Content uploaded by Martins Dogo
Author content
All content in this area was uploaded by Martins Dogo on Aug 21, 2023
Content may be subject to copyright.
MASTER OF PHILOSOPHY
Exploring text representations for online misinformation
Dogo, Martins Samuel
Award date:
2023
Awarding institution:
Queen's University Belfast
Link to publication
Terms of use
All those accessing thesis content in Queen’s University Belfast Research Portal are subject to the following terms and conditions of use
• Copyright is subject to the Copyright, Designs and Patent Act 1988, or as modified by any successor legislation
• Copyright and moral rights for thesis content are retained by the author and/or other copyright owners
• A copy of a thesis may be downloaded for personal non-commercial research/study without the need for permission or charge
• Distribution or reproduction of thesis content in any format is not permitted without the permission of the copyright holder
• When citing this work, full bibliographic details should be supplied, including the author, title, awarding institution and date of thesis
Take down policy
A thesis can be removed from the Research Portal if there has been a breach of copyright, or a similarly robust reason.
If you believe this document breaches copyright, or there is sufficient cause to take down, please contact us, citing details. Email:
openaccess@qub.ac.uk
Supplementary materials
Where possible, we endeavour to provide supplementary materials to theses. This may include video, audio and other types of files. We
endeavour to capture all content and upload as part of the Pure record for each thesis.
Note, it may not be possible in all instances to convert analogue formats to usable digital formats for some supplementary materials. We
exercise best efforts on our behalf and, in such instances, encourage the individual to consult the physical thesis for further information.
Download date: 21. Aug. 2023
EXPLORING TEXT REPRESENTATIONS
for ONLINE MISINFORMATION
F
martins samuel dogo
April 2023
Thesis presented for the degree of
Master of Philosophy
to the
School of Electronics, Electrical Engineering
and Computer Science
colop h o n
This thesis is typeset in L
A
T
E
X. It adopts the typographical style
classicthesis
,
which was created by André Miede and Ivo Pletikosić, and inspired by Robert
Bringhurst’s book, The Elements of Typographic Style.
The text typeface is Sebastian Kosch’s Cochineal. The sans-serif typeface is
Bera Mono
, a modified version of Jim Lyles’ Bitstream Vera Mono font. Chapter
numbers are in Hermann Zapf’s Palatino.
Exploring Text Representations for Online Misinformation
github.com/m-arti/mphil
©April 2023, Martins Samuel Dogo
msdogo.com
For Favour Dogo
ABSTRACT
Mis- and disinformation, commonly collectively called fake news, continue to
menace society. Perhaps, the impact of this age-old problem is presently most
plain in politics and healthcare. However, fake news is affecting an increasing
number of domains. It takes many different forms and continues to shapeshift as
technology advances. Though it arguably most widely spreads in textual form,
e.g., through social media posts and blog articles.
Thus, it is imperative to thwart the spread of textual misinformation, which
necessitates its initial detection. This thesis contributes to the creation of repre-
sentations that are useful for detecting misinformation.
Firstly, it develops a novel method for extracting textual features from news
articles for misinformation detection. These features harness the disparity be-
tween the thematic coherence of authentic and false news stories. In other words,
the composition of themes discussed in both groups significantly differs as the
story progresses.
Secondly, it demonstrates the effectiveness of topic features for fake news
detection, using classification and clustering. Clustering is particularly useful
because it alleviates the need for a labelled dataset, which can be labour-intensive
and time-consuming to amass.
More generally, it contributes towards a better understanding of misinfor-
mation and ways of detecting it using Machine Learning and Natural Language
Processing.
iv
ACKNOWLEDGMENTS
Above all, I am most grateful to the LORD, my rock, my fortress, and my deliverer.
He is my King from of old. Palpable was His steadfast love, grace, and direction
throughout my research.
I am greatly indebted to my supervisors, Dr Deepak Padmanabhan and Dr
Anna Jurek-Loughrey, for their innumerable contributions and insights, patience,
encouragement, supervision and intellectual support.
I express my gratitude to Dr Paul Miller and Dr Barry Devereux, the latter of
whom served as the internal viva examiner, for providing me with constructive
criticism and valuable recommendations during my fair share of Annual Progress
Review meetings. I am grateful to the external viva examiner, Dr Seun Ajao,
for taking the time to examine my thesis, and providing detailed feedback and
suggestions, which have enhanced the quality of this work.
I am grateful for the help I have received from members of staff at EEECS. I
would like to acknowledge the help of Dr Jesus Martinez Del Rincon and Professor
Hans Vandierendonck. Special thanks to Mrs Katie Stewart for her unparalleled
support and guidance throughout my time at the school. I am deeply indebted to
Mrs Kathleen Ingram for her support, and help with proofreading my work.
In developing this thesis I have benefited from the constructive comments
and warm encouragement of friends and colleagues at EEECS. I wish to thank
Abdullah, Alimuddin, Ayesha, Maya, Michael, and Pritam. I enjoyed interacting
and exchanging ideas with you all.
I am also immensely grateful to friends who have helped and supported me
during this journey. Thank you, Abubakar, Akaoma, Chioke, Doreen, Michael,
Victor, Victoria, and Udeme. My heartfelt thanks to Kigwab for your warmth
and mirth.
Completing this thesis would not have been possible without the great personal
support of my family. I am extremely grateful to my parents, who have always
nourished and nurtured my curiosities, and to my siblings, for your profound
belief in my abilities. You have contributed much more to my work than you
realise.
Martins Dogo
February 2023
v
CONTENTS
iFake News
1Overview 2
1.1 InformationDisorder........................ 4
1.1.1 Theecosystem ....................... 5
1.1.2 Whatisfakenews?..................... 6
1.1.3 A Brief History of Fake News . . . . . . . . . . . . . . . 7
1.2 Fake News On Social Media . . . . . . . . . . . . . . . . . . . . 10
1.3 Why Machine Learning and Online Social Networks? . . . . . . 11
1.3.1 Access and participation on OSNs . . . . . . . . . . . . . 11
1.3.2 Are OSNs doing enough to curb misinformation? . . . . 12
1.3.3 Alleviating the strain of labelling . . . . . . . . . . . . . 15
1.3.4 Algorithms are versatile and catalytic . . . . . . . . . . . 15
1.4 Projectaim.............................. 15
1.4.1 Contributions........................ 16
2Related Work 17
2.1 Misinformation detection with Machine Learning . . . . . . . . 17
2.1.1 Featureextraction ..................... 17
2.1.2 Learning and classification . . . . . . . . . . . . . . . . . 17
2.2 Text representations for misinformation detection . . . . . . . . 19
2.3 Use of text in ML-based misinformation detection research . . . 20
2.4 Limitations of existing methods . . . . . . . . . . . . . . . . . . 27
ii Contributions
3Word Embeddings for Misinformation Detection 30
3.1 Background ............................. 30
3.2 Relatedwork............................. 30
3.3 Problemdefinition.......................... 31
3.4 Methodology and materials . . . . . . . . . . . . . . . . . . . . . 32
3.4.1 Experimental procedure . . . . . . . . . . . . . . . . . . 32
3.4.2 Datasets........................... 34
3.4.3 Results and discussion . . . . . . . . . . . . . . . . . . . 34
3.5 Disparities in sentiment . . . . . . . . . . . . . . . . . . . . . . 40
3.6 Conclusion.............................. 42
4Thematic Coherence in Fake News 43
4.1 Background ............................. 43
4.2 Relatedwork............................. 44
4.3 Problemdefinition.......................... 45
4.4 Research Goal and Contributions . . . . . . . . . . . . . . . . . 45
4.5 Methodology and materials . . . . . . . . . . . . . . . . . . . . . 46
vi
contents vii
4.5.1 Latent Dirichlet Allocation . . . . . . . . . . . . . . . . . 46
4.5.2 Distancemeasures ..................... 47
4.6 Experiment ............................. 50
4.6.1 Preprocessing........................ 50
4.6.2 Datasets........................... 50
4.7 Resultsanddiscussion........................ 52
4.7.1 Quantitative analysis of coherence and perplexity . . . . 58
4.8 Conclusion.............................. 61
4.8.1 Futurework......................... 61
5Clustering and Classification using Topics 62
5.1 Clustering .............................. 62
5.2 Classification............................. 69
5.3 Conclusion.............................. 72
iii Epilogue
6Conclusion 75
6.1 Futurework ............................. 76
LIST OF FIGURES
Figure 2.1 Fake news detection taxonomy . . . . . . . . . . . . . . 19
Figure 3.1
Box plots of distributions of
𝑐𝑠𝑎𝑣𝑔
for rumour and non-
rumourreactions...................... 36
Figure 3.2
Distributions of average pairwise cosine similarities be-
tween posts and their comments.
𝑁𝑅 =non-rumours, 𝑅 =
rumours........................... 38
Figure 3.3
Distributions of average pairwise Euclidean distances
between posts and their comments.
𝑁𝑅 =non-rumours, 𝑅 =
rumours........................... 39
Figure 3.4
Average sentiment scores of posts.
𝑁𝑅 =non-rumours, 𝑅 =
rumours........................... 40
Figure 3.5
Average sentiment scores of comments.
𝑁𝑅 =non-rumours, 𝑅 =
rumours........................... 41
Figure 4.1
Graphical representation of Latent Dirichlet Allocation
(LDA)............................ 47
Figure 4.2
Average and median Chebyshev distances in fake and
real news, when comparing topics in the first five sen-
tences to the rest of each article. Error bars show 95%
confidence interval. . . . . . . . . . . . . . . . . . . . . 54
Figure 4.2
Average and median Chebyshev distances in fake and
real news, when comparing topics in the first five sen-
tences to the rest of each article. Error bars show 95%
confidence interval. (Cont.) . . . . . . . . . . . . . . . . 55
Figure 4.2
Average and median Chebyshev distances in fake and
real news, when comparing topics in the first five sen-
tences to the rest of each article. Error bars show 95%
confidence interval. (Cont.) . . . . . . . . . . . . . . . . 56
Figure 4.2 UMass topic coherence scores . . . . . . . . . . . . . . 60
Figure 5.0
2D plots of dimension-reduced topic distributions for
datasetsused. ....................... 65
L I S T O F TA B L E S
Table 1.1 The matrix of misinformation . . . . . . . . . . . . . . 5
Table 2.1
Some of the main text representations for misinforma-
tiondetection ....................... 23
viii
Table 2.2
A breakdown of existing Machine Learning (
ML
) archi-
tectures for misinformation classification. . . . . . . . . 25
Table 3.1 Breakdown of PHEME dataset . . . . . . . . . . . . . . 34
Table 3.2
Summary of experimental and statistical results for
comparisons between sentence embeddings. . . . . . . 35
Table 3.3
Difference between the averages of the Euclidean dis-
tances of rumour and non-rumour comments. . . . . . 37
Table 4.1
Summary of datasets after pre-processing (F – Fake, R
–Real)............................ 52
Table 4.2
Results of T-test evaluation based on different measures
ofdeviationused...................... 53
Table 4.3
Mean and median
𝐷𝐶ℎ
deviations of
𝑁={10,20,30,40,50,100,150,200}
topics combined for fake and real news (F – Fake, R –
Real)............................. 53
Table 5.1
Purity scores for Baseline 1 (B1) and Aggregate (Agg)
methods. 𝑝= perplexity, 𝑛𝑛 = number of neighbours. . 67
Table 5.2 Purity scores for Baseline 2 . . . . . . . . . . . . . . . . 68
Table 5.3
Comparison of clustering purity scores.
𝑡
-distributed
Stochastic Neighbor Embedding (
tSNE
) with
𝑝=200
is
used to obtain 2D data. The best purity scores for each
datasetareinbold. .................... 68
Table 5.4
Evaluation metrics for the best-performing classifier on
each dataset, using the original representation.
𝑙2_𝑟𝑒𝑔 =
𝐿2regularisation..................... 70
Table 5.5
Evaluation metrics for the best-performing classifier
on each dataset, using the 2D representation.
𝑙2_𝑟𝑒𝑔 =
𝐿2regularisation
,
𝑛𝑛 =number of neighbours
,
𝑛𝑢𝑚_𝑡𝑟𝑒𝑒𝑠 =
numberoftrees....................... 71
Table 5.6
Evaluation metrics for the best-performing classifier on
each dataset, using the 300D representation.
𝑙2_𝑟𝑒𝑔 =
𝐿2regularisation, 𝑙𝑟 =learning rate, 𝑚𝑎𝑥_𝑡𝑟_𝑟𝑜𝑢𝑛𝑑𝑠 =
maximum training rounds, 𝑛𝑢𝑚_𝑡𝑟𝑒𝑒𝑠 =number of trees
.
72
LIST OF ALGORITHMS
Algorithm 1
Comparison of rumour and non-rumour comments
using InferSent embeddings.............. 33
Algorithm 2 Evaluation of thematic divergence in news articles . . . 49
ix
ac r onyms x
A C R O N Y M S
ABM agent-based model
ADS American Dialect Society
BERT Bidirectional Encoder Representations from Transformers
BoW bag-of-words
CDT Context Dimension Tree
CNN Convolutional Neural Network
kNN 𝑘-Nearest Neighbour
LDA Latent Dirichlet Allocation
LIWC Linguistic Inquiry and Word Count
LSA Latent Semantic Analysis
LSTM Long Short-Term Memory
ML Machine Learning
NLP Natural Language Processing
NMF Non-negative Matrix factorisation
OED Oxford English Dictionary
OSN Online Social Network
PCA Principal Component Analysis
PCFG Probabilistic Context-Free Grammar
PoS parts-of-speech
RNN Recurrent Neural Network
RST Rhetorical Structure Theory
SVD Singular Value Decomposition
SVM Support Vector Machine
TFIDF Term Frequency-Inverse Document Frequency
tSNE 𝑡-distributed Stochastic Neighbor Embedding
UMAP Uniform Manifold Approximation and Projection
VAE Variational Autoencoder
Part I
F A K E N E W S
1
O V E RV I E W
▶
The impact of news on daily affairs is arguably greater than it has ever been.
Its importance today can hardly be overstated. Although its ubiquity and influx
make news readily available, the news—despite being mostly useful—is increas-
ingly becoming skewed away from truth, towards more sensational headlines, as
competition for readership becomes more difficult.11
Wang (2012), Kavanagh et al.
(2019)
The news plays many roles. It informs the populace, inspires hope, and initiates
conversation, to name but a few. However, it sometimes leans towards rhetoric
rather than facts, so it is not always reliable. The news can also overwhelm with
its dizzying speed and endless breadth of topics. It can be argued that it would be
almost impossible to detach the news from everything else. The world may not
be able to function without it. Consequently, news can be leveraged for virtuous,
or vicious activities.
The fabrication and dissemination of falsehood have become politically and
economically lucrative endeavours, as well as tools for social and ideological
manoeuvre. Thus, the present times have been labelled as a post-truth period.
These endeavours have led to an intricately complex and constantly evolving
phenomenon that is mainly characterised by disinformation (false information
which is created or shared with malicious intent) and misinformation (also false,
but shared with harmless intentions). The pair is commonly collectively referred
to as fake news.22
The uses and misuses of this term
are discussed in more detail in
§
1.1.
▶
Several domains are affected by misinformation, especially in situations that
involve or affect many people, where uncertainty and tensions are high, and
resolutions are not forthcoming. These domains include but are not limited to:
33
Allcott and Gentzkow (2017), U.S.
Department of Homeland Security
(2018), Waldman (2018), Wardle
(2020), Chowdhury et al. (2021)
•Politics
: misinformation has historically and globally played a consequen-
tial role in politics. A recent example of this is during the 2016 U.S. Presi-
dential Elections.
•Healthcare
: during an epidemic such as the Ebola disease outbreak in 2014,
or a pandemic such as COVID-19, misinformation also spreads, particularly
through social media.
•Natural disasters
: examples include sharing falsified information (e.g., fol-
lowing the Fukushima Daiichi nuclear disaster in Japan, in 2011); inade-
2
overview 3
quate (e.g., during the earthquake in Nepal, in 2015); or mis-contextualised
(e.g., during an earthquake in Sicily, in 2014, whereby news of another one
from 1908 was referenced).
Central to this thesis is a computational investigation into some of the char-
acteristics of fake news text that differentiate it from truthful news. The texts
analysed in this work are in short and long forms—tweets related to news events
and full-length news articles, respectively. The datasets cover diverse domains,
including politics, sports, and conflict.
To better understand misinformation, it is important to first understand what
news is. The Oxford English Dictionary (OED) (2022) defines4news as: 4
This is the most relevant defini-
tion in the context of this thesis.
‘The report or account of recent (esp. important or interesting)
events or occurrences, brought or coming to one as new infor-
mation; new occurrences as a subject of report or talk; tidings.’
It is a somewhat subjective matter which events may qualify as important
or interesting. However, it can objectively be stated, that the alteration of new
information can distort a news report to the extent of falsifying events, thereby
rendering the news false. Certainly, even if all the information is true and verified,
the reportage can confuse or mislead, for instance, by means of rhetoric. It is
clear then, that faithfully narrating a vapid event in an engrossing manner, does
not constitute fake news, but perhaps is the result of skill or passion. On the
other hand, regardless of how interesting an event is, its misrepresentation may
misinform or disinform the reader. Therefore, it is important to categorise the
various ways in which—and degrees to which—information can be falsified.
This is because by doing so, a typology for distinguishing the different types of
misinformation can be created. Such a typology helps to identify the specific kind
of problem being dealt with, and in finding the optimum mitigation against it.
In this thesis, a typology called Information Disorder, which captures the essence
and full breadth of the mis- and disinformation landscape, is adopted. This is
discussed in the next section (§1.1).
▶
A plethora of sources now vie for the attention of readers—perhaps, more ‘The news knows how to render its
own mechanics almost invisible and
therefore hard to question.’
— Alain de Botton, “The News: A
User’s Manual”
than ever before. As a result, editors and journalists may be incentivised to
produce more sensational or emotionally charged pieces to invite, maintain
or grow readership. This is not a critique of journalists, nor an assessment of
their practices as measured against the principles and standards which apply to
their field. Rather, it is simply an observation of a trend. In many fields, business
and economic motives can clash with principles, and journalism is no exception.
As discussed later in this chapter,
5
besides sensationalism and bias, advertising is
5See §1.1.3 and §1.3.2.
one of the ways through which misinformation seeps into news pieces.
The lure of misinformation on social media typically begins with the title of
an article—often heightened by accompanying photographs. In the so-called
attention economy, attention is offered primacy because it is scant. With limited
1.1 in f o r mati o n disorde r 4
space, in adjacency with other publications, and within a publication itself, titles
must therefore strive to be eye-catching. This is often done at the expense of
high-quality information; at the same time, readers with limited attention tend
to share tawdry information.66
Menczer and Hills (2020), “The
Attention Economy”
As for the main content of a piece, a compelling story is naturally more captivat-
ing than a list of factual statements. Facts tend to be either bland and predictable,
and therefore boring—or strange and new, and therefore interesting. Most people
possibly prefer the latter.
1.1 informatio n d i s o r d e r
Mis- and disinformation are intertwined but essentially distinct phenomena. They
form a part of the broader landscape of what’s commonly, and often inaccurately,
called ‘fake news’. People, especially on the internet, have become accustomed to
referring to the entire landscape of false information as fake news. Although the
term suffices to indicate various types of false information and even sophistry in
biased articles, its unbridled use is problematic.77
A concise etymology of the term
‘fake news’ is related in
§
1.1.2, as
well as examples of its misuse.
The misinformation landscape as a whole is so complicated, that there is cur-
rently no firm consensus on terminology, nomenclature, and definitions amongst
researchers of the subject. Nonetheless, due to the acceleration of research in the
area, the different types of fake news are becoming more firmly grouped.
Wardle (2017) was among the first to propose a typology of fake news. It ‘When it was reported that Hem-
ingway’s plane had been sighted,
wrecked, in Africa, the New York
Mirror ran a headline saying, "Hem-
ingway Lost in Africa," the word
"lost" being used to suggest he was
dead. When it turned out he was
alive, the Mirror left the headline to
be taken literally.’
— Donald Davidson, “What
Metaphors Mean”
consisted of seven main categories, in increasing order of harmfulness: satire
or parody, false connection, misleading content, false context, imposter content,
manipulated content, and fabricated content. The groups were based on three
criteria: the type of information created and shared, the motivation behind the
creation of the content, and how it is disseminated. Though it received some
pushback, Wardle assiduously defended the inclusion of satire as a category
in a revised edition of her typology.
8
Having acknowledged that satire (when
8
Wardle (2020), “Understanding
Information Disorder”
intelligent) is a form of art, she explained that it is slyly used to veil canards and
conspiracies, and thus divert the attention of fact-checkers. Moreover, should
such a piece be later detected, its authors can simply claim that it was, after all,
not intended to be taken seriously. Table 1.1
9
summarises the types of mis- and
9Adapted from Wardle (2020).
1.1 in f o r mati o n disorde r 5
disinformation and their motivations, according to Wardle (2020).
type description
poor journalism
to pa rody
to provoke/‘punk’
pa s s i on
partisanship
prof i t
political influence
propaganda
Satire/Parody
No intention to cause harm but has
intention to fool
✓ ✓
False
connection
When headlines, visuals or captions
don’t support the content
✓ ✓
Misleading
content
Misleading use of information to
frame an issue or individual
✓ ✓ ✓ ✓
False context When genuine content is shared
with false contextual information
✓ ✓ ✓ ✓ ✓
Imposter
content
When genuine sources are
impersonated
✓ ✓ ✓ ✓
Manipulated
content
When genuine information or
imagery is manipulated to deceive
✓ ✓ ✓
Fabricated
content
New content that is 100% false, made
to deceive and do harm
✓ ✓ ✓ ✓
Table 1 . 1: The matrix of misinformation
1.1.1 The ecosystem
Wardle and Derakhshan (2017b) expand on Wardle’s original typology in what
‘What had gone wrong was the be-
lief in this untiring and unending
accumulation of hard facts as the
foundation of history, the belief that
facts speak for themselves and that
we cannot have too many facts, a
belief at that time so unquestioning
that few historians then thought it
necessary-and some still think it un-
necessary today-to ask themselves the
question: What is history?’
— E.H. Carr, “What Is History?”
can be regarded as one of the most in-depth explorations on the misinformation
landscape to date. In this work, they present a conceptual framework that offers a
useful perspective for understanding the misinformation ecosystem. In contrast
to other authors, they use the term information disorder as a substitute for fake news,
to encapsulate mis-,dis-, and, what they call mal-information—apt for conveying
the mélange of problems faced in a post-truth world.
1.1 in f o r mati o n disorde r 6
Simply put, disinformation involves intentionally creating or sharing false
information to cause harm. In other words, it contains deliberately and verifi-
ably falsified information. Mal-information is genuine information shared with
deceptive intent. Lastly, akin to a rumour, misinformation is false information
but not originally intended to cause harm. This should not be confused with
arumour, which is ‘an unverified or unconfirmed statement or report circulat-
ing in a community.’
10
A rumour may later be verified as true or false, whereas
10
Oxford English Dictionary
(2021), rumour | rumor, n.
disinformation is false from the onset.
Manifold, acceptable definitions of basic terms can be found in the literature of
misinformation; their misdefinitions can also be encountered. Therefore, paying
close attention to definitions of terms related to the problem is critical. Otherwise,
it may compound the problem. For example, Zubiaga et al. (2018) observed that
Cai et al. (2014) and Liang et al. (2015) incorrectly defined a rumour.11 11
Both define a rumour as false
information, whereas the proper
definition is as information whose
veracity is not yet known.
Wardle further corraled a fairly comprehensive glossary in which she defined
common terms and acronyms associated with the misinformation disorder land-
scape.
12
Attempts such as Wardle’s, to clarify misinformation-related terms are
12
Wardle (2018), “Information Dis-
order: The Essential Glossary”
crucial for aiding researchers and the general public in assimilating the scope of
the problem of misinformation. Another such work is that of the media historian
and theorist Caroline Jack, who created a lexicon for media content, aimed at
educators, policymakers and others.13 13
Jack (2017), “Lexicon of Lies:
Terms for Problematic Informa-
tion”
1.1.2 What is fake news?
Fake news essentially means disinformation. It arguably is the term most widely
used to refer to multiple categories of information disorder. Although the term
was once used in a corrective and progressive manner,
14
its positive connotation
14
Gelfert (2018), “Fake news: A
definition”
has since split into a duality—it is now used to refer to disinformation, and
critique and deride mainstream media.
15
Furthermore, ‘fake news’ is also used,
15
Wardle and Derakhshan (2017b),
Caplan et al. (2018)
ironically, to denounce or discredit factual information as misinformation. The
phrase has recently become a tool for tactical subversion from truth and, especially
in politics, for slander against dissenting opposition.
Documented uses of ‘fake news’ in writing date back to the 1890s; however,
other terms denotative of misinformation go as far back as the 16th century.
16 16
Merriam-Webster Dictionary
(2017), How Is ’Fake News’ Defined,
and When Will It Be Added to the
Dictionary?
Gelfert (2018) carried out an in-depth study of the etymology of fake news. The
study lists examples of previous attempts at defining the phenomenon and why
they are inadequate; it also gives historical examples of attempts to define fake
news. The following definition from Gelfert is adopted in this thesis:
‘Fake news is the deliberate presentation of (typically) false or
misleading claims as news, where the claims are misleading by design.’
The term ‘fake news’ may be unideal to refer to all kinds of misinformation.
However, it is popular among the public and researchers of misinformation alike.
Although ‘fake news’ may be a convenient catch-all term, it does not accurately
1.1 in f o r mati o n disorde r 7
reflect the nuances and complexities of misinformation. Therefore, it is crucial to
exercise caution when using this terminology and consider alternative descriptors
that may be more appropriate for specific contexts or types of disinformation.
As it is now used to denote the entire spectrum of information disorder, ‘fake
news’ sometimes educes ambiguity. Understandably, most people will not be fa-
miliar with the minutia of a research area, no matter how relevant it is. Moreover,
people may prefer simpler, more relatable terms for use in conversation. For
these reasons, it may be permissible to call most kinds of misinformation fake
news. However, the term is strictly inadequate and inaccurate. Misinformation
need not even be news in the first place. It is essentially corrupted information.
