues in static and dynamic contexts,” Journal of Management
653. Information Systems, vol. 33, no. 2, pp. 421–455, 2016.
[22] S. Afroz, M. Brennan, and R. Greenstadt, “Detecting hoaxes, frauds, and
deception in writing style online,” in 2012 IEEE Symposium on Security
and Privacy. IEEE, 2012, pp. 461–475.
[23] K. Shu, D. Mahudeswaran, S. Wang, D. Lee, and H. Liu, “Fakenewsnet:
A data repository with news content, social context and dynamic
information for studying fake news on social media,” arXiv preprint
arXiv:1809.01286, 2018.
[24] K. Shu, A. Sliva, S. Wang, J. Tang, and H. Liu, “Fake news detection on
social media: A data mining perspective,” ACM SIGKDD Explorations
Newsletter, vol. 19, no. 1, pp. 22–36, 2017.
[25] J. Ma, W. Gao, P. Mitra, S. Kwon, B. J. Jansen, K.-F. Wong, and M. Cha,
“Detecting rumors from microblogs with recurrent neural networks.” in
Ijcai, 2016, pp. 3818–3824.
[26] J. Ma, W. Gao, and K.-F. Wong, “Detect rumors in microblog posts
using propagation structure via kernel learning,” in Proceedings of the
55th Annual Meeting of the Association for Computational Linguistics
(Volume 1: Long Papers), 2017, pp. 708–717.
[27] “Dataset used in the paper: Separating facts from fiction: Linguistic
models to classify suspicious and trusted news posts on twitter,”
2017. [Online]. Available: http://www.cs.jhu.edu/{\raise.17ex\hbox{$\
scriptstyle\sim$}}svitlana/
[28] “Crowdfuncding platform kickstarter,” 2016. [Online]. Available:
https://www.kickstarter.com
[29] “Brennan-greenstadt adversarial corpus,” 2012. [Online]. Available:
https://psal.cs.drexel.edu
[30] “Hemingway-faulkner imitation corpus,” 2012. [Online]. Available:
http://web.archive.org/web/20051119135221/
[31] “Thomas-amina hoax corpus,” 2012. [Online]. Available: http://www.
joshualandis.com/blog/?p=1831
[32] “Politifact,” 2017. [Online]. Available: https://www.politifact.com
[33] W. Y. Wang, ““liar, liar pants on fire”: A new benchmark
dataset for fake news detection,” in Proceedings of the 55th
Annual Meeting of the Association for Computational Linguistics
(Volume 2: Short Papers). Vancouver, Canada: Association for
Computational Linguistics, Jul. 2017, pp. 422–426. [Online]. Available:
https://www.aclweb.org/anthology/P17-2067
[34] S. R. Bowman, G. Angeli, C. Potts, and C. D. Manning, “A large
annotated corpus for learning natural language inference,” arXiv preprint
arXiv:1508.05326, 2015.
[35] A. Williams, N. Nangia, and S. R. Bowman, “A broad-coverage
challenge corpus for sentence understanding through inference,” arXiv
preprint arXiv:1704.05426, 2017.
[36] “Fake news challenge stage 1 (fnc-1),” 2017. [Online]. Available:
http://www.fakenewschallenge.org