But if one generalises to say ‘fake information’, this is also problematic, because it
negates correct information that is mistakenly shared with a false context.
▶
In response, some have proposed using the term ‘false news’ to refer to disinfor-
mation instead.
17
But what happens when those who use fake news in subversive
17
Oremus (2017), Habgood-Coote
(2018)
ways also begin to use ‘false news’ in the same manner? Quite often, the intent of
an actor who shares problematic information cannot be promptly proven or in-
ferred. It appears, at least in research, that ‘misinformation’ is used as an umbrella
term for information disorder. This is less obscure because although it does not
strictly classify a piece of information, it still insinuates that the information is
problematic.
In this thesis, ‘fake news’, ‘false news’ and ‘misinformation’ are used interchange-
ably, in a broader sense to refer to the scope of information disorder. Moreover,
‘real’, ‘legitimate’ and ‘authentic’ are used to refer to reliable and truthful news.
The focus of this thesis is on finding an algorithmic solution to hindering the
spread of fake news, and not its epistemology. The reader is referred to Tandoc
et al. (2018), Torres et al. (2018) and Zannettou et al. (2019) for further in-depth
studies of the typology and epistemology of fake news.
1.1.3 A Brief History of Fake News
While it is beyond the scope this thesis to expand on the chronology of fake news,
‘Since wit and fancy find easier en-
tertainment in the world than dry
truth and real knowledge, figurative
speeches and allusion in language
will hardly be admitted as an imper-
fection or abuse of it. I confess, in
discourses where we seek rather plea-
sure and delight than information
and improvement, such ornaments as
are borrowed from them can scarce
pass for faults.’
— John Locke, “An Essay
Concerning Human
Understanding, Book III”
some key events may serve to sum up its timelessness. This summary will centre
on a few domains palpably affected by it: war, natural disasters, healthcare, and
politics.
Fake news predates news itself, at least, news conveyed through newspapers.
Dating back to the 17th century and originally called newsletters, newspapers
were simply printed or handwritten letters used to exchange tittle-tattle. This
activity grew and transformed into the production and consumption of modern
newspapers.18
18
Park (1923), “The Natural His-
tory of the Newspaper”
▶
Long before newsletters, however, the disinformation campaign had been
a tactic in use. One example of note was in the Roman Empire. Following the
demise of Julius Ceaser, Octavian and Antony launched disinformation campaigns
1.1 in f o r mati o n disorde r 8
against each other—employing propaganda, through the media of poetry, rhetoric
and newly minted coins—in a bid to become emperor. This led up to the Battle
of Actium, in 31 BC, out of which Octavian emerged the victor.
19
Though not its
19
Kaminska (2017), “A lesson in
fake news from the info-wars of
ancient Rome”
ultimate determiner, propaganda played a crucial role in the war. Misinformation
has since been a prime weapon in the arsenal of warring entities. Or for inciting
conflicts in the first place, as in the case of the Spanish-American War.20 20
Soll (2016), The Long and Brutal
History of Fake News
The media and speed of disseminating fake news have drastically advanced. At
the time of writing, Russia was continuing with its invasion of Ukraine. From
the onset, the Russian state used various disinformation narratives to justify the
invasion.
21
Its current model of propaganda is high-velocity and unremitting,
21
European External Action
Service (2022),U.S. Department of
State (2022)
high-volume and multichannel, and lacking in objective reality or consistency.
This approach has been developing since the Soviet Cold War era—to Russia’s
invasion of Georgia in 2008—to its annexation of the Crimean peninsula in
2014—and it is, in all probability, now deployed in Russia’s invasion of Ukraine.
22 22
Paul and Matthews (2016), “The
Russian "Firehose of Falsehood"
Propaganda Model: Why It Might
Work and Options to Counter It”
Clearly, then, misinformation has had an enduring influence on conflicts, but
so has it on many other areas of life: another is natural disasters. During the 15th
century the readership of news significantly expanded, thanks to the birth of
the printing press. Fake news followed suit, expectedly.
23
After all, the original
23
Soll (2016), The Long and Brutal
History of Fake News
newsletters helped gossip to set sail. After an earthquake in Lisbon, in 1755,
pamphlets containing fake news
24
were circulated around Portugal.
25
Today,
24
To be precise, this was a mix-
ture of witnesses’ accounts, false
context and manipulated content.
It brought forth a new genre of
sensational news called relações de
sucessos.
25
Araújo (2006), “The Lisbon
Earthquake of 1755 – Public
Distress and Political Propaganda”
there are various ways to fact-check news and other information. By contrast,
fact-checking was a rarity then. In want of scientific understanding, several
natural events, including natural disasters, were mystically interpreted.
▶
The term ‘infodemic’ was coined by Rothkopf (2003). It described the surge
of information, true and false, related to the 2003 SARS epidemic. Mindful to
not understate the severity of SARS itself, Rothkopf argued that the ‘information
epidemic’ that resulted from it added a new and more worrisome dimension
to the disease. Future global events would affirm Rothkopf’s ominous piece.
In the wake of the Coronavirus disease 2019 (COVID-19) pandemic, people
all over the world frantically sought information and many hastily acted upon
unverified information. To worsen matters, advice—including from the World
Health Organization, governments, and other trusted and reputable was not
only updated frequently, but at times, inconsistent. Naturally, some were stirred
to doubt, anxiety, and confusion. Meanwhile, a steady flux of misinformation
gushed out through Online Social Networks (
OSN
s) (also known as social media).
This culminated in an infodemic. Its impacts included psychological issues, loss
of public trust, loss of lives due to misinforming protective measures, and panic
purchase.
26
In 2018, the Democratic Republic of Congo experienced multiple
26
Pian et al. (2021), “The causes,
impacts and countermeasures of
COVID-19 “Infodemic”: A system-
atic review using narrative synthe-
sis”
outbreaks of the Ebola virus. The adoption of preventive measures against it
was hampered by misinformation and low institutional trust.
27
Other disease
27
Vinck et al. (2019), “Institutional
trust and misinformation in the
response to the 2018–19 Ebola out-
break in North Kivu, DR Congo: a
population-based survey”
outbreaks such as the Zika virus, Middle East Respiratory Syndrome, and H1N1
Influenza (swine flu) were all adversely affected by misinformation.28
28
Chowdhury et al. (2021), “Un-
derstanding misinformation info-
demic during public health emer-
gencies due to large-scale disease
outbreaks: a rapid review”
1.1 in f o r mati o n disorde r 9
▶
Newspapers have historically carried misinformation—and occasionally, dis-
information. Modern newspapers became more mainstream by the 19th century,
and through them, true and false news travelled faster and farther. The news
became more sensational too. For instance, in 1835, the New York Sun published
multiple false articles claiming that there were aliens on the moon. This was
known as the ‘Great Moon Hoax’.
29
In the 1890s, Joseph Pulitzer and William
29
Soll (2016), The Long and Brutal
History of Fake News
Hearst, rival American news publishers, contended for a larger readership of their
newspapers. Each sought to succeed by dubious practices—blatant reportage of
rumours as facts. This practice was known as ‘yellow journalism.’30 30
Center for Information Technol-
ogy & Society, UCSB (2022), A Brief
History of Fake News
The dynamics of misinformation became more complex when news leapt
from paper onto web pages. News became boundless, and so did misinforma-
tion. Meanwhile, sensationalism reigned on. By and by, news websites became
interlaced with advertisements, and sometimes, shockvertising (i.e., designed to
shock and provoke).
31
And to increase advertisement revenue, some resorted to
31
Oxford English Dictionary
(2022), shockvertising, n.
clickbait—attention-grabbing headlines designed to cajole readers into clicking
links. Clickbait has taken up residence on the internet. To sell advertisements,
‘drive traffic’, ‘increase engagement’, or simply mislead, many websites resort to
clickbait. It often misleads and can be acutely harmful. Yet, it remains inescapable
on the web. In fact, it is believed that misinformation—largely in the form of
clickbait shared on
OSN
s—influenced the outcome of the 2016 U.S. presidential
election.32 32
Allcott and Gentzkow (2017),
“Social media and fake news in the
2016 election”
Perhaps only coinciding with, rather than causing it, greater attention was being
paid to misinformation, as the internet made strides. Or perhaps, misinformation
simply grew too rapidly to be ignored. In 2005, the American Dialect Society (
ADS
)
named truthiness its word of the year.
33
The Merriam-Webster Dictionary did
33
American Dialect Society (2006),
Truthiness Voted 2005 Word of the
Year
the same the following year.
34
In 2016, the
OED
named post-truth its word of the
34
Merriam-Webster Dictionary
(2022), What is ’Truthiness’?
year.
35
The next year, the
ADS
and the Collins Dictionary both announced fake
35
Oxford Languages (2016), Ox-
ford Word of the Year 2016
news as their word of year.
36
In 2018, misinformation was named Dictionary.com’s
36
Wright (2017), American Dialect
Society (2018)
word of the year.37
37
Dictionary.com (2018), Misinfor-
mation | Dictionary.com’s 2018 Word
of the Year
▶
It is unlikely that any medium for sharing information that is open to the
general public will be immune to misinformation. To say nothing of sharing news.
The minimal cost of creating accounts and posting, combined with economic and
social incentives, particularly encourages bad actors.
38
One potential threat in
38
Shu et al. (2017), “Fake News
Detection on Social Media”
the future could be the misuse of generative artificial intelligence. It is likely that
deep fakes—in text, audio, image, and video forms—will become more sinister.
At any rate, they are becoming more realistic.
While exacerbating misinformation, the internet, at the same time, may be
the most effective tool for stifling it. Especially through the strategic use of
OSN
s. Inoculation or ‘prebunking’ is one such strategy. This means pre-empting
oncoming information with facts. According to Pilditch et al. (2022), inoculating
a critical mass of users in a network can inhibit the consolidation of falsehood.
39
Their experiments were carried out using agent-based models (
ABM
s). While
39
They found that inoculating sub-
sets of users at different times is
also effective.
their results are promising, it should be borne in mind that their setup, as well as
1.2 fake news on social media 10
ABM
s, are simplifications of the real-world. The same limitation applies to several
other proposed approaches for stopping misinformation, including
ML
-based
ones. Nonetheless, such research should be spurred on.
Via the internet, suspicious information can be scrutinised in near real-time.
Likewise, facts corrective to them can be dispensed quickly. Indeed, on the in-
ternet, registers of facts abound and are at hand for swift withdrawal. But fake
news is multiplicative. For regarding a single fact, countless false narratives can
sprout up. Therefore, it is easier for lies to accrete than for truth to flow. Another
strategy, nevertheless, is education: on how to spot and retard misinformation,
and how to seek and interpret facts.
It would seem that society is in an endless battle with misinformation; that it
may take one form or another, but can never be fully eradicated; and that it will
always be one step ahead of the safeguards in place. This could be true as long as
there remains insistence on velocity and volume rather than clarity and nuance,
and on flimsy metrics as the measure of the effectiveness of communication. All
these could come true as long as there remains insistence on velocity and volume
rather than clarity and nuance, and on flimsy metrics as the measure of commu-
nication. Not every problem can be solved by a technological breakthrough, or
simply more information, no matter how factual it is. Especially those problems
entangled with people’s identities, tightly held beliefs and opinions, and their daily
lives and bread. It may be necessary to rethink the design of communication tools
(both small and large scale), some of the incentives for these communications,
and the online communities that foster them. Misinformation will like become
an increasingly thorny issue in the future, and it is, therefore, crucial to think
outside the box to find effective solutions.
1.2 fake news on social media
Social media platforms provide a medium where the production and sharing of
news is not limited to established news agencies, but also open to the general
public.
40
News agencies used to be the main creators and distributors of news.
40
Campan et al. (2018), “Fighting
fake news spread in online social
networks: Actual trends and future
research directions”
Today, however, the general public is a lot more involved in that process.
41
In
41
Advances in smartphone and
web technologies, now allow
events to be broadcasted by
members of the public with great
speed and quality.
fact, so-called content creators (i. e., people from various fields who create media
content for consumption, primarily on the internet) are thriving, particularly
in technology news. Furthermore, people of all age groups and from all parts
of the world interact, share and exchange information on
OSN
s. This makes it a
suitable medium for rapidly spreading misinformation. Satisfactory solutions to
counteracting this challenge have not yet been found.
To sum up, misinformation on social media must be tackled. Given that this
problem is multifaceted and dynamic, the ideal solution would equally be holistic
and dynamic in its workings. Given its complexity, it must be approached pen-
sively and with nuance. It is highly unlikely that the solution will be simple, if
there is one at all—misinformation is a ‘wicked problem’.42 42
Rittel and Webber (1973) formu-
lated the idea of a ‘wicked problem’
(in social policy) as one that is oner-
ous or insoluble, characterised by
10 features, which Conklin (2006)
generalised to the following six:
1.
It is understood after a so-
lution is developed.
2. It has no stopping rule.
3.
Its solutions are not right
or wrong.
4. It is novel and unique.
5.
Every solution is a ‘one
shot operation’.
6.
It has no given alternative
solutions.
1.3 why machine learning and online social networks? 11
▶
Whether in research or deployment on the web, a multidisciplinary approach
is ideally needed to combat misinformation. It has traditionally been combatted
through manual fact-checking by experts. In addition to news agencies, other fact-
checking organisations such as FactCheck
43
, Snopes
44
and PolitiFact
45
employ
43 https://www.factcheck.org
44 https://www.snopes.com
45
https://www.politifact.com
such experts. Recently, however, computational techniques such as
ML
have been
used to detect misinformation.
This potentially allows for automated, real-time detection which can alert
people whilst or after engaging with misinformation. Furthermore, it can help in
identifying social media accounts that spread misinformation. A lot of research
work has been done in this area, but there remain limitations which hinder their
application in real-world scenarios. One such limitation is the need for large
news datasets annotated by experts.
1.3 why machine learning and online social networks?
1.3.1 Access and participation on OSNs
News is ubiquitous on
OSN
sand people access news through them. According to a
survey by the Pew Research Center in the United States, 53% and 48% of US adults
got their news from
OSN
sin 2020 and 2021, respectively.
46
The United Kingdom’s
46
Pew Research Center (2021),
News Consumption Across Social
Media in 2021
Office of Communications (Ofcom) stated in its 2021 report on nationwide news
consumption, that about half of adults in the UK access news on social media.
47 47
Ofcom (2021), News consump-
tion in the UK: 2021
This trend transcends the Anglosphere. The Reuters Institute for the Study of
Journalism at the University of Oxford, which aims to understand global news
consumption, has been publishing its Digital News Report for the past decade.
Its research focuses on countries with a high internet penetration and the 2021
report covered data from 46 countries across five continents. The report focused
on the six largest
OSN
s—Facebook, YouTube, Twitter, Instagram, Snapchat, and
TikTok—according to weekly use. In it, the Reuters Institute found that more
than half of the Facebook and Twitter users surveyed encountered news on those
platforms in the past week; for other networks, less than half of the users did.
48
They also found that for many Facebook users, the encounter with news on
48
Reuters Institute (2021), Digital
News Report 2021
the platform is incidental rather than intentional. In fact, some people report
avoiding it altogether.
It should be expected that more people seek and find news on social media.
After all, news sites and blogs share, and nudge people to share content on social
media. Social media is apt for aggregating news from various sources, as well as for
commentary and discussion. Furthermore, people themselves, now create news
online by directly posting onto their profiles. In other words, social media activity
sometimes is the news itself. Therefore, the creation of news is becoming more
democratised and social media is continuously being reinforced as the global
nucleus of news activity—from witnessing to disseminating, to assimilating.
1.3 why machine learning and online social networks? 12
However, along with this new-found voice and power follow ramifications. Most
inimically, noise and lies compete with signal and truth, for space and attention.
▶
Apart from reading news, people are generally spending more time on social
media. It is also used to interact with friends and strangers, engage in public
discourse or dissent, and more recently, to shop and donate to charitable causes,
directly. It is not an overstatement, then, to say that social media has accrued an
enormous value—or cost, depending on how one sees it—for nearly everyone.
Misinformation arguably spreads the fastest on
OSN
samongst news media. Now,
if people’s lives and livelihoods continue to be intricately intertwined with so-
cial media—if people are to find ways of navigating, or escaping, the real-world
through it; to stay in touch and make new friends; to form and maintain commu-
nities and identities; find self-expression; to share memes and commiserate with
one another—then it is worth protecting. Especially when it influences real-world
events and politics. One of the consequences—or benefits, as the case may be—of
wallowing in social media feeds is that it gradually shapes one’s worldview. The
design and resulting dynamics of
OSN
smake their users susceptive of a myriad of
biases—information, political, cognitive, etc.49 Fake news detection is currently 49
Menczer and Hills (2020), Bar-
rett et al. (2021)
mostly done by human experts. This is very expensive and time-consuming given
the deluge of misinformation that parades OSNsdaily. This work contributes to
lessening the cost and effort spent by experts.50 50 See §1.3.3.
1.3.2 Are OSNs doing enough to curb misinformation?
At the ever-rising speed and scale of misinformation dissemination on social
media, and considering that more and more people are reading news on them,
the problem is proving to be insurmountable for human experts alone to deal
with. The situation is critical, and the skills and resources needed for repair are
limited. However,
ML
algorithms can augment the effort of experts combatting the
problem. An example of how this can be done is explained in the next subsection
(
§
1.3.3). Beyond intercepting misinformation, algorithms, more generally—as
can be seen in this thesis and some of the works cited in it—are extending the
capacity for unravelling the tangle of misinformation. A collection of algorithms,
therefore, can act both as tools and as catalysts, matching the speed and scale at
which misinformation propagates on OSNsand its complexity.
▶
Whilst employing people to spot problematic content including misinfor-
mation and false news ensures detection accuracy, this has been found to have
detrimental effects on the moderators of social media content.
51
Firstly, repeated
51
Newton (2019), The secret lives
of Facebook moderators in America
exposure to the kinds of disturbing media moderators scour out, can corrode a
person’s mental well-being. Besides that, reading false information repeatedly
can lead one to believe it is true; this is a phenomenon called the illusory truth
effect.
52
Finally, in spite of their invaluable contributions, content moderators are
52
Pennycook and Rand (2021),
“The Psychology of Fake News”
rather stingily remunerated for their work.
1.3 why machine learning and online social networks? 13
In the case of Facebook, moderators are paid as low as $1.50 and $15 per
hour, in Kenya and the United States, respectively. In both cases, these people
are employed by contractors and not directly by Facebook. However, Facebook’s
own employees audit their work and periodically visit the contractors’ offices
for monitoring. Nonetheless, their pay is meagre and they are treated poorly,
all in sharp contrast to the median salary of $240,000 and numerous additional
perks, which Facebook employees enjoy.
5354
Content moderators have reported
53 Newton (2019), Perrigo (2022)
54
This comparison does not mean
to suggest that moderators should
receive equal pay to other em-
ployees. (Although they definitely
should be paid more and treated
better.) Rather, it serves to eluci-
date that their role is regarded as
subservient to those of others.
struggling with mental trauma and indeed, some have been diagnosed with trau-
matic stress disorders. This is supposedly triggered by the appalling content they
review. However, they have also reported facing intimidation and overwhelming
pressure from their managers at work.
55
This compounds their work-related
55 Ibid.
stresses rather than alleviating them. In 2019, a Facebook content moderator
passed away, at work, at his desk. The management of the contracted company
initially responded by dissuading their employees from discussing the tragedy,
because they worried that it would dwindle productivity.
56
These findings raise
56
Newton (2019b), Bodies in Seats:
Facebook moderators break their
NDAs to expose desperate working
conditions
some serious questions about the earnestness of social media platforms in fighting
misinformation.
Misinformation is a dynamic and convoluted problem. So much so that it
seems misinformation will never be totally eradicated—but will always take one
form or another—and can only be repeatedly extinguished. Such a volatile and
amorphous nature demands supervision and intervention by experts. It is clear
to see, then, the long-term significance of content moderation on
OSN
s, and the
internet as a whole.
▶
It would be unfair to social media companies if their efforts in combatting mis-
information were not recognised. They have undertaken and funded numerous
projects and initiatives, which demonstrate a sincere concern for the safety of
their users. These also throw light on the multifaceted nature of the problem at
issue.
Firstly,
OSN
sprovide access to data for research purposes, through APIs or
competitions. Research activities such as fake news detection using
ML
will not
be practical without datasets, though some researchers have expressed a demand
for additional data, e.g., impression data.
57
In addition to data,
OSN
splatforms
57
Pasquetto et al. (2020), “Tackling
misinformation: What researchers
could do with social media data”
support researchers with grants. Therefore, notable contributions are made in
support of research activities.
Secondly,
OSN
sare making it easier for people to flag or report posts they deem
ill. Twitter, for instance, has taken this one step further through their Birdwatch
pilot programme.
58
Users (in the U.S., for the time being) can directly annotate
58
Coleman (2021), Introducing
Birdwatch, a community-based
approach to misinformation
tweets they believe to be misleading, thereby providing context for the flag. These
notes are publicly viewable by anyone. Beyond simplistic binary labels, this will
provide more insight to Twitter users and researchers alike, for understanding
the roots of misinformation.
Thirdly, these companies have established coalitions and partnerships with
academia, media, fact-checking and other organisations, to work together to-
1.3 why machine learning and online social networks? 14
wards achieving shared goals for the public good. Notable examples of such
consortia include Social Science One,
59
the Content Authenticity Initiative,
60 59
https://socialscience.one
60 https:
//contentauthenticity.org
and the Coalition for Content Provenance and Authenticity.
61
An outstanding,
61 https://c2pa.org
individual example is the Google News Initiative,
62
which boasts more than 7,000
62 https://
newsinitiative.withgoogle.com
partnerships and $300 million in funding to various organisations in over 120
countries.
In addition, OSNsalso:
•
build in-house tools for detecting misinformation; they also incorporate
new tools and expertise through company acquisitions (e.g., Fabula AI
being acquired by Twitter,63 and Bloomsbury AI by Facebook64)63
Agrawal (2019), Twitter acquires
Fabula AI to strengthen its machine
learning expertise
64
Winick (2018), Facebook’s latest
acquisition is all about fighting fake
news
•
create robust, independent and transparent decision-making structures,
which include external experts (e.g., the Oversight Board
65
established by
65 https:
//www.oversightboard.com
Facebook in 2018, which oversees critical content moderation on Facebook
and Instagram)
•
try to adapt their policies to current affairs and adhere to government
policies around the world
All that is mentioned here is not an exhaustive list of measures taken by social
media platforms. But are they doing enough? While some of their efforts are
commendable, there are areas where OSNsought to improve.
▶
It is helpful to constantly bear in mind that profit—primarily through adver-
tising—is a top priority for social media companies. Notwithstanding the Google
News Initiative’s generous funding to various organisations, a noteworthy detail
is that Google itself has a news product, Google News, which helps to drive user
engagement with other Google products, such as search. Moreover, as of 2018
news accounted for 16-40% of Google Search results, and content crawled and
scraped from news publishers drew in an estimated $4.7 billion according to
the News Media Alliance (2019). Debate continues as to whether
OSN
sshould
reward publishers for their images and text which appear in search results, or
if the publishers are better off for the additional web traffic. Publishers initially
received nil from OSNsfor their content, but this is no longer the case.66 66
Google France (2021), Le blog
officiel de Google France: L’Alliance
de la Presse d’Information Générale
et Google France signent un accord
relatif à l’utilisation des publications
de presse en ligne
Are
OSN
swilling to come up with tougher policies, which may hinder misinfor-
mation at the expense of some profit? Misinformation is common in advertising.
According to Chiou and Tucker (2018), advertising makes a significant contribu-
tion to the spread of misinformation. To give some perspective, the U.S. Federal
Trade Commission filed more than 150 instances of misinformation in adverts
between 2015 and 2020; the settlements were as high as $191 million.67 67
Fong et al. (2021), “Debunking
Misinformation in Advertising”
▶
All in all, users have a role—perhaps the biggest role, individually and col-
lectively—to play in curbing misinformation. After all, users—businesses and
individuals—generate most of the content on social media. In fact, according to
1.4 pr o ject aim 15
Vosoughi et al. (2018), real human accounts not bots, are mostly responsible for
sharing misinformation on Twitter.
1.3.3 Alleviating the strain of labelling
Supervised
ML
models for detecting fake news rely on labelled data. As such
datasets are usually large, labelling them can be tedious. This process is expensive,
exhausting and in some cases, detrimental to the well-being of those carrying
out the task. These issues, as well as the low wages paid for labelling tasks, are
discussed in the previous subsection. Further potential problems with labelling
are that it does not scale, and it has an element of subjectivity.
Unsupervised
ML
models, on the other hand, do not rely on labelled data.
Therefore, it can help to alleviate some of these issues. It would be ideal to
minimise the effort required for labelling data while maintaining accuracy. In
that sense, therefore, this thesis explores unsupervised learning as an alternative
to supervised learning.
1.3.4 Algorithms are versatile and catalytic
Will
ML
algorithms someday be able to speedily and single-handedly spot every
problematic content on
OSN
s? This is unlikely, for there will always be many
borderline cases, and even humans sometimes disagree on how content should
be classified. However, when the economic and psychological costs of human
reviewing are considered,
ML
can make significant contributions to curbing
problems such as information disorder. Moreover, it has, by and large, successfully
been used to tackle other issues such as nudity on OSNs.
The scale of
OSN
smake them fertile grounds for the rapid spread of false
news, with billions of people actively using them. History shows that
OSN
shave a
revolutionary power. For instance, social media played a critical role in the Arab
Spring of 2011
68
, and more recently, in the 2016 U.S. Presidential Elections
69
.
68
Brown et al. (2012), The Role of
Social Media in the Arab Uprisings
69
Allcott and Gentzkow (2017),
“Social media and fake news in the
2016 election”
Further,
OSN
sare environments from which new culture (e.g., memes) permeates
into the real world, and they, therefore, influence the lives of individuals. As such,
it is important to rid it of harmful actors and behaviours such as misinforma-
tion. Fortunately, the availability of datasets on false information in
OSN
smakes
research on combatting the issue with algorithms feasible.
1.4 project aim
Existing implementations of semi-supervised and unsupervised
ML
are fewer
and less varied than those of supervised learning. Work has been done aplenty in
the supervised learning space, and good progress has been made. However, there
are limitations which restrict its applicability, such as the need for labelled data.
1.4 pr o ject aim 16
There is also a need for new text representations that are robust for detecting
misinformation. For example, it is common to use text features based on writing
style. However, they may not be robust enough to identify false news written in
a similar style to real news.
▶
The aim of this research is to develop a novel approach for generating text
representations from short and long-form texts. Furthermore, it aims to demon-
strate the efficacy of such representations for misinformation detection, using
unsupervised and supervised ML.
First, in Chapter 2, this thesis explores existing ways of utilising text features
for misinformation detection. Second, in Chapter 3, experiments exploring how
to harness text representations to detect fake news are presented. Chapter 4
introduces the concept of thematic coherence, based on analyses of topic features
in news pieces. Finally, Chapter 5 shows results for detecting misinformation
with topic representations using clustering and classification.
1.4.1 Contributions
Given their influence and harmfulness, a lot of research work has been done to
address the elements of information disorder. This research focuses on mis- and
disinformation. It mainly contributes to the existing body of work on misinfor-
mation detection using Natural Language Processing (NLP) and ML.
Firstly, an exploration of features for misinformation detection is carried out.
A novel feature extraction approach, involving topic modelling, for classifying
and clustering news articles is presented. Topic-based features are advantageous
in situations where labelled data is difficult to acquire, available in a small quantity
or non-existent. Additionally, topic features may be more robust when faced with
machine-generated fake news, unlike the commonly used stylometric ones.70 70
Schuster et al. (2020), “The Limi-
tations of Stylometry for Detecting
Machine-Generated Fake News”
Secondly, supervised and unsupervised
ML
methods are applied to detect
misinformation, in multiple cross-domain datasets.
Lastly, the findings of this research may be applicable in other problem areas
on the spectrum of information disorder in news text, e.g. hate speech detection.
Also, this research more broadly contributes to the field of
NLP
. The experiments
carried out and their results may be informative to other researchers in the
field. The code for all the experiments presented in this thesis is available at
https://github.com/m-arti/mphil.
2
RELATED WORK
As information disorder continues to evolve, so too do surveys of research efforts
in combatting it constantly diversify. This is markedly the case for scientific ap-
proaches. This diversity in perspectives and approaches indicates the complexity
of information disorder. It also signifies the necessity for a holistic view of the
problem.
2.1 misinformation detection with machine learning
Existing approaches to misinformation detection using news text data generally
involve two subtasks: feature extraction, and learning and classification.
2.1.1 Feature extraction
Feature extraction is a process by which attributes of news items on
OSN
posts
are extracted and processed for classification. Shu et al. (2017) categorised these
features into two groups, based on: (i.) news content, i.e., text and image features;
and (ii.) social context, i.e., features based on users, posts, and networks.
Numerous papers use text features to detect fake news. Being the feature of
interest in this work, this is expanded on in
§
2.2 and
§
2.3. In addition to images,
videos and speeches are also used to extract features for fake news detection.
Multiple features can be combined for detection, i.e., in a multimodal fashion.
Alam et al. (2022) surveyed multimodal fake news detection. Similarly, Cao et al.
(2020) gave a comprehensive overview of the role visual content plays in fake
news detection, while Shu et al. (2020b) did the same for user profiles. Zhou and
Zafarani (2019), and Shu et al. (2020c) demonstrate the application and efficacy
of network-based features.
2.1.2 Learning and classification
An
ML
model is then trained using the extracted features, to classify new, unseen
news items or posts. The training process can be:
17
2.1 misinformation detection with machine learning 18
•Supervised
: data with labels (typically ‘real’ or ‘fake’) are applied to train a
classifier, e.g. neural network, decision tree, Support Vector Machine (
SVM
),
etc.
•Semi-Supervised
: this approach primarily aims to attenuate reliance on
labelled data, which may be insufficient. It leverages unlabelled data to
make predictions with higher accuracy than would have been attained
using only labelled data.71 Commonly known examples include:72 71
Ouali et al. (2020), “An Overview
of Deep Semi-Supervised Learn-
ing”
72 Ibid.
–
generative models, which initially learn features from a given task
and are afterwards used in other tasks.
–
proxy-label methods, which utilise a model trained on a labelled
dataset to generate more training data by labelling examples of unla-
belled data.
–
graph-based methods, which model labelled and unlabelled data as
nodes in a graph and try to propagate labels from the former to the
latter.
Semi-supervised learning also allows for a human-in-the-loop detection
process. Some of the data are unlabelled in this case. An example is active
learning, whereby labels for the most ambiguous training examples are
sought from a human—a content moderator, for instance—to progres-
sively improve the classification accuracy.
•Unsupervised
: all data are unlabelled in this case. The task could be one of
clustering, or anomaly detection whereby a fake news item is picked up as
an outlier in the dataset.
Though wanting in visualisation, the overview of information disorder given
by Zannettou et al. (2019), which is based on the work of Wardle (2017),
73
is
73 See §1.1.
sufficiently encompassing. It includes the various types of false information,
its actors, as well as their motivations. Furthermore, works that analyse how
false information propagates via different
OSN
s, as well as those which focus
on how to detect them, are discussed. By comparison, Wardle and Derakhshan
(2017b) equally gives a comprehensive view of the ecosystem—additionally, with
a visual illustration to better the reader’s understanding—of the types, actors,
motives and phases of misinformation. However, a similar illustration for existing
computational approaches is lacking in the literature.
Figure 2.1 shows a taxonomy of the different methods used to detect fake
news and the sub-classification of tasks within each. In the following subsec-
tion, examples from the literature of each method and features utilised will be
discussed.
2.2 text representations for misinformation detection 19
Figure 2.1: Fake news detection taxonomy.
2.2 text representations for misinformation detection
Although today it is produced and consumed in various media, news has pre-
eminently been circulated in textual form. Over time, textual news developed
into a general ossified structure:
•Source: the author and/or publisher of the article.
•
Headline/title: this typically is a short sentence, descriptive of the principal
news topic covered in the article.
•Body: the main text of the article, detailing the news story.
•
Image/video: visual (or audio-visual) cue(s) included in the body of the
article.
▶
In an in-depth study of the structure of news, van Dijk (1983) described the news
as having two kinds of structure: thematic and schematic. The former represents
the topical contents of a news item, while the latter describes the structure of
the item’s discourse. In other words, a news item is composed of themes bound
together by a schema. These themes may vary (in nature and style of presentation)
from one article to another, but the schema is firmly established.74 74
van Dijk’s study and conclusions
were based on an empirical study
of global press coverage of the
1982 assassination of the Lebanese
president-elect, Bachir Gemayel. It
covers 700 news articles published
by 250 newspapers from 100 coun-
tries.
Online news presumably inherits its structure from that of traditional, printed
news. Though the visual layout may be notably different. For instance, news on
the web is typically laid out in a single, rather than multiple columns, has ‘share’
buttons, etc. However, its schema is identical to that of print news. Implicit in this
schema, is a top-to-bottom outline of the news content, ranked from the most to
the least important or newsworthy fragments.
75
It is conventional for journalists
75
van Dijk (1983), “Discourse Anal-
ysis: Its Development and Applica-
tion to the Structure of News”
to produce—as it is for readers to assimilate—news in this manner. It is likely,
2.3 use of text in ml-based misinformation detection research 20
therefore, for the most relevant textual content for news analysis to be found
closer to the top, rather than at the bottom of the news item.
Although published nearly four decades ago—long before the dawn of news
detection using
NLP
and
ML
—van Dijk’s paper provides some interesting and
practical insights that could inform its current modus operandi. For instance, he
gleaned from his study, that:
•the paramount topic of a news item is captured in its headline; and
•
the opening sentences and paragraphs form the top of the schema—containing
crucial details such as the time, location, parties, causes and outcomes of
the main news events.
To summarise, the hierarchical structure of news embodies a linearly decreas-
ing ordering of thematic information in a news item, from top to bottom. van Dijk
attributes this order to “an implicit journalistic rule of the news organization.”
As news is written, so it is read. Thus, one can catch the scope of a news item
by simply reading the headline and the main details in the opening section. This
is true for most authentically produced news articles that adhere to journalistic
standards. However, news produced with bad intent can exploit the hierarchy of
its content for mal-intent.
Some have attempted to take advantage of this hierarchy for misinformation
detection,
76
while most works using text data have relied on the body of articles.
76
For example, Biyani et al. (2016)
used features, including titles, ex-
tracted from news webpages to
detect clickbait; Sisodia (2019) ex-
tracted features from headlines to
do the same; while Yoon et al.
(2019) assessed the congruity be-
tween news headlines and body
texts.
One of the main contributions of this thesis is that it presents a way to harness
the inherent schematic of news, for detecting misinformation.
2.3 use of text in ml-based misinformation detection
research
Whereas photos and videos were once merely accompaniments to news pieces,
they are gradually taking centre stage in news dissemination, especially on
OSN
s.
Nonetheless, text remains the predominant and most abundant form of news.
Similarly, misinformation proliferating through social media and the web is
typically in the form of text, and photos or videos are only recent developments.
Besides, text can be extracted from news items disseminated in pictorial or video
forms for analysis. For example, text extracted from a news video through speech-
to-text technology can be used for
NLP
analysis. Text can similarly be extracted
from photos. Expectedly, research in misinformation detection has mostly utilised
text data as raw material, and
NLP
and
ML
techniques for extraction, enrichment
and categorisation.
This research explores text representations for detecting online misinforma-
tion. In other words, it aims to find effective means of transforming text data
into meaningful representations that can be used to characterize or identify fake
news.
2.3 use of text in ml-based misinformation detection research 21
This section describes an overview of some key papers on text representations
for misinformation detection. Most papers selected for this discussion focus their
approach on two main strategies for exploiting textual data:
1. text-based features, generally extracted from the body text.
2.
the schema of news: i.e., papers which exploit features from a specific
portion of news articles, such as headlines.
This section will focus on text-based features used for misinformation detec-
tion using
ML
. Shu and Liu (2019b) categorises such features into three groups:
(i.) linguistic, (ii.) low-rank, and (iii.) neural text features. More elaborately, Zhou
and Zafarani (2020) additionally categorise linguistic features into four groups:
(i.) lexical, (ii.) syntactic, (iii.) discourse, and (iv.) semantic.
▶
Linguistic features aim to capture the style of writing in a piece of text.
From this style, intent may be inferred (i.e., whether to mislead or not),
77
or
77
Zhou and Zafarani (2020), “A
Survey of Fake News: Fundamen-
tal Theories, Detection Methods,
and Opportunities”
characterisation can be made (since fake news will likely have a style that differs
from that of authentic news).
The following are some linguistic features and their applications for fake news
detection:78 78
Shu and Liu (2019b), Zhou and
Zafarani (2020)
•
Lexical features are generally concerned with the tallies or frequencies of
character- or word-level features. Examples include
𝑛
-grams, bag-of-words
(
BoW
) methods, and Term Frequency-Inverse Document Frequency (
TFIDF
),
which captures the relevance of a given word to a document in a corpus.
Another is the Linguistic Inquiry and Word Count (
LIWC
), which calculates
what percentage of words in a text fall into one of many categories, which
indicate emotional and psychological properties, amongst others.
•
Syntactic features are typically sentence-level features, including counts of
punctuations, words, phrases, parts-of-speech (
PoS
) tagging, and Probabilistic
Context-Free Grammar (
PCFG
) parse trees. Additional examples of these
features are those specific to the news domain, such as quotations and
links.
•
Discourse features include applications of the Rhetorical Structure Theory
(RST) and rhetorical parsers to extract rhetorical features from sentences.
•
Latent features are primarily embeddings created using deep neural net-
works such as Convolutional Neural Networks (
CNN
s), Recurrent Neural
Networks (
RNN
s) (particularly, using the Long Short-Term Memory (
LSTM
)
architecture), and Transformers. These embeddings are dense vector rep-
resentations of text at the word (most commonly), sentence, or document
level. Commonly used word embedding models include
word2vec79
, and
79
Mikolov et al. (2013),
“Distributed Representations of
Words and Phrases and their
Compositionality”
more recently, transformer-based architectures such as BERT
80
and its
80
Devlin et al. (2019), “BERT:
Pre-training of Deep Bidirectional
Transformers for Language
Understanding”
2.3 use of text in ml-based misinformation detection research 22
variants. Another latent feature that is particularly relevant to this work
(in Chapter 4) is the topic feature, extracted using topic modelling. Topic
models identify themes latent in a group of documents by analysing the
distribution of words and/or phrases across them.81 81
Blei (2012), “Probabilistic topic
models”
Casillo et al. (2021) used a combination of topics, syntactic, and semantic
features from news texts in three datasets to detect misinformation. They obtained
the topic features using the
LDA
topic model.
LDA
is also used in this work and an
in-depth explanation of its workings is given in
§
4.5.2. Stopwords were removed
before feature extraction.
82
They used three syntactic features: (i.) the number of
82
stopwords are words such
as ‘just’,‘do’, and ‘it’, which
are non-descriptive, and
therefore, relatively less insightful
with regard to generating or
interpreting topics.
characters; (ii.) the Flesch Index, which is a measure of text readability; and (iii.)
the Gunning Fog Index, which estimates text comprehensibility. These features
are further processed using the Context Dimension Tree (
CDT
), which aids the
selection of topics using temporal context. Next, they incorporate two semantic
features—the probabilities of negative and positive news sentiment. Finally, the
features are fed into a 𝑘-Nearest Neighbour (kNN) classifier for detection.
Another work that uses topic features is by Hosseini et al. (2022). Similar to
the previous work, an
LDA
topic model was used to extract features. Before this,
though, the texts are preprocessed into tokens, and non-English words and stop-
words are removed. Word embeddings are obtained from the original news texts
using the
word2vec
model. These embeddings are input into a bi-directional
LSTM
Variational Autoencoder (
VAE
) to form latent representations of the texts.
The
VAE
representations are combined with the
LDA
topic representations to
form the final features for classification. The combined features improved misin-
formation detection for classifiers compared to individual features.
Topic features can be extracted from non-English news texts. They can also
be used for tasks other than detection. For example, Paixão et al. (2020) used
BoW
, word embedding,
LIWC
,
PoS
, and
TFIDF
features to differentiate between
real and fake news in a Brazilian Portuguese news corpus. However, they further
employed topic modelling to qualitatively study the two groups of articles in the
dataset. They found the optimal number of topics to analyse using the coherence
measure. This is also used in this work, although in a different way, in §4.7.1.
LDA
is not the only topic modelling method available, but it is more commonly
used in the literature. Ajao et al. (2019) experimented with a different method
called Latent Semantic Analysis (
LSA
), but found
LDA
to perform better. They
applied topic modelling to determine the 10 most prevalent topics in rumour
and non-rumour tweets. The sentiment (positive, negative, or neutral) values of
the words in each topic were then computed and used to calculate an emotional
ratio score. This score was combined with linguistic features such as counts of
user mentions, hashtags, and quotations, to form features for rumour detection.
2.3 use of text in ml-based misinformation detection research 23
Table 2.2 shows some of the commonly used text-based features for misinfor-
mation detection and examples of papers wherein they are implemented.
text feature papers implemented in
Lexical BoW: Paixão et al. (2020), Zhou et al. (2020b);
LIWC: Pérez-Rosas et al. (2018), Paixão et al. (2020);
𝑛
-grams: Biyani et al. (2016), Ahmed et al. (2017), Potthast
et al. (2018);
TFIDF: Biyani et al. (2016); Pérez-Rosas et al. (2018)
Others: Biyani et al. (2016), Potthast et al. (2018), Yang et al.
(2019), Paixão et al. (2020)
Syntactic PoS
: Feng et al. (2012), Potthast et al. (2018), Paixão et al.
(2020), Zhou et al. (2020b);
PCFG
: Feng et al. (2012), Pérez-Rosas et al. (2018), Zhou
et al. (2020b);
Others: Potthast et al. (2018)
Discourse RST: Rubin and Lukoianova (2015);
Others: Karimi and Tang (2019), Zhou et al. (2020b)
Latent CNN: Wang (2017b), Ajao et al. (2018), Yang et al. (2018);
RNN
: Rashkin et al. (2017), Ruchansky et al. (2017), Ajao
et al. (2018), Karimi and Tang (2019), Zhang et al. (2019),
Hosseini et al. (2022);
Transformers: Vijjali et al. (2020), Kula et al. (2021), Raza
and Ding (2022)
Topics: Bhattacharjee et al. (2018), Ajao et al. (2019), Be-
namira et al. (2019), Li et al. (2019)
Table 2 . 1: Some of the main text representations for misinformation detection
▶
Similar to Zannettou et al. (2019), Zubiaga et al. (2018) provide a compre-
hensive overview of research in this field, specifically focusing on rumours on
OSN
s. They categorise rumour classification architectures into four main types:
rumour detection,rumour tracking,stance classification, and veracity classification.
Additionally, they discuss examples of scientific approaches taken and datasets
used by researchers to tackle each task—along with the state-of-the-art method
for each task.
This research is primarily concerned with misinformation detection using
machine learning. The availability of data is a prerequisite to achieving this goal.
Furthermore, there are different
ML
approaches that can and have been used
2.3 use of text in ml-based misinformation detection research 24
to solve this problem. In this section, existing datasets and
ML
approaches for
misinformation detection are reviewed.
This thesis extends Zubiaga et al. (2018) by further categorising the
ML
ap-
proaches cited in it—and incorporating those cited in other papers supervised,
semi-supervised and unsupervised, as is laid out in Table 2.2. It also expands on
the applicable datasets for the respective tasks cited in Zubiaga et al. (2018). Their
work focuses on rumours, while this research targets the broader ecosystem of
2.3 use of text in ml-based misinformation detection research 25
misinformation. Note that the information in Table 2.2 does not constitute an
exhaustive list of published research papers or datasets in each category.
ml
approa c h
misinformation
detection
misinformation
tracking
sta nce
classification
veracity
classification
Supervised Wu et al. (2015),
Zubiaga et al.
(2016), Ahmed et al.
(2017), Ruchansky
et al. (2017), Wang
et al. (2018), Wu and
Liu (2018), Zhang
et al. (2020)
Castillo et al. (2011),
Ruchansky et al.
(2017), Wang et al.
(2017)
Kochkina et al.
(2017), Shang et al.
(2018)
Castillo et al. (2011),
Kwon et al. (2017)
Semi-
supervised
Bhattacharjee et al.
(2018), Guacho et al.
(2018), Shu et al.
(2019)
— — —
Unsupervised
Chen et al. (2016),
Zhang et al. (2016),
Zhang et al. (2017),
Chen et al. (2018),
Hosseinimotlagh
and Papalexakis
(2018)
— — —
dataset Mitra and Gilbert
(2015), Zubiaga et al.
(2016), Zubiaga et al.
(2016b), Zubiaga
et al. (2016c), Kwon
et al. (2017),
Kochkina et al.
(2018), Shu et al.
(2018), Rubin (2019)
Kochkina et al.
(2018)
Zubiaga et al.
(2016c),
Mohammad et al.
(2016), Mohammad
et al. (2017),
Kochkina et al.
(2018), Gorrell et al.
(2019)
Zubiaga et al. (2016c),
Kwon et al. (2017),
Kochkina et al. (2018),
Gorrell et al. (2019),
Rubin (2019), Arslan
et al. (2020)
Table 2 . 2: A breakdown of existing ML architectures for misinformation classification.
▶
The literature on misinformation detection has been mostly focused on
supervised learning. Castillo et al. (2011) were among the earliest to evaluate the
veracity of
OSN
content using supervised learning. Their objective was to assess
how believable tweets about global news events were over two months. They
generated a dataset of 747 tweets, manually labelled (‘true’ or ‘false’) by expert
2.3 use of text in ml-based misinformation detection research 26
judges. Extracted features were topic-based (e.g. textual length and sentiment of
tweet), network-based (e.g. the number of users’ followers), propagation-based
(e.g. total number of tweets) and top-element (e.g. fraction of tweets containing the
most popular hashtag). They tried four different supervised
ML
methods including
SVMsand Bayes networks, but Decision Trees yielded the highest accuracy.
Ruchansky et al. (2017) created a deep learning model to detect fake news,
using Twitter and Weibo data. It consists of three modules: Capture,Score and
Integrate. The Capture module is built using a
RNN
which represents the temporal
dynamics of a user’s activities, and a
doc2vec
representation
83
of text posted
83
Le and Mikolov (2014), “Dis-
tributed Representations of Sen-
tences and Documents”
therein. In the Score module, a neural network assigns a score to a user, based
on their tendency of being the source of a fake news article. The third module
combines information from the first two to classify the article. Supervised
ML
has also been used to detect rumours by analysing how they propagate. Wu et al.
(2015) achieved this using an
SVM
classifier, while Wu and Liu (2018) used
RNN
s.
▶
Given that in real-world scenarios, labelled data is—at least immediately—lacking,
some have tried to eliminate this restraint. Shu et al. (2019) proposed a novel semi-
supervised approach, which models the interrelationship between the contents,
publishers, and users (consumers) of news items (of which some are labelled).
It predicts the unlabelled news items, using features extracted from the news
articles, social relations between users, users’ engagements with the news arti-
cles, and publishers’ partisan associations. They collated fact-check data from
BuzzFeed
84
, PolitiFact
85
and Media Bias/Fact-Check
86
, into two new datasets
87 84 https:
//github.com/BuzzFeedNews/
2016-10- facebook-fact-
check/tree/master/data
85
https://www.politifact.com/
factchecks
86 https:
//mediabiasfactcheck.com
, which both included information on news contents, publishers and social in-
87
Shu et al. (2018), “FakeNews-
Net: A Data Repository with News
Content, Social Context and Dy-
namic Information for Studying
Fake News on Social Media”
teractions. They simplified the embeddings of their features using Non-negative
Matrix factorisation (
NMF
) and devised an optimisation algorithm to classify the
news articles.
Bhattacharjee et al. (2018) used active learning to detect the veracity of news,
using partially labelled datasets. Their system comprises two simultaneously
running, independent modules. The first module
𝑀1
begins with a Logistic Re-
gression classifier and a copy of the labelled dataset. It selects and assigns weights
to features by iteratively computing the Joint Mutual Information Maximisation
between features and class labels, and gives higher weights to the most relevant
ones in a greedy way.
𝑀1
’s dataset is updated to include the assigned weights,
and the classifier is retrained. The second module
𝑀2
begins with a copy of
the unlabelled and labelled dataset. The latter was used to train an underlying
classification model which is based on a
CNN
. Both modules iteratively classify
each unlabelled sample, and they request labels from a human if their predictions
do not attain a preset certainty threshold.
𝑀1
and
𝑀2
update their training sets
to include the given labels, and then fine-tune their classification models. Finally,
the predictions from both modules are combined into a decision profile and a
fusion classifier was used to make a final decision on a sample.
The advantage of unsupervised learning is neither labelled data nor human
input is needed. Zhang et al. (2016) considered fake news detection as an outlier
2.4 limitations of existing methods 27
detection problem. The rationale behind this is that the behaviours (related to
style and timing) of a user when posting rumours and non-rumours will differ.
Thus, rumours can be picked up as outliers in the user’s feed. They used Principal
Component Analysis (
PCA
) to detect rumours on Weibo. They initially collected
verified rumour and non-rumour posts for analysis, to determine relevant fea-
tures. The 13 selected features were numerical and categorical. When a post is
flagged as a rumour, their model collects a set of
𝑁
recent posts (between 10 and
100) by the poster and extracts the aforesaid features from them. The model then
performs
PCA
which transforms the
𝑁
posts into a matrix with
𝑁
rows (posts,
the first of which denotes the original flagged post) and eight columns containing
quantitative values. Eight was analytically chosen as the optimal number of pri-
mary components as it is the smallest number which captured at least 85% of the
total variance in the recent posts, using varying
𝑁
sample sizes. The original post
is considered an outlier (i.e., a rumour) if it does not have at least zero neighbours
within a given distance: calculated as the mean distance between pairs of posts
divided by the standard deviation.
2.4 limitations of existing methods
Given the significance of information disorder, a lot of work has been done to
address many of its subproblems. However, some limitations remain unsolved.
The following are some limitations related to this thesis:
1.
One of the open challenges in contemporary fake news research is the
lack of cross-domain, cross-topic, and cross-language studies.
88
This thesis
88
Zafarani et al. (2019), Zhou and
Zafarani (2020)
partly addresses this limitation through the use of cross-domain datasets,
that cover several different news topics, for fake news detection.
2.
Although extensively used to engineer features for fake news detection,
89
stylometric features are ineffective for distinguishing between genuine
89 See §2.3.
news and fake news autogenerated by language models.
90
This limitation
90
Schuster et al. (2020), “The Limi-
tations of Stylometry for Detecting
Machine-Generated Fake News”
may be overcome by exploring features which transcend stylometry, such
as topics, which are used in this work.
3.
Large amounts of labelled data are needed to create accurate models, as
observed by Wu and Liu (2018) and Wang (2017b). This is a motivation for
using unsupervised
ML
, in which case a labelled dataset is not a prerequisite.
The process of manually annotating datasets can be costly and very time-
consuming. Furthermore, while some authors have employed Amazon
Mechanical Turk workers to annotate their datasets, others
91
preferred to
91
Castillo et al. (2011), Mitra and
Gilbert (2015), Vosoughi et al.
(2018), Zhang et al. (2018)
use trained annotators, claiming that they made more informed judgements
on the veracity of examples. This ascribes an element of doubt to the
reliability of manually labelled datasets.
Part II
CONTRIBUTIONS
3
WORD EMBEDDINGS FOR MISINFORMATION
DETECTION
3.1 backgro u n d
One of the prevailing approaches to solving tasks in
NLP
is based on the hypothesis
that words which appear in close proximity tend to have a similar meaning.
92
This
92
Levy and Goldberg (2014), “Neu-
ral Word Embedding As Implicit
Matrix Factorization”
is known as the distributional hypothesis and was originally posited in 1954.
93
The
93
Harris (1954), “Distributional
Structure”
distributional hypothesis has since led to the development of various methods
to encode text in numeric form. Building on it, distributed representations for
computing elements were introduced by Hinton et al. (1986) about three decades
later. They were among the first to create numerical representations of words.
More recent approaches in
NLP
problem-solving are based on neural net-
work word embedding.
94
These models are constructed using neural nets that
94
Collobert and Weston (2008),
Mikolov et al. (2013)
represent the similarity between words using dense continuous, real vectors of
numbers. Today, the representations of words as vectors are generally referred to
as word embeddings.
95
Semantically similar words will have numerically similar
95
Levy and Goldberg (2014), “Neu-
ral Word Embedding As Implicit
Matrix Factorization”
vectors—known as learned distributed feature vectors—which ideally have a
much smaller dimension than the vocabulary. Embeddings can also be generated
for whole sentences by aggregating word embeddings or using neural nets trained
specifically for this task. When the dimensions of the learned vectors of words
are reduced to two or three and visualised on a Cartesian plane, relationships
between them become apparent. In this chapter, experiments were set up using
word and sentence embeddings to find semantic differences between reactions to
rumours and non-rumours. The experimental procedure, results, and conclusions
are explained in the following subsections.
3.2 related wor k
Word embedding models have shown a better performance than classical methods,
such as
BoW
, Term Frequency-Inverse Document Frequency and Distributional
Embeddings.
96
However, according to Goldberg and Levy (2014), it remains
96
Mikolov et al. (2011), Mikolov
et al. (2013)
unknown exactly why some models produce good word representations. Word
embeddings have been used effectively for fake news detection, as demonstrated
29
3.3 pr o blem defi n i t ion 30
by Bhattacharjee et al. (2018), Shang et al. (2018) and Shu et al. (2019). They have
also been used for automated fact-checking.97 97
Konstantinovskiy et al. (2018),
“Towards Automated Factchecking:
Developing an Annotation Schema
and Benchmark for Consistent Au-
tomated Claim Detection”
Mikolov et al. (2013) observed that a limitation of word representations is
their inability to capture idioms. For example, the phrase ‘Washington Post’ refers
to a newspaper, and its meaning is not directly deducible by simply combining
the individual meanings of the words ‘Washington’ and ‘Post’. They, therefore,
suggest using a Skip-gram model
98
to learn vector representations, as such a
98
Mikolov et al. (2013b), “Efficient
Estimation of Word Representa-
tions in Vector Space”
model is capable of representing phrases as vectors and is highly efficient, in
terms of training time and accuracy. Nonetheless, word embeddings are now
established and have been successfully applied to improve performance in various
NLP
tasks.
99
Examples of open-sourced word embedding tools include
word2vec 99
Collobert et al. (2011), “Natu-
ral Language Processing (Almost)
from Scratch”
by Mikolov et al. (2013), Global Vectors for Word Representation (or
GloVe
)
by Pennington et al. (2014), and
FastText
by Bojanowski et al. (2016). Word
embeddings have effectively been used for fake news detection. Some papers
which used these latent text features are listed in Table 2.2, in §2.3.
▶
False news and rumour content tend to be semantically distinct from authentic
or non-rumour content.
100
The observation that the semantics of the two tend
100
Parikh and Atrey (2018), Pot-
thast et al. (2018)
to differ partly motivates this investigation. Additionally, Choi et al. (2020) found
that echo chambers tend to increase virality and accelerate the spread of rumours.
They define an echo chamber as a collection of users that have shared at least
two rumours in common. They analysed more than one hundred rumours from
six fact-checking platforms. These rumours were the subject of nearly 300,000
tweets made by over 170,000 users. Therefore, it can be argued that those who
retweet rumours are likely to be more driven to amplify a common message,
than those who retweet non-rumour tweets. This amplification may also be in
the form of replies that express agreement and may, therefore, be semantically
similar.
The experiment presented in this chapter differs from some previous studies
101
—it focuses not on the rumours or non-rumours posted, but on the reactions
101
Wu et al. (2015), Zhang et al.
(2016), Zhang et al. (2017)
which they attract. The goal here is to find out whether there is a difference in
dispersion between people’s reactions to rumours and non-rumour in tweets.
Note that in this section, ‘reactions’ and ‘comments’ refer to the replies received
by tweets.
3.3 problem definition
In this experiment, the aim is to determine whether or not there is any evidence to
differentiate between rumours and non-rumours tweets, based on the reactions
they receive. Latent text representations are used as the discriminant between
the two groups. This study is carried out using statistical hypothesis testing. The
hypotheses can be stated as follows:
3.4 methodology and materials 31
Hypothesis H
0
(Null):the semantic similarities between rumour and non-rumour
tweet reactions are equal.
Hypothesis H
1
(Alternative):the semantic similarities of rumour tweet reactions
are greater than those of non-rumour tweet reactions.
For both Hypotheses H
0
and H
1
, the semantic similarities are measured using
InferSent102
sentence embeddings. It is expected that there will be a greater
102
Conneau et al. (2017), “Super-
vised Learning of Universal Sen-
tence Representations from Natu-
ral Language Inference Data”
similarity amongst rumour tweet reactions, compared with non-rumours ones.
This is in line with the aforementioned observations. The method through which
Hypothesis H
1
will be tested against Hypothesis H
0
is explained in the next
section.
3.4 methodology and materials
3.4.1 Experimental procedure
Let
𝑃𝐴𝐿𝐿 ∈𝑃𝑅, 𝑃 𝐹
represent a dataset containing all
𝑀
rumour
𝑁
and non-
rumour (factual) posts,
𝑃𝑅=𝑝𝑅
1, . . . , 𝑝𝑅
𝑀
and
𝑃𝐹=𝑝𝐹
1, . . . , 𝑝𝐹
𝑁
, respectively.
Each rumour post
𝑝𝑅
𝑖
has received comments
𝑐𝑅
𝑖=h𝑐𝑅
𝑖,1, 𝑐𝑅
𝑖,2, . . . , 𝑐𝑅
𝑖,𝑚𝑖i
, where
𝑚𝑖
is the total number of comments that follow, and
𝑖={1. . . 𝑀 }
. Likewise,
each factual tweet
𝑝𝐹
𝑖
has received
𝑐𝐹
𝑖=h𝑐𝐹
𝑖,1, 𝑐𝐹
𝑖,2, . . . , 𝑐𝐹
𝑖,𝑛𝑖i
comments (a total
of
𝑛𝑖
), with
𝑖={1. . . 𝑁 }
. Therefore,
𝐶𝑅=
𝑀
Ð
𝑖=1
𝑐𝑅
𝑖
and
𝐶𝐹=
𝑁
Ð
𝑖=1
𝑐𝐹
𝑖
are all the
reactions to rumours and non-rumours, respectively. Algorithm 1summarises
the computations for this experiment.
Before the experiment, the data was cleaned as follows: (i.) all datasets were
cleaned to remove usernames and hashtags; (ii.) comments less than three words
long were removed.
103
Next, a pre-trained
InferSent
word embedding model
103
This helps to make the word
embedding more accurate. A post
may have only a single comment,
but all comments must be more
than three words long, or else that
post is excluded.
was used to generate 4096-length vectors for each rumour and non-rumour
reaction. This gives us matrices for rumour and non-rumour embeddings,
𝐸𝑅=
(𝑒𝑖,𝑗 ) ∈ R𝑀×4096 and 𝐸𝐹=(𝑒𝑖, 𝑗) ∈ R𝑁×4096 , respectively.
▶
The average pairwise cosine similarities,
𝑐𝑠𝑎𝑣𝑔
, between the embeddings for
rumour and non-rumour reactions are separately calculated. Self-comparisons
between items (having a similarity of
1
) are excluded. Therefore, an
𝑙×4096
matrix of embeddings is inputted and the output is a
𝑙
-length vector in return,
104 104
The row-wise mean is calcu-
lated first in Line 6 of Algorithm
1to compare each comment with
every other comment except itself.
after finding the mean. In this vector, each item is the mean of cosine distances
between the embedding of a comment and all other comments.
Lastly, the average of each vector is calculated, as
𝐴𝑅
and
𝐴𝐹
, and the difference
between the two is found as
Δ𝐴=𝐴𝑅−𝐴𝐹
. The higher
Δ𝐴
is, the more similar
rumours are as compared with non-rumours, and vice versa.
InferSent
is trained on natural inference data and it generates semantic
representations for sentences in English.
105
Embeddings of phrases and sentences
105
Conneau et al. (2017), “Super-
vised Learning of Universal Sen-
tence Representations from Natu-
ral Language Inference Data”
3.4 methodology and materials 32
Algorithm 1 Comparison of rumour and non-rumour comments using
InferSent embeddings
Input: Comments 𝐶𝑅and 𝐶𝐹
Output: Δ𝐴
1: function InferSentEmbed(𝑡𝑒𝑥𝑡)
2: return InferSent embedding vector for 𝑡𝑒𝑥𝑡,e⊲|e|=4096
3: end function
4: function PairwiseCosSim(𝑋=(𝑒𝑖,𝑗 ) ∈ R𝑙×4096 )
5: Pairwise cosine similarity matrix for 𝑋is 𝑆=(𝑎𝑖,𝑗 ) ∈ R𝑙×𝑙
6: return 𝑚𝑒𝑎𝑛(𝑆)⊲do row-wise mean first
7: end function
8: for all 𝑐𝑖∈𝐶𝐹do
9: 𝐸𝐹
𝑖=InferSentEmbed(𝑐𝑖)
10: end for
11: for all 𝑐𝑖∈𝐶𝑅do
12: 𝐸𝑅
𝑖=InferSentEmbed(𝑐𝑖)
13: end for
14: 𝐸𝐹=𝐸𝐹
1, . . . , 𝐸𝐹
𝑀=(𝑒𝑖,𝑗 ) ∈ R𝑀×4096
15: 𝐸𝑅=𝐸𝑅
1, . . . , 𝐸𝑅
𝑁=(𝑒𝑖,𝑗 ) ∈ R𝑁×4096
16: 𝐴𝐹=PairwiseCosSim(𝐸𝐹)
17: 𝐴𝑅=PairwiseCosSim(𝐸𝑅)
18: Δ𝐴=𝐴𝑅−𝐴𝐹
19: return Δ𝐴
3.4 methodology and materials 33
are typically obtained by averaging their constituting word embedding vectors.
However,
InferSent
is advantageous because it takes the order of words into
account to produce embeddings for whole sentences, as it is built on a RNN.106 106
Conneau et al. (2017), Kon-
stantinovskiy et al. (2018)
3.4.2 Datasets
The PHEME dataset created by Zubiaga et al. (2016b) was used to evaluate Hy-
potheses H
0
and H
1
. It contains nearly 6,000 tweets concerning five fatal incidents
that occurred in North America and Europe between 2014 and 2015. A break-
down of this dataset is shown in Table 3.1.
event rumours
non -
ru m o u r s
ru m o u r
comments
non - r u m o u r
comments
Charlie Hebdo
Shooting (Jan.
2015)
458 (22%) 1620 (78%) 422 (22.0%) 1493 (88%)
Ferguson Unrest
(Aug. 2014)
284 (24.8%) 231 (75.2%) 257 (25.8%) 740 (74.2%)
Germanwings
Crash (Mar.
2015)
238 (50.7%) 231 (49.3%) 160 (49.0%) 166 (51.0%)
Ottawa
Shooting (Oct.
2014)
470 (52.8%) 420 (47.2%) 426 (54.2%) 360 (45.8%)
Sydney Siege
(Dec. 2014)
522 (42.8%) 699 (57.2%) 486 (42.6%) 656 (57.4%)
Table 3 . 1: Breakdown of PHEME dataset
3.4.3 Results and discussion
The tendency for rumour and non-rumour content to differ semantically was
introduced in
§
3.2. This experiment hypothesises that rumour reactions will
generally be similar to each other—rather than to non-rumour reactions—and
vice versa. Similarity, here, is evaluated by computing and comparing the sentence
embeddings of the two groups of tweets. It is expected, therefore, that the mean
of the pairwise distances between rumours will generally be greater than that
between non-rumour comments.
To verify this scientifically, a statistical test was carried out on the experimental
results. The differences in the similarities of rumour and non-rumour reactions
3.4 methodology and materials 34
were evaluated using the Wilcoxon–Mann–Whitney test, at 5% significance level.
This test was chosen because the resulting data, for all datasets, did not pass the
test for normality, and therefore, could not be assumed to be normally distributed.
A Shapiro-Wilk test showed that the distribution of rumour and non-rumour
similarity values departed significantly from a normal distribution (R — rumour,
NR — non-rumour):
•Charlie Hebdo: R (𝑊=0.971, 𝑝 <0.01), NR (𝑊=0.915, 𝑝 <0.01)
•Ferguson: R (𝑊=0.890, 𝑝 <0.01), NR (𝑊=0.915, 𝑝 <0.01)
•Germanwings: R (𝑊=0.929, 𝑝 <0.01), NR (𝑊=0.970, 𝑝 <0.01)
•Ottawa: R (𝑊=0.914, 𝑝 <0.01), NR (𝑊=0.901, 𝑝 <0.01)
•Sydney: R (𝑊=0.918, 𝑝 <0.01), NR (𝑊=0.883, 𝑝 <0.01)
Therefore, the differences between the median similarities, rather than the
mean, were conclusively analysed, by applying the Wilcoxon–Mann–Whitney
test.
Table 3.2 summarises the results of this experiment, while detailed plots are
presented in Figure 3.1. Recall from Algorithm 1, that
𝐴𝑅
and
𝐴𝐹
are the mean
similarities between rumour and non-rumour (factual) tweet reactions, respec-
tively.
𝑀𝑒𝑑𝑅
and
𝑀𝑒𝑑𝐹
(
Δ𝑀𝑒𝑑 =𝑀𝑒𝑑𝑅−𝑀𝑒𝑑𝐹
) are the median similarities
between rumours and non-rumours, respectively.
event 𝐴𝑅𝐴𝐹Δ𝐴 𝑀 𝑒 𝑑𝑅𝑀 𝑒 𝑑 𝐹Δ𝑀 𝑒 𝑑 𝑝 -𝑣𝑎 𝑙𝑢𝑒
Charlie
Hebdo
0.603 0.599 0.004 0.603 0.608 -0.005 0.487
Ferguson
0.632 0.617 0.015 0.629 0.618 0.012 1.686 ×10-4
German-
wings
0.603 0.593 0.001 0.608 0.599 0.009 0.114
Ottawa
0.615 0.613 0.002 0.614 0.604 0.009 0.136
Sydney
0.606 0.614 -0.008 0.607 0.616 -0.009 0.998
Table 3 . 2: Summary of experimental and statistical results for comparisons between
sentence embeddings.
The results show that there are no significant, consistent differences between
rumours and factual comments when comparing the two using sentence embed-
dings. Most values of
Δ𝐴
are positive as expected, except for the Sydney Siege
dataset. The
Δ𝑀𝑒𝑑
values are also positive, except for the Charlie Hebdo and
Sydney Siege datasets. The two measures may suggest that the rumour reactions
3.4 methodology and materials 35
◦◦ ◦ ◦◦◦ ◦ ◦◦ ◦◦◦◦◦◦◦◦ ◦◦ ◦ ◦◦◦◦ ◦◦ ◦◦ ◦◦◦ ◦◦ ◦◦◦◦ ◦◦◦◦◦◦◦ ◦◦ ◦◦◦◦ ◦ ◦◦ ◦◦ ◦×× ×× ××× ×× ×× × × × ×× ×× ××
◦ ◦◦◦◦ ◦◦◦ ◦ ◦◦◦ ◦◦ ◦◦ ◦◦ ◦ ◦◦ ◦◦◦ ◦◦◦◦ ◦×× × ×× ×××
0.0 0.2 0.4 0.6 0.8 1.0
non-rumours
rumours
Distribution of csavg for Charlie Hebdo
rumour mean: 0.603
non-rumour mean: 0.599
◦ ◦◦◦◦ ◦◦◦◦ ◦◦ ◦◦ ◦◦◦◦ ◦◦ ◦◦◦ ◦ × ×× ×× ×
◦ ◦ ◦◦ ◦◦◦ ◦◦ ◦◦ ◦ ◦◦◦◦◦× × ×× ×× ×× ×××× ××
0.0 0.2 0.4 0.6 0.8 1.0
non-rumours
rumours
Distribution of csavg for Ottawa Shooting
rumour mean: 0.615
non-rumour mean: 0.613
◦◦ ◦◦◦◦◦◦◦◦ ◦◦ ◦◦◦ ◦◦ ◦◦ ◦◦◦◦◦◦ ◦◦◦ ◦ ◦◦ ◦◦ ◦◦◦× ×× ×× ×× ×× ××××× ×
◦◦ ◦◦ ◦◦◦ ◦◦
0.0 0.2 0.4 0.6 0.8 1.0
non-rumours
rumours
Distribution of csavg for Ferguson
rumour mean: 0.632
non-rumour mean: 0.617
◦◦ ◦◦ ◦◦ ◦◦ ◦◦ ◦ ◦ ◦◦◦ ◦◦◦◦◦ ◦◦ ◦× ×× ×× × ××× ×
◦◦◦◦ ◦ ◦◦◦ ◦◦ ◦◦◦ ◦◦ ◦◦◦◦ ◦× × × ××
0.0 0.2 0.4 0.6 0.8 1.0
non-rumours
rumours
Distribution of csavg for Sydney Siege
rumour mean: 0.606
non-rumour mean: 0.614
◦◦◦◦ ◦ ◦◦ ◦◦
◦ ◦◦◦ ◦◦◦ ◦× ××
0.0 0.2 0.4 0.6 0.8 1.0
non-rumours
rumours
Distribution of csavg for Germanwings Crash
rumour mean: 0.603
non-rumour mean: 0.593
Figure 3.1: Box plots of distributions of 𝑐𝑠𝑎𝑣𝑔 for rumour and non-rumour reactions.
are generally semantically more similar to each other than non-rumours, but
Δ𝐴
and
Δ𝑀𝑒𝑑
indicate that such a conclusion cannot be made. Except for the
Ferguson dataset, the null hypothesis (Hypothesis H
0
)—that the average semantic
similarities between rumour and non-rumour tweet reactions are equal—is not
rejected at the 5% level, based on the Wilcoxon–Mann–Whitney test. Therefore,
in summary, the Hypothesis H0tested in this experiment is not rejected.
There are some possible factors which may have affected the outcome of this
experiment. First, sentence embeddings work better with well-written sentences.
However, short texts of just 140 characters—not necessarily forming complete
sentences—were used here. Second, the datasets used are concerned with separate
events in different countries, which may have been discussed in different ways. For
example, the Charlie Hebdo event occurred in France, and some of the tweets in
that dataset are in French. Similarly, the Germanwings Crash data contains some
tweets in German. However, an
InferSent
model for English texts was used
to get the sentence embeddings, as most of the tweets are in English. Lastly, the
small size of the dataset, coupled with the imbalance of the number of examples
in each class, possibly influenced the outcome of this experiment.
3.4 methodology and materials 36
▶
Further experiments were conducted with minor changes made to the method-
ology. In one of the follow-up studies, the aim was to determine whether rumour
posts are more similar to the comments they attract, compared with non-rumours.
In other words, the goal is to compare the semantic differences between rumours
and their reactions, and non-rumours and their comments. The following steps
were carried out, for each event:
•Embed each post (rumour or non-rumour) into a 300-length vector.
•Embed its corresponding comments also into 300-length vectors.
•
Find the averages of pairwise cosine similarities and Euclidean distances
between each post and its set of comments. Here,
𝑐𝑠𝑎𝑣𝑔
represents the
similarity between a post and the comments it attracted.
The results of this experiment (see Figure 3.2,Figure 3.3 and Table 3.3) also did
not yield conclusive results. They show that for some events, the rumour posts
are more similar to their comments, while the opposite is the case for others.
Even when the embeddings of the comments are averaged before being compared
with the embeddings of their original post, significant differences were not found
between rumour and non-rumour tweets.
event euclidean distance
Charlie Hebdo Shooting -1.365
Ferguson Unrest 0.924
Germanwings Crash 1.156
Ottawa Shooting 0.076
Sydney Siege -1.608
Table 3 . 3: Difference between the averages of the Euclidean distances of rumour and
non-rumour comments.
3.4 methodology and materials 37
0-0.1
0.1 -0.2
0.2 -0.3
0.3 -0.4
0.4 -0.5
0.5 -0.6
0.6 -0.7
0.7 -0.8
0.8 -0.9
0.9 -1
0
20
40
60
80
100
Percent
Distribution of csavg (post-comment)forCharlie Hebdo
NR R
0-0.1
0.1 -0.2
0.2 -0.3
0.3 -0.4
0.4 -0.5
0.5 -0.6
0.6 -0.7
0.7 -0.8
0.8 -0.9
0.9 -1
0
20
40
60
80
100
Percent
Distribution of csavg (post-comment)forOttawa Shooting
NR R
0-0.1
0.1 -0.2
0.2 -0.3
0.3 -0.4
0.4 -0.5
0.5 -0.6
0.6 -0.7
0.7 -0.8
0.8 -0.9
0.9 -1
0
20
40
60
80
100
Percent
Distribution of csavg (post-comment)forFerguson
NR R
0-0.1
0.1 -0.2
0.2 -0.3
0.3 -0.4
0.4 -0.5
0.5 -0.6
0.6 -0.7
0.7 -0.8
0.8 -0.9
0.9 -1
0
20
40
60
80
100
Percent
Distribution of csavg (post-comment)forSydney Siege
NR R
0-0.1
0.1 -0.2
0.2 -0.3
0.3 -0.4
0.4 -0.5
0.5 -0.6
0.6 -0.7
0.7 -0.8
0.8 -0.9
0.9 -1
0
20
40
60
80
100
Percent
Distribution of csavg (post-comment)forGermanwings Crash
NR R
Figure 3.2: Distributions of average pairwise cosine similarities between posts and
their comments. 𝑁 𝑅 =non-rumours, 𝑅 =rumours.
3.4 methodology and materials 38
0-1
1-2
2-3
3-4
4-5
5-6
6-7
7-8
8-9
9-10
0
20
40
60
80
100
Percent
Distribution of avg. euclidean dist. (post-comment)forCharlie Hebdo
NR R
0-1
1-2
2-3
3-4
4-5
5-6
6-7
7-8
8-9
9-10
0
20
40
60
80
100
Percent
Distribution of avg. euclidean dist. (post-comment)forOttawa Shooting
NR R
0-1
1-2
2-3
3-4
4-5
5-6
6-7
7-8
8-9
9-10
0
20
40
60
80
100
Percent
Distribution of avg. euclidean dist. (post-comment)forFerguson
NR R
0-1
1-2
2-3
3-4
4-5
5-6
6-7
7-8
8-9
9-10
0
20
40
60
80
100
Percent
Distribution of avg. euclidean dist. (post-comment)forSydney Siege
NR R
0-1
1-2
2-3
3-4
4-5
5-6
6-7
7-8
8-9
9-10
0
20
40
60
80
100
Percent
Distribution of avg. euclidean dist. (post-comment)forGermanwings Crash
NR R
Figure 3.3: Distributions of average pairwise Euclidean distances between posts and
their comments. 𝑁 𝑅 =non-rumours, 𝑅 =rumours.
3.5 disparities in sentiment 39
3.5 disparities in sentiment
As mentioned earlier (see
§
1.2), authors of false news sometimes seek to arouse
emotional responses from readers, as has been observed through studies of their
writing style. This observation served as the basis for an experiment which aimed
to distinguish between rumours and non-rumours by analysing the sentiment
expressed in both sets of tweets.
In this experiment, the Stanford NLP tool
107
was used to analyse the sentiment
107
Socher et al. (2013), “Recursive
Deep Models for Semantic Compo-
sitionality Over a Sentiment Tree-
bank”
scores of posts and comments in the PHEME dataset. For a given text, the tool
computes one of the following sentiment scores: 0 (very negative), 1 (negative), 2
(neutral), 3 (positive), or 4 (very positive). In the first variant of this experiment,
the sentiment scores of the posts and comments of rumour and non-rumour
tweets were computed and compared.
The results (plotted in Figure 3.4 and Figure 3.5) do not show significant varia-
tions between the sentiment scores of rumours and non-rumours, for posts or
comments. Further analyses were carried out but the results did not significantly
distinguish between rumour and factual comments or posts based on inferred
sentiment.
Sentiment
0 1 2 3 4
0
20
40
60
80
100
Percent
Average post sentiment score for Charlie Hebdo
NR R
Sentiment
0 1 2 3 4
0
20
40
60
80
100
Percent
Average post sentiment score for Ottawa Shooting
NR R
Sentiment
0 1 2 3 4
0
20
40
60
80
100
Percent
Average post sentiment score for Ferguson
NR R
Sentiment
0 1 2 3 4
0
20
40
60
80
100
Percent
Average post sentiment score for Sydney Siege
NR R
Sentiment
0 1 2 3 4
0
20
40
60
80
100
Percent
Average post sentiment score for Germanwings Crash
NR R
Figure 3.4: Average sentiment scores of posts. 𝑁 𝑅 =non-rumours, 𝑅 =rumours.
▶
One final experiment was performed regarding text embeddings and sen-
timent. In it,
𝐾
-Means clustering (with
𝐾=2
) was used to analyse the senti-
ment scores for rumour and non-rumour comments. This was repeated with
3.6 concl u s ion 40
Sentiment
0 1 2 3 4
0
20
40
60
80
100
Percent
Average comment sentiment score for Charlie Hebdo
NR R
Sentiment
0 1 2 3 4
0
20
40
60
80
100
Percent
Average comment sentiment score for Ottawa Shooting
NR R
Sentiment
0 1 2 3 4
0
20
40
60
80
100
Percent
Average comment sentiment score for Ferguson
NR R
Sentiment
0 1 2 3 4
0
20
40
60
80
100
Percent
Average comment sentiment score for Sydney Siege
NR R
Sentiment
0 1 2 3 4
0
20
40
60
80
100
Percent
Average comment sentiment score for Germanwings Crash
NR R
Figure 3.5: Average sentiment scores of comments.
𝑁𝑅 =non-rumours, 𝑅 =rumours.
InferSent
, as well as
word2vec
(300-length vectors) word embeddings, instead
of sentiment scores. The results from clustering showed that there is no clear
distinction between rumour and non-rumour tweets.
3.6 conclusion
The series of experiments discussed in this section do not conclude that word
or sentence embeddings, or sentiment, can reliably distinguish rumour tweets
from factual ones (at least, in the datasets used here), or the reactions that either
receive. Nonetheless, these findings are limited, and probably only apply, to the
set-up of the experiments presented here. Others have successfully used these
text representations in different ways to differentiate between the two types of
tweets. To further test the methodology followed here will require substantially
larger datasets. Furthermore, as opposed to using simplistic sentiment measures
(positive, neutral, and negative), a more granular and precise measure of emotions
in tweets can be explored. For example, Vosoughi et al. (2018) examined a range of
positive emotions (such as joy and trust) as well as negative ones (such as fear and
anger) in both false and true news. They found that comments to false rumours
had greater surprise and disgust expressed in them, while reactions to true ones
expressed more sadness and anticipation. Similarly, Kolev et al. (2022) carried
3.6 concl u s ion 41
out fake news detection by using the predicted six emotions (anger, disgust, fear,
joy, sadness, and surprise) in the titles of news articles as features.
4
THEMATIC COHERENCE IN FAKE NEWS
4.1 backgro u n d
This chapter deals with the exploration of thematic coherence of fake news. News
‘The construction of life is at present
in the power far more of facts than
of convictions, and of such facts as
have scarcely ever become the basis
of convictions.’
— Walter Benjamin, “One-Way
Street”
readers are often enticed by the headlines of articles, or their opening sentence(s).
False news is written for many different reasons, including propaganda, provo-
cation and profit,
108
and therefore, often in catchy or emotive language. Given
108
Shu et al. (2017), “Fake News
Detection on Social Media”
the deluge of information which competes daily for people’s attention, most
people would now skim through news pieces that they would otherwise carefully
read—perhaps to save their time—or in an attempt to spend it on stories of
greater interest to them. However, this inattention can be exploited by propa-
gators of misinformation, as they can make the headlines or openings of false
news captivating. An indication of a misleading article could, therefore, be that its
headline or starting paragraph thematically deviates from the rest of the article.
In this chapter, the focal point is fake news that appears in the form of long
online articles and explores the extent of internal consistency within fake news
vis-à-vis legitimate news. In particular, these experiments aim to determine
whether thematic deviations—i.e., a measure of how dissimilar topics discussed in
different parts of an article are—between the opening and remainder sections of
texts can be used to distinguish between fake and real news across different news
domains. Put simply, this is a measure of the distance between the distributions of
topics extracted from two sections of an article, the opening and the remainder.
The dissemination of fake news is increasing, and because it appears in various
forms and self-reinforces,
109
it is difficult to erode. Hence, there is an urgent need
109
Wardle (2017), Waldman (2018),
Zhou and Zafarani (2018)
for increased research in understanding and curbing it.
▶
One study by Gabielkov et al. (2016) found that, as of 2016, 59% of links shared
on OSNs have never been clicked before. This indicates that people share in-
formation without actually reading it. A more recent study by Anspach et al.
(2019) suggests that some readers may skim through an article instead of reading
the whole content because they overestimate their political knowledge, while
others may hastily share news without reading it fully, for emotional affirma-
tion. This presents bad actors with the opportunity to deftly intersperse news
content with falsity. Moreover, the production of fake news typically involves
the collation of disjointed content and lacks a thorough editorial process.
110
42
4.2 re l ated wo r k 43
The limitation of existing misinformation detection methods not adequately
110
Karimi and Tang (2019),
“Learning Hierarchical
Discourse-level Structure
for Fake News Detection”
capturing the subtle differences between false and legitimate news motivates the
experiments presented in this section.
▶
Topics discussed in news pieces can be studied to ascertain whether an article
thematically deviates between its opening and the rest of the story, or if it remains
coherent throughout. In other words, does an article open with one topic and
finish with a different, unrelated topic? Thematic analysis is useful here for two
reasons. First, previous studies show that the coherence between units of dis-
course (such as sentences) in a document is useful for determining its veracity.
111 111
Rubin and Lukoianova (2015),
Karimi and Tang (2019)
Second, analysis of thematic deviation can identify general characteristics of fake
news that persist across multiple news domains.
Topics have been employed as features for misinformation detection using
ML
.
112
However, they have not been applied to study the unique characteristics
112
Bhattacharjee et al. (2018), Be-
namira et al. (2019), Li et al. (2019)
of fake news. Research efforts in detecting fake news through thematic deviation
have thus far focused on spotting incongruences between pairs of headlines and
body texts.
113
Yet, thematic deviation can also exist within the body text of a
113
Chen et al. (2015), Ferreira
and Vlachos (2016), Sisodia (2019),
Yoon et al. (2019)
news item. The focus is to examine these deviations to distinguish fake from real
news.
To the best of the author’s knowledge, this is the first work that explores
thematic deviations in the body text of news articles to distinguish between fake
and legitimate news.
4.2 related wor k
The coherence of a story may be indicative of its veracity. For example, Rubin
and Lukoianova (2015) demonstrated this by applying
RST114
to study the dis-
114
Mann and Thompson (1988),
“Rhetorical Structure Theory: To-
ward a functional theory of text
organization”
course of deceptive stories posted online. They found that a major distinguishing
characteristic of deceptive stories is that they are disjunctive. Furthermore, while
truthful stories provide evidence and restate information, deceptive ones do not.
This suggests that false stories may tend to thematically deviate more due to dis-
junction, while truthful stories are likely to be more coherent due to restatement.
Similarly, Karimi and Tang (2019) investigated the coherence of fake and real
news by learning hierarchical structures based on sentence-level dependency
parsing. Their findings also suggest that fake news documents are less coherent.
▶
Topic models are unsupervised algorithms that aid the identification of themes
discussed in large corpora. With them, these texts can be understood, organized,
summarised and searched for automatically.
115
One example of topic models is
115
Blei (2012), “Probabilistic topic
models”
LDA
, which is a generative probabilistic model that aids the discovery of latent
themes or topics in a corpus.
116
Vosoughi et al. (2018) used
LDA
to show that false
116
Blei et al. (2003), “Latent Dirich-
let Allocation”
rumour tweets tend to be more novel than true ones. Novelty was evaluated
using three measures: Information Uniqueness, Bhattacharyya Distance, and
Kullback-Leibler Divergence. Likewise, Ito et al. (2015) used
LDA
to assess the
4.3 pr o blem defi n i t ion 44
credibility of Twitter users by analyzing the topical divergence of their tweets
from those of other users. They also assessed the veracity of users’ tweets by
comparing the topic distributions of new tweets against historically discussed
topics. Divergence was computed using the Jensen-Shannon Divergence, Root
Mean Squared Error, and Squared Error. This work primarily differs from those
two, in that here, full-length articles are analysed instead of tweets.
4.3 problem definition
Building on the previous subsections, the aim is to establish whether or not there is
evidence to distinguish between fake and authentic news, based on the coherence
of topics discussed in them. Similar to Chapter 3, the statistical hypothesis testing
approach is found to be appropriate for carrying out this study. The following
hypotheses are tested:
Hypothesis H
0
(Null):False and authentic news articles are similarly coherent
thematically.
Hypothesis H
1
(Alternative):the thematic coherence of authentic news articles is
greater than that of false news articles.
Specifically, the thematic drift between the opening part and the remaining
part of an article is measured, to see how they differ. The primary tool used to
measure this is
LDA
topic modelling. The opening section of an article is defined
using a hyperparameter,
𝑙
, which is the number of sentences at the start of it. To
test Hypotheses H
0
and H
1
, experiments are carried out in the manner outlined
in Algorithm 2. The differences in mean and median coherence values of fake and
real articles are evaluated using an Independent Samples T-test, at 5% significance
level.
4.4 research goal and contributions
The research presented in this chapter aims to assess the importance of internal
consistency within articles as a high-level feature to distinguish between fake
and real news stories across different domains. This chapter sets out to explore
whether the opening segments of fake news thematically deviate from the rest of
it, significantly more than in authentic news. Experiments are conducted using
seven datasets which collectively cover a wide variety of news domains, from
business to celebrity, to warfare. Deviations are evaluated by calculating the
distance between the topic distribution of the opening part of an article, to that
of its remainder. The first five sentences of an article are taken as its opening
segment.
The following summarise the contributions of this chapter:
4.5 methodology and materials 45
•
It presents new insights towards understanding the underlying character-
istics of fake news, based on thematic deviations between the opening and
remainder parts of news body text.
•
Experiments are carried out on five cross-domain misinformation datasets;
the results demonstrate the effectiveness of thematic deviation for distin-
guishing fake from real news.
4.5 methodology and materials
4.5.1 Latent Dirichlet Allocation
Given a text document, an
LDA
model generates words by selecting a topic from
the document-topic distribution, and then selecting a word from the topic-word
distribution.
117
A brief description of how
LDA
works is given here, following
117
Blei (2012), “Probabilistic topic
models”
the notation used by Maiya and Rolfe (2015) for its clarity.
Let
𝐷={𝑑1, . . . , 𝑑𝑁}
be a corpus, consisting of
𝑁
documents which collec-
tively cover
𝑇={𝑡1, . . . , 𝑡𝐾}
latent topics. Each document,
𝑑𝑖
, is made up of a
sequence of words. That is,
𝑑𝑖=⟨𝑤𝑖,1, 𝑤𝑖,2, . . . , 𝑤𝑖,𝑊𝑖⟩
, where
𝑖∈ {1. . . 𝑁 }
and
𝑊𝑖
is the total number of words in—called the vocabulary of—
𝑑𝑖
. Therefore, the
vocabulary of 𝐷is 𝒲=
𝑁
Ð
𝑖=1
𝑑𝑖.
In addition to some hyperparameters, probabilistic topic models such as
LDA
require only two inputs: (i.) corpus
𝐷
; and (ii.) desired number of topics
𝐾
. They
output two matrices: (i.) the document-topic distribution matrix,
𝜃∈R𝑁×𝐾
,
which represents the topics drawn from each document;
118
and (ii.) the topic-
118 This is the ‘Allocation’ in LDA.
word distribution matrix,
𝜙∈R𝐾×|𝒲|
, which represents the distribution of
words within each topic. The model assumes that each row in both matrices
is a Dirichlet probability distribution, hence its name. The optimal value of
𝐾
is typically found by iteration. If
𝐾
is overly high, the resulting topics may be
uninterpretable, and should ideally have been merged; and if it is too low, the
topics will be too broad, i.e., covering many differing concepts.119 119
Syed and Spruit (2018),
“Full-Text or abstract? Examining
topic coherence scores using latent
dirichlet allocation”
A topic found in a document
𝑑𝑖
is usually shown as a combination of a word
𝑤𝑖
and its probability
𝑝𝑖
in the distribution
𝜙𝑖
, as
(𝑝𝑖∗𝑤𝑖)
. For example,
(𝑓 𝑎𝑐𝑡 ∗0.01)
or
(𝑓 𝑎𝑘𝑒 ∗0.001)
. Each topic distribution contains the entire vocabulary, with
varying probabilities assigned to each word. The word with the highest probability
in the distribution is usually used to label a topic.
120
Words that have higher
120
Maiya and Rolfe (2015), “Topic
similarity networks: Visual analyt-
ics for large document sets”
probabilities within a topic would tend to co-occur in the corpus as a whole.
LDA
generates document-topic distributions
𝜃𝑑
and word-topic distributions
𝜙𝑡
.Figure 4.1
121
shows a graphical model of
LDA
. The box labelled
𝐷
represents
121 Adapted from Blei et al. (2003)
(Fig. 1) and Blei (2012) (Fig. 4).
the documents in a corpus. While boxes
𝒲
and
𝐾
represent the repeatedly
selected words and topics within a document, respectively. The circles are random
variables in the generative process. The Dirichlet parameter
𝛼
controls the sparsity
of topics within documents, while
𝛽
controls the sparsity of words within topics.
4.5 methodology and materials 46
Figure 4.1: Graphical representation of LDA
The hidden variables (topics, topic proportions, and assignments) are unshaded,
while the observed variable (words in a document) is shaded.
4.5.2 Distance measures
Distributional similarity/distance measures are commonly used to compare the
similarities, differences and overlaps between topics extracted from corpora.
122 122
Omar et al. (2015), Vosoughi et
al. (2018)
All articles
𝑆𝑏𝑔
are split into two parts: its first
𝑥
sentences
123
, and the remaining
123
Only articles with at least
𝑥+1
sentences are used.
𝑦
. Next,
𝑁
topics are obtained from
𝑥
and
𝑦
from an
LDA
model trained on
the entire dataset. For
𝑖=(1, . . . , 𝑚)
topics, let
𝑝𝑥=(𝑝𝑥1, . . . , 𝑝𝑥 𝑚)
and
𝑝𝑦=
(𝑝𝑦1, . . . , 𝑝 𝑦𝑚)
be two vectors of topic distributions, which denote the prevalence
of a topic
𝑖
in the opening text
𝑥
and remainder
𝑦
of an article, respectively.
Finally, the average and median values of each distance are calculated across all
fake (
𝑆𝑓
) and real (
𝑆𝑟
) articles. These steps were repeated with varying values of
𝑁(from 10 to 200 topics) and 𝑥(from 1 to 5 sentences).
The following are the data required for this procedure: a corpus
𝑆𝑏𝑔 =𝑆𝑓Ð𝑆𝑟
of full-length fake (
𝑆𝑓=n𝑑𝑓
1, 𝑑𝑓
2, . . . , 𝑑𝑓
𝐹o
) and real (
𝑆𝑟=𝑑𝑟
1, 𝑑𝑟
2, . . . , 𝑑𝑟
𝑅
) docu-
ments.
The following measures were considered for calculating the topical divergence
between parts 𝑥and 𝑦of an article:
1. Cosine distance (𝐷𝐶):
𝐷𝐶𝑝𝑥, 𝑝𝑦=
𝑝𝑥·𝑝𝑦
𝑝𝑥
𝑝𝑦
=1−Í𝑚
𝑖=1𝑝𝑥𝑖𝑝𝑦𝑖
Í𝑚
𝑖=1𝑝𝑥𝑖2Í𝑚
𝑖=1𝑝𝑦𝑖2(4.1)
4.5 methodology and materials 47
2. Chebyshev distance (𝐷𝐶ℎ):
𝐷𝐶ℎ 𝑝𝑥𝑖, 𝑝𝑦𝑖=max
𝑖=1...𝑚 𝑝𝑥𝑖−𝑝𝑦𝑖(4.2)
3. Euclidean distance (𝐷𝐸):
𝐷𝐸𝑝𝑥𝑖, 𝑝𝑥𝑖=
𝑝𝑥𝑖−𝑝𝑥𝑖
=v
t𝑚
𝑖=1𝑝𝑥𝑖−𝑝𝑥𝑖2(4.3)
4. Hellinger distance (𝐷𝐻):
𝐷𝐻𝑝𝑥, 𝑝𝑦=
1
√2v
t𝑚
𝑖=1√𝑝𝑥𝑖−√𝑝𝑦𝑖2(4.4)
5. Jensen-Shannon divergence (𝐷𝐽 𝑆 ):
𝐷𝐽𝑆 𝑝𝑥
𝑝𝑦=
1
2𝐷𝐾𝐿 𝑝𝑥
𝑝𝑧+𝐷𝐾𝐿 𝑝𝑦
𝑝𝑧(4.5)
where 𝑝𝑧=1
2𝑝𝑥+𝑝𝑦
6. Kullback-Leibler (KL) divergence (𝐷𝐾 𝐿)124 :124 𝐷𝐾 𝐿
is not symmetric and
therefore not a metric, but it can be
transformed into one—to form the
Jensen-Shannon divergence,
𝐷𝐽𝑆
(Equation 4.5).
𝐷𝐾𝐿 𝑝𝑥
𝑝𝑦=
𝑚
𝑖=1
𝑝𝑥𝑖log 𝑝𝑥𝑖
𝑝𝑦𝑖
(4.6)
7. Squared Euclidean distance (𝐷𝑆𝐸 ):
𝐷𝐸𝑝𝑥𝑖, 𝑝𝑥𝑖=
𝑝𝑥𝑖−𝑝𝑥𝑖
2
=
𝑚
𝑖=1𝑝𝑥𝑖−𝑝𝑥𝑖2(4.7)
These measures were all used in the preliminary explorations carried out for
this chapter. Eventually, however, only three (
𝐷𝐶ℎ
,
𝐷𝐸
, and
𝐷𝑆𝐸
) were proceeded
with in the main experiments. Intuitively, the cosine distance indicates the angular
gap between two vectors (distributions of topics, in this case). Chebyshev distance
is the greatest difference found between any two topics in
𝑥
and
𝑦
. The Euclidean
distance measures how far the two topic distributions are from one another, while
the Squared Euclidean distance is simply the square of that farness. The other
measures (
𝐷𝐻
,
𝐷𝐽𝑆
, and
𝐷𝐾𝐿
) were considered as they were originally developed
to deal directly with probability distributions.
4.5 methodology and materials 48
Algorithm 2 Evaluation of thematic divergence in news articles
Input:
(i.) Pairs of first
𝑙=[1,2,...,5]
sentences and remainder
𝑦
of each
fake (
𝑑𝑓
𝑖=D𝑑𝑓
𝑖𝑥, 𝑑𝑓
𝑖𝑦E
;
𝑑𝑓
𝑖𝑥
=𝑙
) and real article (
𝑑𝑟
𝑖=D𝑑𝑟
𝑖𝑥, 𝑑𝑟
𝑖𝑦E
;
𝑑𝑟
𝑖𝑥
=𝑙);
(ii.) LDA model ℳ𝑏𝑔 generated using 𝑆𝑏𝑔 ;
(iii.) Number of topics 𝑁∈ {10,20,30,4050,100,150,200};
(iv.) Divergence function 𝒟∈ {𝐷𝐶ℎ, 𝐷𝐸, 𝐷𝑆𝐸}
Output: n𝐷𝑓
𝑎𝑣𝑔 , 𝐷𝑟
𝑎𝑣𝑔 , 𝐷𝑓
𝑚𝑒𝑑 , 𝐷𝑟
𝑚𝑒𝑑 o
1: for all 𝑙=[1,2,...,5]do
2: for all fake articles D𝑑𝑓
𝑖𝑥, 𝑑𝑓
𝑖𝑦Edo
3: get 𝑁topics in 𝑑𝑓
𝑖𝑥and 𝑑𝑓
𝑖𝑦using ℳ𝑏𝑔
4: end for
5: 𝑇𝑓
𝑖𝑥
=𝑝𝑥
𝑖, . . . , 𝑝𝑥
𝑁𝑓⊲Topics in opening of fake article
6: 𝑇𝑓
𝑖𝑦
=𝑝𝑦
𝑖, . . . , 𝑝 𝑦
𝑁𝑓
⊲Topics in remainder of fake article
7: 𝐷𝑓
𝑖=𝒟𝑇𝑓
𝑖𝑥, 𝑇 𝑓
𝑖𝑦
8: for all real articles D𝑑𝑟
𝑖𝑥, 𝑑𝑟
𝑖𝑦Edo
9: get 𝑁topics in 𝑑𝑟
𝑖𝑥and 𝑑𝑟
𝑖𝑦using ℳ𝑏𝑔
10: end for
11: 𝑇𝑟
𝑖𝑥
=𝑝𝑥
𝑖, . . . , 𝑝𝑥
𝑁𝑟⊲Topics in remainder of real article
12: 𝑇𝑟
𝑖𝑦
=𝑝𝑦
𝑖, . . . , 𝑝 𝑦
𝑁𝑟
⊲Topics in remainder of real article
13: 𝐷𝑟
𝑖=𝒟𝑇𝑟
𝑖𝑥, 𝑇 𝑟
𝑖𝑦
14: 𝐷𝑓
𝑎𝑣𝑔 =𝑚𝑒𝑎𝑛 𝐷𝑓
𝑖𝑖∈{1,...,𝐹 };𝐷𝑟
𝑎𝑣𝑔 =𝑚𝑒𝑎𝑛 𝐷𝑟
𝑖𝑖∈{1,...,𝑅 }
15: 𝐷𝑓
𝑚𝑒𝑑 =𝑚𝑒𝑑𝑖𝑎𝑛 𝐷𝑓
𝑖𝑖∈{1,...,𝐹 };𝐷𝑟
𝑚𝑒𝑑 =𝑚𝑒𝑑𝑖𝑎𝑛 𝐷𝑟
𝑖𝑖∈{1,...,𝑅 }
16: end for
17: return n𝐷𝑓
𝑎𝑣𝑔 , 𝐷𝑟
𝑎𝑣𝑔 , 𝐷𝑓
𝑚𝑒𝑑 , 𝐷𝑟
𝑚𝑒𝑑 o
4.6 experiment 49
4.6 experiment
4.6.1 Preprocessing
All computational operations in this experiment were performed using Python
and freely available packages. Preprocessing for each dataset was done in the
following steps:
1.
Articles are split into sentences using the
NTLK125
package. Each sentence is
125 https://www.nltk.org
tokenised, lowercased and normalised (i.e., accentuations are removed) to
form a list of words, from which stopwords are removed. The union of the
built-in stopwords in the
NLTK
and
spaCy
toolkits, as well as the MySQL
Reference Manual,
126
was used to filter irrelevant words. Furthermore,
126 https://dev.mysql.com/
doc/refman/8.0/en/fulltext-
stopwords.html
additional words typically found in news text but can be considered to be
unimportant were added. Examples of such words include long and short
forms of days of the week, and months, and others such as ‘says’,‘said’,
‘Reuters’,‘Mr’, and ‘Mrs’.
2.
Bigrams were created from two consecutive words which appeared several
times in the corpus. A minimum count of such instances was set to five
and a threshold score as explained in Mikolov et al. (2013b) of 100 was
used. The bigrams are then added to the vocabulary.
3.
Next, each document is lemmatized using
spaCy127
, and only noun, ad-
127
https://spacy.io/models/en
jective, verb, and adverb lemmas are retained. A dictionary is formed by
applying these steps to 𝑆𝑏𝑔 .
4.
Each document is converted into a
BoW
format,
128
which is used to create
128 A list of (token_id,
token_count) tuples.
an LDA model ℳ𝑏𝑔. The models were created with Gensim.129
129 https:
//radimrehurek.com/gensim
5.
Fake and real articles are subsequently preprocessed likewise (i.e., from
raw text data to BoW format) before topics are extracted from them.
Although there is no consensus on whether the inclusion or omission of stop-
words yields better topic models,
130
stopwords can affect the interpretability of
130
Shi et al. (2019), “A new evalu-
ation framework for topic model-
ing algorithms based on synthetic
corpora”
topics as they can diminish the appearance of other more important words. In
this experiment, the goal is to find differences between topics extracted from
legitimate and false news. As false news content often cunningly mimics true
news, it is important to remove words which are contextually irrelevant and
focus on words which can help us tell the two apart.
4.6.2 Datasets
Table 4.1 summarizes the datasets (after preprocessing) used in this study and
lists the domains (as stated by the dataset provider) covered by each. An article’s
4.6 experiment 50
sentence length (Avg. sent. length) is measured by the number of words that
remain after preprocessing. The article’s maximum sentence length (Max. sent
length) is measured in terms of the number of sentences. The following datasets
were used:
1. BuzzFeed-Webis Fake News Corpus 2016 (BuzzFeed-Web)131 131 Potthast et al. (2018),
https://zenodo.org/record/
1239675
2. BuzzFeed Political News Data (BuzzFeed-Political)132
132 Horne and Adali (2017),
https://github.com/
BenjaminDHorne/
fakenewsdata1
3. FakeNewsAMT + Celebrity (AMT+C)133
133
Pérez-Rosas et al. (2018), “Au-
tomatic Detection of Fake News”
4. Falsified and Legitimate Political News Database (POLIT)134
134 http:
//victoriarubin.fims.uwo.ca/
news-verification/access-
polit-false- n-legit- news-
db-2016- 2017
5. George McIntire’s fake news dataset (GMI)135
135 https://github.com/
GeorgeMcIntire/
fake_real_news_dataset
(Accessed 5 November 2018)
6.
University of Victoria’s Information Security and Object Technology (ISOT)
Research Lab136
136 Ahmed et al. (2017),
https://www.uvic.ca/
engineering/ece/isot
7. Syrian Violations Documentation Centre (SVDC)137
137 Salem et al. (2019), https://
zenodo.org/record/2532642
4.7 results and discussion 51
dataset
(d o m a i n)
no. of
fake
no. of
real
avg. sent
lengt h ( f )
avg. sent
lengt h ( r)
max. sent
lengt h ( f )
max. sent
lengt h ( r)
AMT+C
(business, education,
entertainment, politics,
sports, tech)
324 317 14.7 23.2 64 1,059
BuzzFeed-Political
(politics)
116 127 18.9 43.9 76 333
BuzzFeed-Web
(politics)
331 1,214 21.7 26.4 117 211
GMI
(politics)
2,695 2,852 33.9 42.8 1,344 406
ISOT
(government, politics)
19,324 16,823 18.0 20.3 289 324
POLIT
(politics)
122 134 19.2 34.9 96 210
SVDC
(conflict, war)
312 352 14.0 14.6 62 49
Table 4 . 1: Summary of datasets after pre-processing (F – Fake, R – Real).
4.7 results and discussion
As can be seen in the first line of Algorithm 2, experimented with were varying
values of hyperparameter for the length of the opening section,
𝑙
, from 1 to 5.
Results for
𝑙=5
are reported in this section because during initial analyses it
yielded the best results (i.e., the greatest disparity between fake and real deviations)
for most datasets and measures. This is likely due to the first five sentences
containing more information. For example, five successive sentences are likely to
entail one another and contribute more towards a topic than a single sentence.
The outcomes of the experimental evaluation using the different divergence
measures are shown in Table 4.2.
138
It was observed that fake news is generally
138
The average of each
𝑁
group
was found before doing the T-test.
likely to show greater thematic deviation (lesser coherence) than real news in all
datasets. Nonetheless, the mean and median values for fake news are lower than
those of real news for these datasets. Table 4.3, shows the mean and median
𝐷𝐶ℎ
deviations of fake and real articles across all values of
𝑁
, while Figure 4.2 shows
results for comparing topics in the first five and remaining sentences. Results for
values of
𝑁
not shown are similar (with
𝐷𝐶ℎ
gradually decreasing as
𝑁
increases).
As the results for all three measures are alike,
𝐷𝐶ℎ
is focused on for the rest of
the analysis. This is because the choice of divergence measure is not critical to
4.7 results and discussion 52
the outcome of the experiment. Rather it is only a means for estimating thematic
divergence. Table 4.3 shows the mean
𝐷𝐶ℎ
deviations of fake and real articles
across 𝑁={10,20,30,40,50,100,150,200}topics.
AMT+C and BuzzFeed-Web are not statistically significant according to the
T-test. However, the results for all other datasets are. Therefore, for all datasets
except AMT+C and BuzzFeed-Web, the null hypothesis (Hypothesis H
0
)—that
false and authentic news articles are similarly coherent thematically—is rejected
at the 5% level, based on the T-test. In summary, it has been shown statistically
that thematic coherence is generally greater in real news articles compared to
fake ones.
dataset p-value ( 𝐷𝐶 ℎ ) p-va lue ( 𝐷𝐸) p - value (𝐷𝑆 𝐸 )
AMT+C 0.144 0.126 0.116
BuzzFeed-Political 0.0450 0.0147 0.0287
BuzzFeed-Web 0.209 0.209 0.207
GMI 0.0480 0.00535 0.0106
ISOT 0.00319 0.000490 0.000727
POLIT 0.000660 0.0000792 0.0000664
SVDC 0.000684 0.0000112 0.0000789
Table 4 . 2: Results of T-test evaluation based on different measures of deviation used.
dataset mean (𝐷𝐶 ℎ ) (f) mean ( 𝐷𝐶 ℎ ) (r) m e d i a n ( 𝐷𝐶 ℎ ) (f) me d i a n ( 𝐷𝐶 ℎ ) (r)
AMT+C 0.2568 0.2379 0.2438 0.2285
BuzzFeed-Political 0.2373 0.2149 0.2345 0.2068
BuzzFeed-Web 0.2966 0.2812 0.2863 0.2637
GMI 0.4580 0.4241 0.4579 0.4222
ISOT 0.3372 0.2971 0.3369 0.2989
POLIT 0.2439 0.1939 0.2416 0.1894
SVDC 0.2975 0.2517 0.2934 0.2435
Table 4 . 3: Mean and median 𝐷𝐶ℎ deviations of 𝑁={10,20,30,40,50,100,150,200}
topics combined for fake and real news (F – Fake, R – Real).
4.7 results and discussion 53
10 20 30 40 50 100 150 200
N0
0.1
0.2
0.3
0.4
DCh
Avg. and median DCh between first five sentences and rest in AMT+C
Davg
fDavg
rDmed
fDmed
r
(a) Results for AMT+C
10 20 30 40 50 100 150 200
N0
0.1
0.2
0.3
DCh
Avg. and median DCh between first five sentences and rest in BuzzFeed-Political
Davg
fDavg
rDmed
fDmed
r
(b) Results for BuzzFeed-Political
10 20 30 40 50 100 150 200
N0
0.1
0.2
0.3
0.4
DCh
Avg. and median DCh between first five sentences and rest in BuzzFeed-Web
Davg
fDavg
rDmed
fDmed
r
(c) Results for BuzzFeed-Web
Figure 4.2: Average and median Chebyshev distances in fake and real news, when
comparing topics in the first five sentences to the rest of each article. Error
bars show 95% confidence interval.
4.7 results and discussion 54
10 20 30 40 50 100 150 200
N0
0.1
0.2
0.3
0.4
0.5
DCh
Avg. and median DCh between first five sentences and rest in GMI
Davg
fDavg
rDmed
fDmed
r
(d) Results for GMI
10 20 30 40 50 100 150 200
N0
0.1
0.2
0.3
0.4
DCh
Avg. and median DCh between first five sentences and rest in ISOT
Davg
fDavg
rDmed
fDmed
r
(e) Results for ISOT
10 20 30 40 50 100 150 200
N0
0.1
0.2
0.3
DCh
Avg. and median DCh between first five sentences and rest in POLIT
Davg
fDavg
rDmed
fDmed
r
(f) Results for POLIT
Figure 4.2: Average and median Chebyshev distances in fake and real news, when
comparing topics in the first five sentences to the rest of each article. Error
bars show 95% confidence interval. (Cont.)
4.7 results and discussion 55
10 20 30 40 50 100 150 200
N0
0.1
0.2
0.3
0.4
DCh
Avg. and median DCh between first five sentences and rest in SVDC
Davg
fDavg
rDmed
fDmed
r
(g) Results for SVDC
Figure 4.2: Average and median Chebyshev distances in fake and real news, when
comparing topics in the first five sentences to the rest of each article. Error
bars show 95% confidence interval. (Cont.)
4.7 results and discussion 56
It is worth highlighting the diversity of datasets used here, in terms of domain,
size, and the nature of articles. For example, the fake and real news in the SVDC
dataset have a very similar structure. Both types of news were mostly written
with the motivation to inform the reader of conflict-related events that took
place across Syria. However, fake articles are labelled as such primarily because
the reportage (e.g., on locations and the number of casualties recorded) in them
is insufficiently accurate.
▶
To gain insight into possible causes of greater deviation in fake news, the five
most and least diverging fake and real articles (according to
𝐷𝐶ℎ
) were qualita-
tively inspected. A small set of low and high numbers of topics (
𝑁≤30
and
𝑁≥100
) were also compared. It was observed that fake openings tend to be
shorter, vaguer, and less congruent with the rest of the text. By contrast, real
news openings generally give a better narrative background to the rest of the
story. Horne and Adali (2017) reported similar findings regarding the comparison
in length, between authentic and false news articles: i.e., the former is gener-
ally longer than the latter, as shown in Table 4.1. Furthermore, the same study
also found that fake articles are highly redundant and contain less substantial
information.
Although the writing style in fake news is sometimes unprofessional, this is an
unlikely reason for the higher deviations in fake news. Additionally, both fake and
real news may open with the most newsworthy content, and expand on it with
more context and explanation. This is the conventional hierarchical structure of
news,139 as discussed in §2.2.139
van Dijk (1983), “Discourse
Analysis: Its Development and Ap-
plication to the Structure of News”
Indeed, in this study, it was observed that real news tends to have longer
sentences, which give more detailed information about a story and are more
narrative. It can be argued that the reason behind this is that fake articles are
designed to get readers’ attention, whereas legitimate ones are written to inform
the reader. For instance, social media posts which include a link to an article
are sometimes displayed with a short snippet of the article’s opening text or its
summary. This section can be designed to capture readers’ attention.
It was also observed that fake articles include more question and exclamation
marks, as well as words and phrases in all capitals. Although this is inconsequential
to forming topics, it supports the claim that false news is written in an attention-
grabbing style. While Horne and Adali (2017) state that punctuation is less likely
to be found in fake news text, Rubin et al. (2016) suggest that it is a differentiating
factor between fake and real news. Punctuation marks, including question and
exclamation marks, have also been used as a feature in fake news detection.140 140
Pérez-Rosas et al. (2018), “Au-
tomatic Detection of Fake News”
Furthermore, it is conceivable that a bigger team of people working to produce
a fake piece may contribute to its vagueness. They may input different perspectives
that diversify the story and make it less coherent. This may be compared with
real news, whereby there is one professional writer, perhaps two, and therefore,
better coherence.
4.7 results and discussion 57
4.7.1 Quantitative analysis of coherence and perplexity
The observation of greater thematic deviation was further explored experimen-
tally. The qualitative findings previously discussed can be more reliably verified
through quantitative analyses, using empirical measures for topic coherence.141 141
This should not be confused
with the concept of thematic coher-
ence earlier introduced in this chap-
ter. Whereas the former is used in
this thesis to denote consistency
in the subject(s) discussed through-
out a news article, the latter is a
measure for evaluating topics in
general.
Topic coherence assigns a score to each topic by evaluating the semantic
similarity between top words in the topic.
142
It is capable of reflecting people’s
142
Stevens et al. (2012), “Exploring
topic coherence over many models
and many topics”
perception of latent topics in a given text.
143
Thus, topic coherence is adopted
143
Blair et al. (2019), “Aggregated
topic models for increasing social
media topic coherence”
here as an indicator of the amount of vagueness in an article. The intuition behind
this is that topics with high coherence constitute words which allow a reader
to infer the general topic(s) the text is about. Conversely, those with very low
coherence are hardly interpretable,
144
and hence, are likely to arise from vaguer
144
Röder et al. (2015), “Exploring
the space of topic coherence mea-
sures”
text.
Topic coherence measures fall into two groups:
1.
Intrinsic measures, which capture model semantics and are based on hu-
man evaluation of topics’ interpretability.
2.
Extrinsic measures, which indicate how good a topic model is at performing
predefined tasks such as classification.
As intrinsic measures are based on human evaluations, they are more apt
for indicating how a person might assess the coherence of an article they are
reading. Moreover, intrinsic measures have been shown to correlate better with
human judgement.
145
Therefore, one such measure called UMass
146
is used. This
145
Chang et al. (2009), “Reading
tea leaves: How humans interpret
topic models”
146
Mimno et al. (2011), “Optimiz-
ing semantic coherence in topic
models”
is defined in Equation 4.8 as was done by Stevens et al. (2012). From a set of top
words used to describe a topic, UMass measures the extent to which a common
word is a good predictor of a less common word on average.147
147
Mimno et al. (2011), Hemma-
tian et al. (2019)
𝑠𝑐𝑜𝑟𝑒𝑈 𝑀 𝑎𝑠𝑠 𝑣𝑖, 𝑣𝑗, 𝜖=𝑙𝑜𝑔 𝐷𝑣𝑖, 𝑣𝑗+𝜖
𝐷𝑣𝑗(4.8)
where: 𝐷(𝑥)=number of documents which contain word 𝑥
𝐷(𝑥, 𝑦)=number of documents containing words 𝑥and 𝑦
𝜖=smoothing factor that ensures 𝑠𝑐𝑜𝑟𝑒𝑈 𝑀 𝑎𝑠𝑠 is a real number
Topic coherence was evaluated in two ways: (i.) the openings of fake (
𝑆1
) and
authentic articles (
𝑆2
); and (i.) the whole articles. In both cases, the numbers of
topics (𝑁) studied are 10,20,...,140,150,200.148 148
Note that a wider range of
𝑁
is used here, compared with Algo-
rithm 2.
▶𝑆2
articles in the AMT+C dataset have greater coherence than
𝑆1
ones. This
becomes more apparent when
𝑁≥40
. However, focusing on the opening
sections, it can be seen that
𝑆1
opening sentences are only slightly more coherent
4.7 results and discussion 58
than
𝑆2
ones. This means that although
𝑆2
in this dataset are more topically
coherent, the opening sections of 𝑆1are more coherent.
In the BuzzFeed dataset, it can be seen that
𝑆2
articles are more coherent than
𝑆1
articles, both for opening and whole texts. For the openings in BuzzFeed
Political,
𝑆2
articles are only slightly more coherent. Nonetheless, whole
𝑆2
articles are noticeably more coherent than
𝑆1
articles. Considering the opening
sections, it is clear that the
𝑆2
coherence scores are generally—though only
marginally in most cases—higher than those of 𝑆1.
Figure 4.2 shows UMass scores for the first five sentences of
𝑆1
and
𝑆2
articles,
calculated over the training set (a combination of all
𝑆1
and
𝑆2
full articles).
Higher values indicate higher topic coherence, i.e., words associated with each
topic in that model are more likely to co-occur. As expected, more topics are
generally less coherent than fewer ones.
(a) (b)
(c) (d)
4.7 results and discussion 59
(e) (f)
(g) (h)
Figure 4.2: UMass topic coherence scores
▶
In summary, the topic coherence of authentic news is generally greater than that
of misinformation, in all datasets except AMT+C. This is the case in the articles’
opening sections, and when considering the whole article. The UMass coherence
scores suggest that true articles are less vague, compared with fake ones, as they
form more coherent topics. This corroborates earlier qualitative findings on
the coherence of real and fake articles. Nonetheless, manual inspection of the
top words in each topic may still be required. While some datasets show a clear
distinction between false and true articles’ coherence scores, the disparity is not
clear in others.
Ideally, applying insights from Röder et al. (2015), a different topic coherence
score called
𝐶𝑉
, should be used. The authors found this to be the best amongst
topic coherence measures in their study. The
𝐶𝑉
score uses Normalized Point-
wise Mutual Information and cosine similarity measure (see Equation 4.1) in
its workings. One drawback of the
𝐶𝑉
score is its runtime—it takes more than
twenty times longer to run compared with UMass. In any case, the performance
of UMass suffices in this exploration.
The datasets analysed here cover a broad range of themes and contain articles
with different structures and writing styles. Furthermore, their constituent false
and real articles have sentences of varying lengths and vocabularies of varying
4.8 concl u s ion 60
sizes. These findings show that regardless of the individual attributes of datasets,
fake news articles appear to have some high-level features which can be used to
systematically tell them apart from real news.
4.8 conclusion
Fake news and deceptive stories tend to open with sentences which may be
incoherent with the rest of the text. It is worth exploring if the consistency of
fake and real news can distinguish between the two. Accordingly, the thematic
deviations of seven cross-domain fake and real news, using topic modelling were
investigated. The findings presented in this chapter suggest that the opening
sentences of fake articles topically deviate more from the rest of the article, in
contrast to real news. The next step is to find possible reasons behind these
deviations through in-depth analyses of topics. In conclusion, this paper presents
valuable insights into thematic differences between fake and authentic news,
which may be exploited for fake news detection.
4.8.1 Future work
Future work can extend this research in two main ways. Firstly, experimenting
with topic modelling methods other than
LDA
may improve the results. One
example that has been demonstrated to outperform
LDA
in the task of learning
insightful topics is
top2vec149
. This topic modelling algorithm combines doc-
149
Angelov (2020), “Top2Vec: Dis-
tributed Representations of Top-
ics”
ument and word vectors to find topics. Both are commonly used features for
fake news detection. With
top2vec
, topic vectors are indicative of the semantic
similarity between documents. Additionally, it automatically finds the optimal
number of topics. By comparison, this is done iteratively with
LDA
, by evaluating
metrics such as perplexity across a varying number of topics.
Secondly, other techniques beyond splitting an article into multiple parts can
also be investigated. Ideally, the feature extraction method should be resilient
to minor alterations in the news text. It is worth stating that experiments in
parallel with this idea were also carried out in this research. For example, the
semantic coherence between the extractive summaries and opening paragraphs
of articles was evaluated, using word embeddings obtained from Bidirectional En-
coder Representations from Transformers (BERT).150 Further experiments were 150
Devlin et al. (2019), “BERT:
Pre-training of Deep Bidirectional
Transformers for Language Under-
standing”
carried out to find the amount of overlap between the summaries and opening
paragraphs, and the positions of sentences in each article, which constitute its
extractive summary. This branch of experiments did not show semantic coher-
ence—found in this particular way—to be a robust marker of misinformation
detection. Nonetheless, it can be further investigated in future research.
5
CLUSTERING AND CLASSIFICATION USING TOPICS
The various experiments in the preceding chapters culminated in demonstrating
that topics can be used to tell apart fake from authentic news texts. This can
be achieved using the unique representation of topics from the opening and
remainder sections of news articles. This conclusion was arrived at following
statistical tests, which showed that there is some evidence that fake and real news
may differ thematically.
In this study, the utility of topic representations using simple methods is anal-
ysed. First, through unsupervised learning, clustering; and second, through super-
vised learning, classification. The most straightforward possible
ML
methods are
selected here because the goal is to evaluate the utility of these representations. It
should be noted that the topic distributions themselves are used as features in this
case, rather than the divergence scores calculated from them. Therefore, the ex-
periments in this chapter are not based on the calculated variance between topics
in the opening and remainder parts of fake and real articles per se. Nonetheless,
this information is still retained in the distributions.
As related in Chapter 2, both approaches (i.e., clustering and classification) have
been used by several works in the literature on misinformation detection. Their
advantages and shortcomings, particularly in the context of misinformation, are
also discussed therein.
Clustering is done using the
𝐾
-means algorithm. Having performed feature
extraction in an unsupervised way, it is additionally beneficial to further detect
misinformation likewise. A wide range of classifiers was experimented with,
including Decision Trees, Random Forest and SVM.
5.1 clustering
This experiment was carried out on whole topic distributions from the opening
and remainder sections of articles, as well as their reduced 2D vectors (called the
Aggregate method here).
problem definition: The
𝐾
-means algorithm requires us to specify the
number of clusters outputted,
𝐾
. The evaluation for this experiment, detailed later
61
5.1 clustering 62
in this subsection, will focus on determining whether the clusters are becoming
pure or not.
datasets and data: The datasets used in this experiment are the same as in
Chapter 4
151
except for GMI because it was only considered as a first step in the
151 See §4.6.2.
other experiments. The same topic data was used too, except that it is shortened
here, which is to say,
𝑁={10,20,30,40,50}
.
152
Therefore, there is a
(1×150)152 See Algorithm 2in §4.5.2.
topic distribution for each article.
conjectures and baselines: Three baselines were formulated to assess
and compare the utility (based on the clustering metric used hereinafter) of
extracting topic features from articles in different ways. For example, whether
using reduced dimensions of the topic features improve the clustering. Or, if
extracting features from two sections of an article is any better than taking topics
from the entire text.
Conjecture 1: Topics extracted from the opening and remainder sections of articles
improve clustering (Aggregate method), compared with topics extracted from the whole
document.
Baseline 1 was created to evaluate Conjecture 1. Here, topics (10, 20, 30, 40,
and 50) are extracted from entire documents, instead of from their openings and
remainders. Therefore, each document is represented as a single 150-dimensional
vector.
𝐾
-means (with
𝐾=2
and the maximum number of iterations set to 500) is
run on the original 150D topic distribution, as well as on their reduced dimension
(2D) vectors. The projection-based dimensionality reduction methods experi-
mented with are: Autoencoder,
tSNE
, and Uniform Manifold Approximation and
Projection (
UMAP
). The component-based methods used are: Linear,
NMF
,
PCA
,
and Singular Value Decomposition (SVD).
Conjecture 2: Combining multiple topic distributions also improves clustering,
compared with individual topics on their own.
Baseline 2 tests Conjecture 2: individual topic distributions (for 10, 20, 50, and
100 topics) from the opening and remainder sections form the clustering data.
For example, the vector for 10 topics will be a (1×20)vector.
Conjecture 3: Clustering performs better than simply assigning examples to classes
randomly.
Baseline 3 tests Conjecture 3: here, the quality of clustering is calculated based
on random assignment to each class, i.e., half of each type of news forms a cluster.
evaluat i on: There are two main ways to evaluate the quality of clustering:
internal and external criteria.
153
Ideally, articles within a given cluster should
153
Manning et al. (2008), Introduc-
tion to Information Retrieval
be similar (inter-cluster similarity), and those from different clusters should be
dissimilar (intra-cluster similarity). This is the basis of the internal criterion. On
5.1 clustering 63
the other hand, the external criterion requires a benchmark created by people
who can expertly categorise each item. As it takes account of the nuances of a
given application, the external criterion is more reliable, especially if a model is
to be deployed in the real world. It is applicable in this case since the data is fully
labelled. One such criterion, Purity, was used in this experiment to evaluate the
aforementioned baselines. Following Manning et al. (2008), it can be defined as:
𝑝𝑢𝑟𝑖𝑡 𝑦 (Ω, 𝐶)=
1
𝐷
𝐾
𝑘=1
max
𝑗=1...𝐽 𝜔𝑘∩𝑐𝑗(5.1)
where: Ω={𝜔1, 𝜔2, . . . , 𝜔𝐾}is a set of clusters
𝐶={𝑐1, 𝑐2, . . . , 𝑐𝐽}is a set of classes
𝐷=total number of documents
𝜔𝑘=set of documents in 𝜔𝑘
𝑐𝑗=set of documents in c𝑗
The most frequent class of articles–fake or real–in a cluster is assigned as
the label of that cluster. Therefore, the accuracy of the clustering is the sum of
fractions of correct assignments in each cluster. This summarises purity as a
metric.
Figure 5.0 shows plots of the concatenated 300D data, with their dimensions
reduced to 2D, using the Linear method for dimensionality reduction in Wolfram
Mathemematica.
154
Each data point is coloured according to its class. It can be
154
Wolfram Research (2021), “Lin-
ear” (Machine Learning Method)
observed from the figure that reducing the dimensions removes superfluous
information while preserving the essential information that apparently differen-
tiates the two types of news. Significant variations can be observed in the topic
distributions of fake and real news in all datasets, except for BuzzFeed.
Put simply, when it comes to both fake and real articles talking about the same
subject, even across various domains, there are noticeable variations in how they
approach the topic in the beginning and the rest of the articles. This observation
is in agreement with the outcomes of previous experiments, that topics are an
effective feature for detecting misinformation. Naturally, the boundaries between
the clusters are not clear-cut. For example, there is a discernible overlap among
the clusters in the ISOT dataset, indicating the presence of both counterfeit and
genuine articles that nsimilarly narrate stories. However, notable differences can
be seen when examining multiple fabricated and legitimate articles in general.
5.1 clustering 64
(a) (b)
(c) (d)
(e) (f)
Figure 5.0: 2D plots of dimension-reduced topic distributions for datasets used.
5.1 clustering 65
results and discussion: Table 5.1 shows results for the evaluation of
Baseline 1 on the different data dimensions experimented with. Two values, 100
and 200, were used for the
tSNE
parameter perplexity.
155
For
UMAP
, two values,
155
Note that this notion of per-
plexity is different from the one
discussed in 4.7.1.
50 and 100, were used for the number of neighbours. Table 5.2 shows results for
the evaluation of Baseline 2.
5.1 clustering 66
dimension
amt+c buzzfeed
buzzfeed-
political isot polit svdc
ag g b 1 a g g b 1 a g g b 1 agg b1 agg b1 agg b1
150D / 300D 0.5179 0.5367 0.7858 0.7858 0.5205 0.5226 0.5679 0.5346 0.6211 0.5234 0.7169 0.5452
Autoencoder 0.5179 0.5320 0.7858 0.7858 0.5220 0.6872 0.5453 0.5345 0.5234 0.6523 0.6221 0.5301
2D, Linear 0.5133 0.5211 0.7858 0.7858 0.8150 0.7078 0.5897 0.5579 0.7031 0.6602 0.5949 0.5301
2D, NMF 0.5086 0.5226 0.7858 0.7858 0.5410 0.5225 0.5844 0.5409 0.5820 0.5234 0.7334 0.5557
2D, PCA 0.5195 0.5413 0.7858 0.7858 0.6271 0.5226 0.5346 0.5346 0.6211 0.5234 0.7500 0.5437
2D, SVD 0.5117 0.5288 0.7858 0.7858 0.5574 0.5226 0.5608 0.5346 0.6094 0.5234 0.7425 0.5422
2D, TSNE
[𝑝=100]
0.5055 0.5273 0.7858 0.7858 0.9016 0.5226 0.5591 0.5346 0.7648 0.5625 0.7892 0.5723
2D, TSNE
[𝑝=200]
0.5179 0.5273 0.7858 0.7858 0.8893 0.5597 0.5826 0.5346 0.7617 0.5430 0.8298 0.5723
2D, UMAP
[𝑛𝑛 =50]
0.5226 0.5445 0.7858 0.7858 0.5205 0.5597 0.5828 0.5346 0.5742 0.5391 0.5301 0.5768
2D, UMAP
[𝑛𝑛 =100]
0.5304 0.5445 0.7858 0.7858 0.5205 0.5597 0.5815 0.5346 0.5547 0.5391 0.6054 0.5768
Table 5 . 1: Purity scores for Baseline 1 (B1) and Aggregate (Agg) methods. 𝑝= perplexity, 𝑛𝑛 = number of neighbours.
5.1 clustering 67
dataset 𝑇=10 𝑇=20 𝑇=50
AMT+C 0.5101 0.5070 0.5663
BuzzFeed 0.7858 0.7858 0.7858
BuzzFeed-Political 0.5820 0.5328 0.5697
ISOT 0.5373 0.5595 0.5346
POLIT 0.5898 0.5430 0.5860
SVDC 0.5979 0.5919 0.7063
Table 5 . 2: Purity scores for Baseline 2
With regards to Baseline 1, clustering on combined (i.e., concatenated) topic
distributions from the opening and remainder of documents (Aggregate method)
generally performs better, than clustering on whole documents. The only excep-
tions to this are in AMT+C where the opposite result is observed; and BuzzFeed
Political, where there is no significant difference, probably owing to class imbal-
ance. As for Baseline 2, the results show that the combination of multiple topic
distributions also gives better clustering performance, than individual topics,
except in BuzzFeed Political. Baseline 3 results show that clustering outperforms
random assignment.
dataset baseline 1 baseline 2 baseline 3 aggregate
150d 2d 𝑇=1 0 𝑇=2 0 𝑇=5 0 300d 2d
AMT+C 0.5367 0.5273 0.5101 0.5070 0.5663 0.5023 0.5179 0.5179
BuzzFeed 0.7858 0.7858 0.7858 0.7858 0.7858 0.7864 0.7858 0.7858
BuzzFeed Political 0.5226 0.5597 0.5820 0.5328 0.5697 0.5204 0.5205 0.8893
ISOT 0.5346 0.5346 0.5373 0.5595 0.5346 0.5346 0.5679 0.5826
POLIT 0.5234 0.5430 0.5898 0.5430 0.5860 0.5234 0.6211 0.7617
SVDC 0.5452 0.5723 0.5979 0.5919 0.7063 0.5301 0.7169 0.8298
Table 5 . 3:
Comparison of clustering purity scores.
tSNE
with
𝑝=200
is used to obtain
2D data. The best purity scores for each dataset are in bold.
All results for clustering are summarised in Table 5.3. They show that clus-
tering on a combination of multiple topic distributions from the opening and
remainder of articles generally performs better than the formulated baselines. The
exceptions are AMT+C and BuzzFeed. The latter dataset has a significant class
imbalance (331 fake articles and 1,214 real ones), which is a likely explanation
5.2 classification 68
for the absence of variation in its results. Dimensionality reduction appears to
retain important thematic information and improve clustering. Misinformation
detection is typically done using a variety of high-level and low-level (shallow)
features. For example, combining thematic features with semantic or linguistic
ones may yield improvements in the current clustering results.
In conclusion, the clustering experiments presented in this chapter have demon-
strated that features obtained through topic modelling may be exploited for
misinformation detection. Crucially, unsupervised learning is advantageous for
problems such as this. Therefore, the combination of multiple unsupervised learn-
ing methods, such as topic modelling, dimensionality reduction, and clustering,
allows for an end-to-end unsupervised pipeline for detecting misinformation.
However, such an implementation would not be without constraints. For though
clustering algorithms such as
𝐾
-means are efficient, topic modelling and dimen-
sionality reduction can be time-consuming, depending on the size of the dataset.
In future work, the experimental methods can be improved. Firstly, as men-
tioned in Chapter 4, the topic modelling method for feature extraction can be
improved. Secondly, other clustering methods can also be experimented with. In
this research, spectral clustering was also considered, but
𝐾
-means gave better
results. In a semi-supervised scenario, rather than setting
𝐾=2
, articles may be
clustered into three authentic, false, and indeterminable—and human experts
can review and label items in the third group.
5.2 classification
Classification was also applied to assess the effectiveness of topic representations
as markers for distinguishing between authentic and false news. Classification is
the most prominent ML method in the literature.
The models used are Decision Trees, Gradient Boosted Trees, Logistic Regres-
sion, Markov Model, Naive Bayes,
kNN
, Neural Network, Random Forest, and
SVM
. First, a dataset is trained using all types of classifiers simultaneously. To do
this, the data is split 80% for training and validation, and 20% for testing. Multiple
versions of some classifiers, with different parameters, were created and trained
using the 80% portion. Next, the best classifier (based on loss) and its parameters
are selected. Note that the test set was not used in selecting the hyperparameters
of the classifiers, but only for testing later on.
Finally, this classifier is recreated, applying its parameters, to train the dataset
afresh, using five-fold cross-validation. Classification was done using the Wolfram
Language
Classify156
function. Unless stated otherwise, the default parameters
156
Wolfram Research (2021), Clas-
sify
for all classifier types were used.
datasets and data: The datasets used here are the same as for clustering.
157
This time though, an additional dataset, FakeNewsNet,
158
was also used. This
157
See 4.6.2,Table 4.1 for more
information on datasets.
158 Shu et al. (2020),
https://github.com/
KaiDMML/FakeNewsNet
dataset contains 4,443 and 13,433 fake and real articles, respectively. Similar to
5.2 classification 69
FakeNewsNet, the BuzzFeed dataset also has a significant class imbalance—with
331 and 1,214 fake and real articles, respectively). These two datasets were bal-
anced, by randomly sampling the bigger class to select the same number of articles
from the smaller one. Therefore, the final FakeNewsNet and BuzzFeed datasets
used had 4,443 and 331 articles, respectively, in both classes.
The original representation used for clustering contained
𝑁={10,20,30,40,50}
topics, extracted from the opening and remaining text of each article. Therefore,
for
𝑚
articles, the dimension of the data is
(5×2×𝑚)
. This data was modified
to obtain the following topic data representations for classification:
1. The original topic representation, i.e., a (5×2×𝑚)tensor.
2.
Flattened topic representation, i.e., the original tensor concatenated to a
300D vector. The sum of
𝑁
dimensions for each section of the article is
150; concatenating the distributions for both sections gives 300D.)
3.
Dimension reduced representation, i.e., the 300D vector reduced to 2D
using tSNE (with the parameter 𝑝𝑒𝑟 𝑝𝑙𝑒 𝑥𝑖𝑡 𝑦 =200).
results and discussion: Table 5.4 shows results for the original topic
representation. After the initial training, the best classifier on each dataset is
a variant of a logistic regressor. Four datasets (FakeNewsNet, GMI, ISOT, and
SVDC) were classified with more than 90% accuracy, and all others with more
than 80%. Accuracy is satisfactory as an indicator of the overall classification
performance because the datasets do not have huge class imbalances.
dataset accuracy f1 precision recall
AMT+Ca0.8393 0.8384 0.8380 0.8413
BuzzFeedb0.8384 0.8376 0.8375 0.8411
BuzzFeed-Politicalc0.8933 0.8912 0.8953 0.8906
FakeNews Netd0.9487 0.9314 0.9307 0.9323
GMIe0.9171 0.9169 0.9172 0.9168
ISOTf0.9352 0.9346 0.9364 0.9335
POLITg0.8479 0.8451 0.8494 0.8462
SVDCh0.9353 0.9345 0.9358 0.9343
𝑎
Logistic Regressor (
𝑙2_𝑟𝑒𝑔 =100
),
𝑏
Logistic Regressor (
𝑙2_𝑟𝑒𝑔 =100
),
𝑐
Logistic Regressor (
𝑙2_𝑟𝑒𝑔 =10
),
𝑑
Logistic Regressor
(
𝑙2_𝑟𝑒𝑔 =100
),
𝑒
Logistic Regressor (
𝑙2_𝑟𝑒𝑔 =100
),
𝑓
Logistic Regressor (
𝑙2_𝑟𝑒𝑔 =100
),
𝑔
Logistic Regressor (
𝑙2_𝑟𝑒𝑔 =10
),
ℎ
Logistic Regressor (𝑙2_𝑟𝑒𝑔 =10).
Table 5 . 4: Evaluation metrics for the best-performing classifier on each dataset, using
the original representation. 𝑙2_𝑟𝑒𝑔 =𝐿2regularisation
5.2 classification 70
Table 5.5 shows the results for the dimension-reduced (2D) topic represen-
tation. The accuracy scores are generally lower compared with when using the
original dimensions, but they remain at 80% or higher in the BuzzFeed Political,
FakeNewsNet, POLIT, and SVDC datasets.
dataset accuracy f1 precision recall
AMT+Ca0.5289 0.5261 0.5307 0.5301
BuzzFeedb0.6285 0.6274 0.6281 0.6283
BuzzFeed-Politicalc0.8853 0.8838 0.8885 0.8845
FakeNews Netd0.8004 0.8002 0.8017 0.8005
GMIe0.6358 0.6285 0.6410 0.6322
ISOTf0.6779 0.6778 0.6799 0.6803
POLITg0.8515 0.8495 0.8556 0.8597
SVDCh0.9217 0.9208 0.9248 0.9195
𝑎
Random Forest (
𝑙𝑒𝑎𝑓 _𝑠𝑖𝑧𝑒 =2, 𝑛𝑢𝑚_𝑡𝑟𝑒𝑒𝑠 =100
),
𝑏
Logistic Regressor (
𝑙2_𝑟𝑒𝑔 =1
),
𝑐
Logistic Regressor (
𝑙2_𝑟𝑒𝑔 =0.01
),
𝑑kNN
(
𝑛𝑛 =20, 𝑚𝑒𝑡ℎ 𝑜𝑑 =𝐾 𝐷𝑡𝑟𝑒𝑒
),
𝑒kNN
(
𝑛𝑛 =50, 𝑚𝑒𝑡ℎ 𝑜𝑑 =𝐾 𝐷𝑡𝑟𝑒𝑒
),
𝑓kNN
(
𝑛𝑛 =500, 𝑚𝑒𝑡ℎ 𝑜𝑑 =𝐾 𝐷𝑡𝑟𝑒𝑒
),
𝑔kNN
(
𝑛𝑛 =20, 𝑚𝑒𝑡ℎ 𝑜𝑑 =
𝐾𝐷𝑡𝑟 𝑒𝑒), ℎkNN (𝑛𝑛 =10, 𝑚𝑒𝑡 ℎ𝑜𝑑 =𝐾 𝐷𝑡𝑟𝑒𝑒).
Table 5 . 5: Evaluation metrics for the best-performing classifier on each dataset, using
the 2D representation. 𝑙2_𝑟𝑒𝑔 =𝐿2regularisation,
𝑛𝑛 =number of neighbours, 𝑛𝑢𝑚_𝑡𝑟𝑒𝑒𝑠 =number of trees.
Table 5.6 shows the results for the flattened 300D topic representation. They
are generally better than the results of the 2D representation but not as good as
those of the original representation. The accuracy scores only drop below 80% in
the AMT+C and BuzzFeed datasets.
5.3 concl u s ion 71
dataset accuracy f1 precision recall
AMT+Ca0.7628 0.7607 0.7624 0.7609
BuzzFeedb0.7372 0.7360 0.7376 0.7372
BuzzFeed-Politicalc0.8813 0.8768 0.8798 0.8765
FakeNews Netd0.9102 0.9101 0.9102 0.9101
GMIe0.9037 0.9036 0.9042 0.9034
ISOTf0.9242 0.9238 0.9240 0.9236
POLITg0.8124 0.8111 0.8160 0.8140
SVDCh0.9247 0.9241 0.9237 0.9260
𝑎
Gradient Boosted Trees (
𝑙𝑒𝑎𝑓 _𝑠𝑖𝑧𝑒 =35, 𝑚𝑎𝑥_𝑑𝑒𝑝𝑡ℎ =6, 𝑛𝑢𝑚_𝑙𝑒𝑎𝑣𝑒𝑠 =110, 𝑙2_𝑟𝑒𝑔 =0, 𝑚 𝑎𝑥_𝑡𝑟_𝑟𝑜𝑢𝑛𝑑𝑠 =50, 𝑙 𝑟 =0.04
),
𝑏
Logistic
Regressor (
𝑙2_𝑟𝑒𝑔 =1×10−6
),
𝑐
Logistic Regressor (
𝑙2_𝑟𝑒𝑔 =1
),
𝑑
Logistic Regressor (
𝑙2_𝑟𝑒𝑔 =100
),
𝑒
Logistic Regressor
(
𝑙2_𝑟𝑒𝑔 =100
),
𝑓
Logistic Regressor (
𝑙2_𝑟𝑒𝑔 =10
),
𝑔
Logistic Regressor (
𝑙2_𝑟𝑒𝑔 =1
),
ℎ
Gradient Boosted Trees (
𝑙𝑒𝑎𝑓 _𝑠𝑖𝑧𝑒 =
35, 𝑚𝑎𝑥_𝑑𝑒𝑝𝑡ℎ =6, 𝑛𝑢𝑚_𝑙𝑒𝑎𝑣𝑒𝑠 =110, 𝑙2_𝑟 𝑒𝑔 =0, 𝑚 𝑎𝑥_𝑡𝑟_𝑟 𝑜𝑢𝑛𝑑𝑠 =50, 𝑙𝑟 =0.1).
Table 5 . 6: Evaluation metrics for the best-performing classifier on each dataset, using
the 300D representation.
𝑙2_𝑟𝑒𝑔 =𝐿2regularisation, 𝑙𝑟 =learning rate, 𝑚𝑎𝑥_𝑡𝑟_𝑟 𝑜𝑢𝑛𝑑𝑠 =
maximum training rounds, 𝑛𝑢𝑚_𝑡𝑟𝑒𝑒𝑠 =number of trees.
In summary, using the original topic representation, without dimension re-
duction or flattening, gives the best classification performance. However, this
requires a noticeably longer training time compared with the two other represen-
tations. Dimension reduction appears to lose some important information that
could enhance classification. Nonetheless, the retained information is adequate
for classification on four of the datasets used. Topics are yet to be fully exploited
as features in the misinformation detection literature. The classification results
suggest that topic data representations can be used in different ways as features
for this task. They can be used as standalone features, as demonstrated here, or
combined with other kinds of features to improve the generalization ability of
an
ML
model for misinformation detection. The latter is a worthwhile direction
to pursue in future work.
5.3 conclusion
In the previous chapter, topic representations of articles were introduced and
explored as potentially viable features for misinformation detection. This chapter
has presented experiments aimed at demonstrating the utility of topic represen-
tations. Simple implementations of clustering and classification have been used
to separate authentic news articles from false ones. The results suggest that these
representations are efficacious for this task.
Part III
EPILOGUE
6
C O N C L U S I O N
One way to detect misinformation is by manual fact-checking. This task is typ-
ically done by trained experts who tend to be accurate in spotting fake news.
There are, however, a few issues with this approach. Firstly, the high quantity
and speed of fake news make it difficult for fact-checkers to keep up. Secondly,
continuous exposure to misinformation can be harmful to an individual, or even
lead them to believe it is true. Finally, fact-checkers have a degree of subjectivity,
which can lead to inconsistent results, especially when dealing with complex or
controversial topics.
An alternative way of identifying fake news is by using computational methods.
The application of
NLP
and
ML
techniques to do so is delivering increasingly better
results so far. Supervised
ML
models are more commonly used to detect fake
news, but they rely on a large amount of labelled data.
Fake news is becoming more cunning and even more similar to authentic
news. Some of the features currently used to tell apart fake from real may not
work well when the two are highly similar. For instance, it has been shown
that representations based on writing style, are impractical when applied to
machine-generated fake news. Therefore, there is also a need to innovate new
text representations for distinguishing this type of news from legitimate news.
▶
This thesis makes the following main contributions to the current state of fake
news detection:
1.
It develops a novel approach for obtaining robust text features from news
articles based on the topics they discuss. This is particularly useful in
circumstances where labelled data is scant or unavailable.
2.
It demonstrates the effectiveness of this new representation in distinguish-
ing between fake and real news articles. This is shown using both supervised
(classification) and unsupervised (clustering)
ML
. The latter approach helps
in minimising the reliance on labelled datasets.
These contributions were achieved through the design and implementation of
three main experiments:
1.
This thesis explored word embeddings and sentiment features on short
rumour and non-rumour texts. They did not show evidence of being
74
6.1 fu t u r e work 75
capable of differentiating between the groups. Nonetheless, this study can
be extended in different ways to better understand semantic and sentiment
relations between rumours and non-rumours.
2.
It investigated the coherence in the themes discussed in the opening and
remaining sections of fake and authentic news articles. The themes were
represented in the form of latent topics. This study culminated in the
development of a novel text representation, which showed evidence of
being able to distinguish fake from real news.
3.
It exploited the topic features for misinformation detection, using classifica-
tion and clustering methods. Although these experiments are preliminary,
the results are promising and to some degree, substantiate the efficacy of
topic representations.
In its totality, this thesis contributes to researchers’ ability to detect fake news
computationally. However, further and deeper studies remain to be done.
6.1 future work
There now exist word embedding models that are more advanced than
word2vec
and
InferSent
. The experiments carried out in Chapter 3 can be extended to
take advantage of state-of-the-art language models such as
BERT
. Language models
pre-trained on short texts or tweets, such as
BERTweet159
, may perform better
159
Nguyen et al. (2020),
“BERTweet: A pre-trained
language model for English
Tweets”
than the ones used in this research. Furthermore, concerning sentiment, it has
only explored limited categories of it (positive, neutral, and negative). In future
work, an expanded range of emotions can be studied.
Other topic modelling tools may perform better than
LDA
for a study similar to
the one in Chapter 4. Such a tool may generate better topics as assessed through
intrinsic and extrinsic measures. This will increase confidence in ascertaining the
robustness of thematic coherence as a text representation. For example, Egger
and Yu (2022) carried out a detailed study of the strengths and weaknesses of
different topic modelling methods for investigating
OSN
text data. In this work,
LDA
,
NMF
,
top2vec
,
160
and
BERTopic161
were compared. Additionally, other
160
Angelov (2020), “Top2Vec: Dis-
tributed Representations of Top-
ics”
161
Grootendorst (2022),
“BERTopic: Neural topic modeling
with a class-based TF-IDF
procedure”
ways of extracting topics from articles can be explored. For instance, the articles
could be split into multiple sections, rather than just two. This may improve the
robustness of topic text representations, to make it more resilient to changes that
can be made to increase the coherence of topics in fake news.
Chapter 5 presents preliminary yet promising results on the evaluation of the
utility of topic representations. Simple classification and clustering algorithms
were used for this. In future work, however, a more novel technique can be
devised to fully take advantage of the features used in this work. For example,
beyond evaluating the purity of clusters, a complete unsupervised fake news
detection model can be created and its performance can be evaluated against the
state-of-the-art.
6.1 fu t u r e work 76
Finally, the text representations explored in this thesis can be combined, or
used with those in other studies, to detect fake news. Experiments can be set
up to compare the performances of the topic and stylometric features, or to
evaluate the utility of their combination. Future research can adopt multimodal
misinformation detection to combine the text representations presented in this
work with features from other types of media. Especially image and video,
162 162
Mirsky and Lee (2021), “The
Creation and Detection of Deep-
fakes”
which are becoming increasingly easier to fabricate using tools such as Generative
Adversarial Networks.
163
Future work may also evaluate if topic features are
163
Goodfellow et al. (2014), “Gen-
erative Adversarial Nets”
robust enough to accurately detect machine-generate fake news. This type of
misinformation has the potential of becoming the primary way to create mis-
and disinformation in the future.
REFERENCES
Agrawal, Parag (June 2019). Twitter acquires Fabula AI to strengthen its machine
learning expertise.url:
https:// blog . twitter . com / en _us /topics/
company/2019/Twitter-acquires-Fabula-AI.
Ahmed, Hadeer, Issa Traore, and Sherif Saad (2017). “Detection of Online Fake
News Using N-Gram Analysis and Machine Learning Techniques.” In: do i:
10.1007/978-3-319-69155-8_9.
Ajao, Oluwaseun, Deepayan Bhowmik, and Shahrzad Zargari (2018). “Fake News
Identification on Twitter with Hybrid CNN and RNN Models.” In: Proceedings
of the 9th International Conference on Social Media and Society. SMSociety
’18. Copenhagen, Denmark: Association for Computing Machinery, 226–230.
isb n: 9781450363341. d o i:
10 . 1145 / 3217804 . 3217917
.url:
https :
//doi.org/10.1145/3217804.3217917.
Ajao, Oluwaseun, Deepayan Bhowmik, and Shahrzad Zargari (2019). “Sentiment
Aware Fake News Detection on Online Social Networks.” In: ICASSP 2019 -
2019 IEEE International Conference on Acoustics, Speech and Signal Processing
(ICASSP), pages 2507–2511. d o i:10.1109/ICASSP.2019.8683170.
Alam, Firoj, Stefano Cresci, Tanmoy Chakraborty, Fabrizio Silvestri, Dimiter
Dimitrov, Giovanni Da San Martino, Shaden Shaar, Hamed Firooz, and Preslav
Nakov (Oct. 2022). “A Survey on Multimodal Disinformation Detection.” In:
Proceedings of the 29th International Conference on Computational Linguistics.
Gyeongju, Republic of Korea: International Committee on Computational
Linguistics, pages 6625–6643. url:
https: //aclanthology .org/2022 .
coling-1.576.
Allcott, Hunt and Matthew Gentzkow (May 2017). “Social media and fake news
in the 2016 election.” In: Journal of Economic Perspectives 31 (2), pages 211–236.
doi:
10.1257/jep.31.2.211
.url:
http://pubs.aeaweb.org/doi/10.
1257/jep.31.2.211.
American Dialect Society (Jan. 2006). Truthiness Voted 2005 Word of the Year.
url:
https://www.americandialect.org/truthiness_voted_2005_
word_of_the_year.
American Dialect Society (Jan. 2018). “Fake news” is 2017 American Dialect Society
word of the year.url:
https://www.americandialect.org/fake-news-
is-2017-american-dialect-society-word- of-the-year.
77
references 78
Angelov, Dimo (Aug. 2020). “Top2Vec: Distributed Representations of Topics.”
In: doi:
10.48550/arxiv.2008.09470
.url:
https://arxiv.org/abs/
2008.09470v1.
Anspach, Nicolas M., Jay T. Jennings, and Kevin Arceneaux (2019). “A little bit
of knowledge: Facebook’s News Feed and self-perceptions of knowledge.” In:
Research and Politics 6 (1). doi:10.1177/2053168018816189.
Araújo, Ana Christina (2006). “The Lisbon Earthquake of 1755 – Public Distress
and Political Propaganda.” In: European Journal of Portuguese History 4 (1),
pages 1–25.
Arslan, Fatma, Naeemul Hassan, Chengkai Li, and Mark Tremayne (Jan. 2020).
“ClaimBuster: A Benchmark Dataset of Check-worthy Factual Claims.” In:
doi:
10.5281/ZENODO.3836810
.url:
https://zenodo.org/record/
3836810.
Barrett, Paul M., Justin Hendrix, and J. Grant Sims (Sept. 2021). Fueling the Fire:
How Social Media Intensifies U.S. Political Polarization — And What Can Be
Done About It. NYU Stern Center for Business and Human Rights. url:
https:
//bhr.stern.nyu.edu/polarization-report-page.
Benamira, Adrien, Benjamin Devillers, Etienne Lesot, Ayush K Ray, Manal Saadi,
and Fragkiskos D Malliaros (Aug. 2019). “Semi-Supervised Learning and Graph
Neural Networks for Fake News Detection.” In: pages 568–569. url:
https:
//hal.archives-ouvertes.fr/hal-02334445.
Bhattacharjee, Sreyasee Das, Ashit Talukder, and Bala Venkatram Balantrapu (Dec.
2018). “Active learning based news veracity detection with feature weighting
and deep-shallow fusion.” In: volume 2018-Janua. IEEE, pages 556–565. d oi:
10.1109/BigData.2017.8257971.
Biyani, Prakhar, Kostas Tsioutsiouliklis, and John Blackmer (Feb. 2016). “"8 Amaz-
ing Secrets for Getting More Clicks": Detecting Clickbaits in News Streams
Using Article Informality.” In: Proceedings of the Thirtieth AAAI Conference on
Artificial Intelligence 30 (1), pages 94–100. issn: 2374-3468. d oi:
10.1609/
AAAI.V30I1 . 9966
.url:
https:/ / ojs.aaai.org/ index .php/AAAI/
article/view/9966.
Blair, Stuart J., Yaxin Bi, and Maurice D. Mulvenna (July 2019). “Aggregated topic
models for increasing social media topic coherence.” In: Applied Intelligence.
issn: 15737497. doi:
10 . 1007 / s10489 - 019 - 01438 - z
.url:
https :
//doi.org/10.1007/s10489-019-01438-z.
Blei, David M. (Apr. 2012). “Probabilistic topic models.” In: Communications of the
ACM 55 (4), pages 77–84. d o i:
10.1145/2133806.2133826
.url:
https:
//dl.acm.org/doi/abs/10.1145/2133806.2133826.
references 79
Blei, David M., Andrew Y. Ng, and Michael I. Jordan (2003). “Latent Dirichlet
Allocation.” In: Journal of Machine Learning Research 3 (Jan), pages 993–1022.
url:http://jmlr.csail.mit.edu/papers/v3/blei03a.html.
Bojanowski, Piotr, Edouard Grave, Armand Joulin, and Tomas Mikolov (July
2016). “Enriching Word Vectors with Subword Information.” In: url:
http:
//arxiv.org/abs/1607.04606.
Brown, Heather, Emily Guskin, and Amy Mitchell (Nov. 2012). The Role of So-
cial Media in the Arab Uprisings.url:
https: //www .pewresearch. org/
journalism/2012/11/28/role-social-media-arab-uprisings.
Cai, Guoyong, Hao Wu, and Rui Lv (2014). “Rumors detection in Chinese via
crowd responses.” In: ASONAM 2014 - Proceedings of the 2014 IEEE/ACM
International Conference on Advances in Social Networks Analysis and Mining.
doi:10.1109/ASONAM.2014.6921694.
Campan, Alina, Alfredo Cuzzocrea, and Traian Marius Truta (2018). “Fighting fake
news spread in online social networks: Actual trends and future research direc-
tions.” In: volume 2018-January, pages 4453–4457. d oi:
10.1109/BigData.
2017.8258484.
Cao, Juan, Peng Qi, Qiang Sheng, Tianyun Yang, Junbo Guo, and Jintao Li (2020).
“Exploring the role of visual content in fake news detection.” In: Disinformation,
Misinformation, and Fake News in Social Media: Emerging Research Challenges
and Opportunities, pages 141–161.
Caplan, Robyn, Lauren Hanson, and Joan Donovan (2018). “Dead Reckoning: Navi-
gating Content Moderation After Fake News.” In: url:
https://datasociety.
net/pubs/oh/DataAndSociety_Dead_Reckoning_2018.pdf.
Casillo, Mario, Francesco Colace, Brij B. Gupta, Domenico Santaniello, and
Carmine Valentino (2021). “Fake News Detection Using LDA Topic Mod-
elling and K-Nearest Neighbor Classifier.” In: edited by David Mohaisen and
Dr. Ruoming Jin. Springer International Publishing, pages 330–339. i s b n:
978-3-030-91434-9. doi:10.1007/978-3-030-91434- 9_29.
Castillo, Carlos, Marcelo Mendoza, and Barbara Poblete (2011). “Information
credibility on twitter.” In: ACM Press, page 675. do i:
10.1145 / 1963405 .
1963500.
Center for Information Technology & Society, UCSB (2022). A Brief History of
Fake News.url:
https :/ / www. cits .ucsb .edu / fake- news / brief-
history.
Chang, Jonathan, Jordan Boyd-Graber, Sean Gerrish, Chong Wang, and David
M. Blei (2009). “Reading tea leaves: How humans interpret topic models.” In:
pages 288–296.
references 80
Chen, Weiling, Chai Kiat Yeo, Chiew Tong Lau, and Bu Sung Lee (Oct. 2016).
“Behavior deviation: An anomaly detection view of rumor preemption.” In:
IEEE, pages 1–7. d oi:
10. 1109 / IEMCON.2016.7746262
.url:
http: / /
ieeexplore.ieee.org/document/7746262.
Chen, Weiling, Yan Zhang, Chai Kiat Yeo, Chiew Tong Lau, and Bu Sung Lee
(Apr. 2018). “Unsupervised rumor detection based on users’ behaviors using
neural networks.” In: Pattern Recognition Letters 105 (C), pages 226–233. d o i:
10.1016/j.patrec.2017.10.014.
Chen, Yimin, Niall J. Conroy, and Victoria L. Rubin (2015). “Misleading online
content: Recognizing clickbait as “false news”.” In: pages 15–19. d o i:
10.1145/
2823465.2823467.
Chiou, Lesley and Catherine Tucker (Nov. 2018). Fake News and Advertising on
Social Media: A Study of the Anti-Vaccination Movement. National Bureau of
Economic Research. d o i:
10.3386/w25223
.url:
http://www.nber.org/
papers/w25223.
Choi, Daejin, Selin Chun, Hyunchul Oh, Jinyoung Han, and Ted “Taekyoung”
Kwon (2020). “Rumor Propagation is Amplified by Echo Chambers in Social
Media.” In: Scientific Reports 10 (1), page 310. issn: 2045-2322. doi:
10.1038/
s41598-019-57272-3
.url:
https://doi.org/10.1038/s41598-019-
57272-3.
Chowdhury, Nashit, Ayisha Khalid, and Tanvir C. Turin (2021). “Understanding
misinformation infodemic during public health emergencies due to large-scale
disease outbreaks: a rapid review.” In: Zeitschrift Fur Gesundheitswissenschaften
(Journal of public health), pages 1–21. issn: 16132238. do i:
10.1007/S10389-
021 - 01565 - 3
.url:
/pmc / articles / PMC8088318 / /pmc / articles /
PMC8088318/?report= abstracthttps : / / www .ncbi.nlm . nih . gov /
pmc/articles/PMC8088318/.
Coleman, Keith (Jan. 2021). Introducing Birdwatch, a community-based approach
to misinformation.url:
https:/ / blog.twitter.com/ en _us /topics/
product/2021/introducing-birdwatch-a-community-based-approach-
to-misinformation.
Collobert, Ronan and Jason Weston (2008). “A unified architecture for natu-
ral language processing.” In: ACM Press, pages 160–167. d oi:
10 . 1145 /
1390156 . 1390177
.url:
http : / / portal . acm . org / citation . cfm ?
doid=1390156.1390177.
Collobert, Ronan, Jason Weston, Léon Bottou, Michael Karlen, Koray Kavukcuoglu,
and Pavel Kuksa (Nov. 2011). “Natural Language Processing (Almost) from
Scratch.” In: J. Mach. Learn. Res. 12, pages 2493–2537. url:
http://dl.acm.
org/citation.cfm?id=1953048.2078186.
references 81
Conklin, Jeffrey (2006). Dialogue Mapping: Building Shared Understanding of
Wicked Problems. Wiley. is b n: 978-0-470-01768-5.
Conneau, Alexis, Douwe Kiela, Holger Schwenk, Loïc Barrault, and Antoine
Bordes (2017). “Supervised Learning of Universal Sentence Representations
from Natural Language Inference Data.” In: Proceedings of the 2017 Conference
on Empirical Methods in Natural Language Processing. Copenhagen, Denmark:
Association for Computational Linguistics, pages 670–680. url:
https://
www.aclweb.org/anthology/D17-1070.
Devlin, Jacob, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova (June 2019).
“BERT: Pre-training of Deep Bidirectional Transformers for Language Under-
standing.” In: Proceedings of the 2019 Conference of the North, pages 4171–4186.
doi:
10.18653/V1/N19-1423
.url:
https://aclanthology.org/N19-
1423.
Dictionary.com (Nov. 2018). Misinformation | Dictionary.com’s 2018 Word of the
Year.url:
https://www.dictionary.com/e/word-of-the-year-2018
.
Egger, Roman and Joanne Yu (2022). “A Topic Modeling Comparison Between
LDA, NMF, Top2Vec, and BERTopic to Demystify Twitter Posts.” In: Frontiers
in Sociology 7. issn: 22977775. d o i:10.3389/fsoc.2022.886498.
European External Action Service (Jan. 2022). Disinformation About the Current
Russia-Ukraine Conflict – Seven Myths Debunked.
Feng, Song, Ritwik Banerjee, and Yejin Choi (2012). “Syntactic Stylometry for
Deception Detection.” In: Proceedings of the 50th Annual Meeting of the Asso-
ciation for Computational Linguistics (Volume 2: Short Papers). Jeju Island, Ko-
rea: Association for Computational Linguistics, pages 171–175. url:
https:
//aclanthology.org/P12-2034.
Ferreira, William and Andreas Vlachos (2016). “Emergent: A novel data-set for
stance classification.” In: pages 1163–1168. doi:10.18653/v1/n16-1138.
Fong, Jessica, Tong Guo, and Anita Rao (June 2021). “Debunking Misinformation
in Advertising.” In: SSRN Electronic Journal.do i:
10.2139/SSRN.3875665
.
url:https://papers.ssrn.com/abstract=3875665.
Gabielkov, Maksym, Arthi Ramachandran, Augustin Chaintreau, and Arnaud
Legout (June 2016). “Social clicks: What and who gets read on twitter?” In:
pages 179–192. d oi:
10.1145 / 2896377 . 2901462
.url:
https:/ / hal .
inria.fr/hal-01281190.
Gelfert, Axel (2018). “Fake news: A definition.” In: Informal Logic 38 (1), pages 84–
117. doi:10.22329/il.v38i1.5068.
references 82
Goldberg, Yoav and Omer Levy (Feb. 2014). “word2vec Explained: deriving
Mikolov et al.’s negative-sampling word-embedding method.” In: url:
http:
//arxiv.org/abs/1402.3722.
Goodfellow, Ian J., Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-
Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio (2014). “Generative
Adversarial Nets.” In: volume 3, pages 2672–2680. d oi:
10. 1007 /978 - 3-
658-40442-0_9.
Google France (Jan. 2021). Le blog officiel de Google France: L’Alliance de la Presse
d’Information Générale et Google France signent un accord relatif à l’utilisation
des publications de presse en ligne.url:
https://france.googleblog.com/
2021/01/APIG-Google.html.
Gorrell, Genevieve, Elena Kochkina, Maria Liakata, Ahmet Aker, Arkaitz Zubi-
aga, Kalina Bontcheva, and Leon Derczynski (2019). “SemEval-2019 Task 7:
RumourEval, Determining Rumour Veracity and Support for Rumours.” In: Pro-
ceedings of the 13th International Workshop on Semantic Evaluation, pages 845–
854. doi:
10.18653/V1/S19-2147
.url:
https://aclanthology.org/
S19-2147.
Grootendorst, M. (2022). “BERTopic: Neural topic modeling with a class-based
TF-IDF procedure.” In: ArXiv.do i:10.48550/arxiv.2203.05794.
Guacho, Gisel Bastidas, Sara Abdali, Neil Shah, and Evangelos E. Papalexakis
(Aug. 2018). “Semi-supervised content-based detection of misinformation via
tensor embeddings.” In: IEEE, pages 322–325. doi:
10.1109/ASONAM.2018.
8508241.
Habgood-Coote, Joshua (2018). “The term ‘fake news’ is doing great harm.” In:
The Conversation (July 27). url:
https:// theconversation . com / the -
term-fake-news-is-doing-great- harm-100406.
Harris, Zellig S. (Aug. 1954). “Distributional Structure.” In: Word 10 (2-3), pages 146–
162. doi:
10 . 1080 / 00437956 . 1954 . 11659520
.url:
http : / / www .
tandfonline.com/doi/full/10.1080/00437956.1954.11659520.
Hemmatian, Babak, Sabina J. Sloman, Uriel Cohen Priva, and Steven A. Sloman
(Aug. 2019). “Think of the consequences: A decade of discourse about same-
sex marriage.” In: Behavior Research Methods 51 (4), pages 1565–1585. d oi:
10.3758/s13428-019-01215-3
.url:
http://link.springer.com/10.
3758/s13428-019-01215-3.
Hinton, G E, J L McClelland, and D E Rumelhart (1986). Parallel Distributed
Processing: Explorations in the Microstructure of Cognition, Vol. 1. Edited by
David E Rumelhart, James L McClelland, and CORPORATE PDP Research
Group. url:http://dl.acm.org/citation.cfm?id=104279.104287.
references 83
Horne, Benjamin D. and Sibel Adali (2017). “This Just In: Fake News Packs a Lot
in Title, Uses Simpler, Repetitive Content in Text Body, More Similar to Satire
than Real News.” In: url:http://arxiv.org/abs/1703.09398.
Hosseini, Marjan, Alireza Javadian Sabet, Suining He, and Derek Aguiar (2022).
Interpretable Fake News Detection with Topic and Deep Variational Models. arXiv:
2209.01536 [cs.CL].
Hosseinimotlagh, Seyedmehdi and Evangelos E Papalexakis (2018). “Unsupervised
Content-Based Identification of Fake News articles with Tensor Decomposition
Ensembles.” In.
Ito, Jun, Hiroyuki Toda, Yoshimasa Koike, Jing Song, and Satoshi Oyama (2015).
“Assessment of tweet credibility with LDA features.” In: pages 953–958. d oi:
10.1145/2740908.2742569.
Jack, Carloine (2017). “Lexicon of Lies: Terms for Problematic Information.” In:
url:https://datasociety.net/library/lexicon-of-lies.
Kaminska, Izabella (Jan. 2017). “A lesson in fake news from the info-wars of
ancient Rome.” In: Financial Times.url:
https://www.ft.com/content/
aaf2bb08-dca2-11e6-86ac-f253db7791c6.
Karimi, Hamid and Jiliang Tang (2019). “Learning Hierarchical Discourse-level
Structure for Fake News Detection.” In: Association for Computational Lin-
guistics, pages 3432–3442. d oi:
10 . 18653 / v1 / N19 - 1347
.url:
http :
//aclweb.org/anthology/N19-1347.
Kavanagh, Jennifer, William Marcellino, Jonathan S Blake, Shawn Smith, Steven
Davenport, and Mahlet Gizaw (2019). News in a Digital Age: Comparing the
Presentation of News Information over Time and Across Media Platforms. RAND
Corporation. doi:
10 . 7249 / RR2960
.url:
https : / / www . rand . org /
pubs/research_reports/RR2960.html.
Kochkina, Elena, Maria Liakata, and Isabelle Augenstein (Apr. 2017). “Turing at
SemEval-2017 Task 8: Sequential Approach to Rumour Stance Classification
with Branch-LSTM.” In: url:http://arxiv.org/abs/1704.07221.
Kochkina, Elena, Maria Liakata, and Arkaitz Zubiaga (Mar. 2018). “PHEME
dataset for Rumour Detection and Veracity Classification.” In: d o i:
10.6084/
m9.figshare.6392078.v1
.url:
https://figshare.com/articles/
dataset/ PHEME _dataset _for _Rumour_Detection _and _Veracity _
Classification/6392078.
Kolev, Vladislav, Gerhard Weiss, and Gerasimos Spanakis (Feb. 2022). “FOREAL:
RoBERTa Model for Fake News Detection based on Emotions.” In: Proceedings
of the 14th International Conference on Agents and Artificial Intelligence (ICAART
2022). Edited by Ana Paula Rocha, Luc Steels, and Jaap van den Herik. Volume 2.
references 84
Portugal: Scitepress - Science And Technology Publications, pages 429–440.
isb n: 978-989-758-547-0. doi:10.5220/0010873900003116.
Konstantinovskiy, Lev, Oliver Price, Mevan Babakar, and Arkaitz Zubiaga (Sept.
2018). “Towards Automated Factchecking: Developing an Annotation Schema
and Benchmark for Consistent Automated Claim Detection.” In: url:
http:
//arxiv.org/abs/1809.08193.
Kula, Sebastian, Rafał Kozik, Michał Choraś, and Michał Woźniak (June 2021).
“Transformer Based Models in Fake News Detection.” In: Lecture Notes in
Computer Science 12745, pages 28–38. issn: 16113349. d oi:
10.1007/978-
3 - 030 - 77970 - 2 _3 / COVER
.url:
https : / / link . springer . com /
chapter/10.1007/978-3-030-77970-2_3.
Kwon, Sejeong, Meeyoung Cha, and Kyomin Jung (Jan. 2017). “Rumor Detection
over Varying Time Windows.” In: PLOS ONE 12 (1). Edited by Zhong-Ke
Gao, e0168344. d oi:
10 . 1371 / journal . pone . 0168344
.url:
https :
//dx.plos.org/10.1371/journal.pone.0168344.
Le, Quoc V. and Tomas Mikolov (May 2014). “Distributed Representations of
Sentences and Documents.” In: url:http://arxiv.org/abs/1405.4053.
Levy, Omer and Yoav Goldberg (2014). “Neural Word Embedding As Implicit
Matrix Factorization.” In: MIT Press, pages 2177–2185. url:
http:/ / dl .
acm.org/citation.cfm?id=2969033.2969070.
Li, Songqian, Kun Ma, Xuewei Niu, Yufeng Wang, Ke Ji, Ziqiang Yu, and Zhenxiang
Chen (Aug. 2019). “Stacking-based ensemble learning on low dimensional
features for fake news detection.” In: IEEE, pages 2730–2735. doi:
10.1109/
HPCC/SmartCity/DSS.2019.00383
.url:
https://ieeexplore.ieee.
org/document/8855557.
Liang, Gang, Wenbo He, Chun Xu, Liangyin Chen, and Jinquan Zeng (2015).
“Rumor Identification in Microblogging Systems Based on Users’ Behavior.”
In: IEEE Transactions on Computational Social Systems.doi:
10.1109/TCSS.
2016.2517458.
Maiya, Arun S. and Robert M. Rolfe (2015). “Topic similarity networks: Visual
analytics for large document sets.” In: pages 364–372. d o i:
10.1109/BigData.
2014.7004253.
Mann, William C. and Sandra A. Thompson (1988). “Rhetorical Structure Theory:
Toward a functional theory of text organization.” In: Text 8 (3), pages 243–
281. issn: 16134117. doi:
10 . 1515 / text . 1 . 1988 . 8 . 3 . 243
.url:
https : / / www . degruyter . com / view / j / text . 1 . 1988 . 8 . issue -
3/text.1.1988.8.3.243/text.1.1988.8.3.243.xml.
references 85
Manning, Christoper D., Prabhakar Raghavan, and Hinrich Schütze (2008). In-
troduction to Information Retrieval. Cambridge University Press. isb n: 978-0-
521-86571-5. url:https://nlp.stanford.edu/IR-book.
Menczer, Filippo and Thomas Hills (Dec. 2020). “The Attention Economy.” In: Sci-
entific American 6 (323), pages 54–61. d oi:
10.1038/scientificamerican1220-
54
.url:
https://www.scientificamerican.com/article/information-
overload-helps-fake-news-spread-and- social-media-knows-it.
Merriam-Webster Dictionary (Mar. 2017). How Is ’Fake News’ Defined, and When
Will It Be Added to the Dictionary? url:
https://www.merriam-webster.
com/words-at-play/the-real-story-of- fake-news.
Merriam-Webster Dictionary (Mar. 2022). What is ’Truthiness’? url:
https://
www. merriam- webster.com /words- at- play/ truthiness- meaning-
word-origin.
Mikolov, Tomas, Kai Chen, Gregory S. Corrado, and Jeffrey Dean (Jan. 2013b).
“Efficient Estimation of Word Representations in Vector Space.” In.
Mikolov, Tomas, Stefan Kombrink, Lukas Burget, Jan Cernocky, and Sanjeev
Khudanpur (May 2011). “Extensions of recurrent neural network language
model.” In: IEEE, pages 5528–5531. doi:
10.1109/ICASSP.2011.5947611
.
url:http://ieeexplore.ieee.org/document/5947611.
Mikolov, Tomas, Ilya Sutskever, Kai Chen, Greg S. Corrado, and Jeff Dean (2013).
“Distributed Representations of Words and Phrases and their Compositional-
ity.” In: pages 3111–3119. url:
https://papers.nips.cc/paper/5021-
distributed-representations-of-words-andphrases.
Mimno, David, Hanna M. Wallach, Edmund Talley, Miriam Leenders, and An-
drew McCallum (2011). “Optimizing semantic coherence in topic models.” In:
Conference on Empirical Methods in Natural Language Processing, Proceedings
of the Conference (2), pages 262–272. url:
https : / / www . aclweb . org /
anthology/D11-1024.
Mirsky, Yisroel and Wenke Lee (2021). “The Creation and Detection of Deepfakes.”
In: ACM Computing Surveys 54 (1). issn: 15577341. do i:
10.1145/3425780
.
Mitra, T and E Gilbert (2015). “CREDBANK: A large-scale social media corpus
with associated credibility annotations.” In: pages 258–267.
Mohammad, Saif M, Parinaz Sobhani, and Svetlana Kiritchenko (2017). “Stance
and Sentiment in Tweets.” In: volume 17. d o i:
10 . 1145 / 3003433
.url:
https://doi.org/10.1145/3003433.
Mohammad, Saif, Svetlana Kiritchenko, Parinaz Sobhani, Xiaodan Zhu, and
Colin Cherry (May 2016). “A Dataset for Detecting Stance in Tweets.” In:
references 86
European Language Resources Association (ELRA), pages 3945–3952. url:
https://aclanthology.org/L16-1623.
News Media Alliance (June 2019). Google Benefit from News Content. News Media
Alliance. url:
http : / / www . newsmediaalliance . org / wp - content /
uploads/2019/06/Google-Benefit-from-News-Content.pdf.
Newton, Casey (2019). The secret lives of Facebook moderators in America.url:
https : / / www . theverge . com / 2019 / 2 / 25 / 18229714 / cognizant -
facebook-content-moderator-interviews-trauma-working- conditions-
arizona.
Newton, Casey (June 2019b). Bodies in Seats: Facebook moderators break their NDAs
to expose desperate working conditions.url:
https://www.theverge.com/
2019 / 6 / 19 / 18681845 / facebook - moderator - interviews - video -
trauma-ptsd-cognizant-tampa.
Nguyen, Dat Quoc, Thanh Vu, and Anh Tuan Nguyen (Oct. 2020). “BERTweet: A
pre-trained language model for English Tweets.” In: Association for Computa-
tional Linguistics, pages 9–14. d oi:
10.18653/v1/2020.emnlp-demos.2
.
url:https://aclanthology.org/2020.emnlp-demos.2.
Ofcom (July 2021). News consumption in the UK: 2021.url:
https://www.ofcom.
org . uk / research - and - data / tv - radio - and - on - demand / news -
media/news-consumption.
Omar, Muhammad, Byung Won On, Ingyu Lee, and Gyu Sang Choi (Oct. 2015).
“LDA topics: Representation and evaluation.” In: Journal of Information Science
41 (5), pages 662–675. d oi:
10.1177/0165551515587839
.url:
http://
journals.sagepub.com/doi/10.1177/0165551515587839.
Oremus, Will (2017). “Facebook has stopped saying “fake news.” Is “false news”
any better?” In: Slate.url:
https://slate.com/technology/2017/08/
facebook- has - stopped - saying- fake - news- is - false - news- any -
better.html.
Ouali, Yassine, Céline Hudelot, and Myriam Tami (2020). “An Overview of Deep
Semi-Supervised Learning.” In: ArXiv abs/2006.05278. url:
https://arxiv.
org/abs/2006.05278.
Oxford English Dictionary (Sept. 2021). rumour | rumor, n. url:
https://www.
oed.com/view/Entry/168836.
Oxford English Dictionary (Feb. 2022a). news, n. url:
https://www.oed.com/
view/Entry/126615.
Oxford English Dictionary (July 2022b). shockvertising, n. url:
https://www.
oed.com/view/Entry/94769467.
references 87
Oxford Languages (2016). Oxford Word of the Year 2016.url:
https://languages.
oup.com/word-of-the-year/2016/.
Paixão, "Maik, Rinaldo Lima, and Bernard Espinasse" (2020). “Fake News Classifi-
cation and Topic Modeling in Brazilian Portuguese.” In: 2020 IEEE/WIC/ACM
International Joint Conference on Web Intelligence and Intelligent Agent Technology
(WI-IAT), pages 427–432. do i:10.1109/WIIAT50758.2020.00063.
Parikh, Shivam B. and Pradeep K. Atrey (2018). “Media-Rich Fake News Detection:
A Survey.” In: pages 436–441. doi:10.1109/MIPR.2018.00093.
Park, Robert E. (1923). “The Natural History of the Newspaper.” In: American
Journal of Sociology 29 (3), pages 273–289. issn: 15375390.
Pasquetto, Irene, Briony Swire-Thompson, Michelle A Amazeen, Fabrício Ben-
evenuto, Nadia M Brashier, Robert M Bond, Lia C Bozarth, Ceren Budak,
Ullrich K H Ecker, Lisa K Fazio, Emilio Ferrara, Andrew J Flanagin, Alessandro
Flammini, Deen Freelon, Nir Grinberg, Ralph Hertwig, Kathleen Hall Jamieson,
Kenneth Joseph, Jason J Jones, R Kelly Garrett, Daniel Kreiss, Shannon Mc-
Gregor, Jasmine McNealy, Drew Margolin, Alice Marwick, Filippo Menczer,
Miriam J Metzger, Seungahn Nah, Stephan Lewandowsky, Philipp Lorenz-
Spreen, Pablo Ortellado, Gordon Pennycook, Ethan Porter, David G Rand,
Ronald E Robertson, Francesca Tripodi, Soroush Vosoughi, Chris Vargo, Onur
Varol, Brian E Weeks, John Wihbey, Thomas J Wood, and Kai-Cheng Yang (Dec.
2020). “Tackling misinformation: What researchers could do with social media
data.” In: Harvard Kennedy School Misinformation Review 1 (8). d o i:
10.37016/
MR - 2020 - 49
.url:
https : / / misinforeview . hks . harvard . edu /
article/tackling-misinformation-what- researchers- could-do-
with-social-media-data.
Paul, Christopher and Miriam Matthews (2016). “The Russian "Firehose of False-
hood" Propaganda Model: Why It Might Work and Options to Counter It.” In:
RAND Corporation.doi:
10.7249/PE198
.url:
https://www.rand.org/
pubs/perspectives/PE198.html.
Pennington, Jeffrey, Richard Socher, and Christopher Manning (2014). “Glove:
Global Vectors for Word Representation.” In: Association for Computational
Linguistics, pages 1532–1543. d oi:
10 .3115 /v1 / D14- 1162
.url:
http:
//aclweb.org/anthology/D14-1162.
Pennycook, Gordon and David G Rand (2021). “The Psychology of Fake News.”
In: Trends in Cognitive Sciences 25 (5), pages 388–402. issn: 1364-6613. d o i:
https://doi.org/10.1016/j.tics.2021.02.007
.url:
https://www.
sciencedirect.com/science/article/pii/S1364661321000516.
Perrigo, Billy (Feb. 2022). Inside Facebook’s African Sweatshop.url:
https :
/ / time . com / 6147458 / facebook - africa - content - moderation -
employee-treatment.
references 88
Pew Research Center (Sept. 2021). News Consumption Across Social Media in 2021.
url:
https://www.pewresearch.org/journalism/2021/09/20/news-
consumption-across-social-media-in-2021.
Pian, Wenjing, Jianxing Chi, and Feicheng Ma (Nov. 2021). “The causes, impacts
and countermeasures of COVID-19 “Infodemic”: A systematic review using nar-
rative synthesis.” In: Information Processing & Management 58 (6), page 102713.
issn: 0306-4573. doi:10.1016/J.IPM.2021.102713.
Pilditch, Toby D., Jon Roozenbeek, Jens Koed Madsen, and Sander van der Lin-
den (Aug. 2022). “Psychological inoculation can reduce susceptibility to mis-
information in large rational agent networks.” In: Royal Society Open Sci-
ence 9 (8). issn: 2054-5703. d oi:
10 . 1098 / RSOS . 211953
.url:
https :
//royalsocietypublishing.org/doi/10.1098/rsos.211953.
Potthast, Martin, Johannes Kiesel, Kevin Reinartz, Janek Bevendorff, and Benno
Stein (2018). “A stylometric inquiry into hyperpartisan and fake news.” In:
volume 1. Association for Computational Linguistics, pages 231–240. d o i:
10.18653/ v1/p18-1022
.url:
http:// aclweb.org/anthology/P18 -
1022.
Pérez-Rosas, Verónica, Bennett Kleinberg, Alexandra Lefevre, and Rada Mihalcea
(2018). “Automatic Detection of Fake News.” In: Association for Computa-
tional Linguistics, pages 3391–3401. url:
https : / / www . aclweb . org /
anthology/C18-1287.
Rashkin, Hannah, Eunsol Choi, Jin Yea Jang, Svitlana Volkova, and Yejin Choi
(2017). “Truth of Varying Shades: Analyzing Language in Fake News and Po-
litical Fact-Checking.” In: Association for Computational Linguistics (ACL),
pages 2931–2937. i sbn: 9781945626838. d oi:10.18653/V1/D17- 1317.
Raza, Shaina and Chen Ding (May 2022). “Fake news detection based on news
content and social contexts: a transformer-based approach.” In: International
Journal of Data Science and Analytics 13 (4), pages 335–362. issn: 23644168.
doi:
10.1007/S41060-021-00302-Z/FIGURES/11
.url:
https://link.
springer.com/article/10.1007/s41060-021-00302-z.
Röder, Michael, Andreas Both, and Alexander Hinneburg (2015). “Exploring
the space of topic coherence measures.” In: pages 399–408. do i:
10.1145/
2684822.2685324.
Reuters Institute (2021). Digital News Report 2021. Reuters Institute for the Study
of Journalism, Oxford University. url:
https : / / reutersinstitute .
politics.ox.ac.uk/digital-news-report/2021.
Rittel, Horst W.J. and Melvin M. Webber (June 1973). “Dilemmas in a general
theory of planning.” In: Policy Sciences 1973 4:2 4 (2), pages 155–169. d oi:
references 89
10. 1007/BF01405730
.url:
https: //link .springer. com/article /
10.1007/BF01405730.
Rothkopf, David (May 2003). When the Buzz Bites Back.url:
https:/ / www.
washingtonpost .com / archive/ opinions /2003 / 05/ 11 /when - the -
buzz-bites-back/bc8cd84f-cab6-4648-bf58- 0277261af6cd.
Rubin, Victoria L. and Tatiana Lukoianova (May 2015). “Truth and deception
at the rhetorical structure level.” In: Journal of the Association for Information
Science and Technology 66 (5), pages 905–917. d o i:
10.1002/asi.23216
.
url:http://doi.wiley.com/10.1002/asi.23216.
Rubin, Victoria (2019). News Verification Project: Datasets for Share.url:
http:
//victoriarubin.fims.uwo.ca/news-verification/data-to-go.
Rubin, Victoria, Niall Conroy, Yimin Chen, and Sarah Cornwell (June 2016).
“Fake News or Truth? Using Satirical Cues to Detect Potentially Misleading
News.” In: Association for Computational Linguistics, pages 7–17. doi:
10.
18653/v1/W16-0802.url:https://aclanthology.org/W16-0802.
Ruchansky, Natali, Sungyong Seo, and Yan Liu (2017). “CSI: A Hybrid Deep
Model for Fake News Detection.” In: Proceedings of the 2017 ACM on Conference
on Information and Knowledge Management - CIKM ’17, pages 797–806. d o i:
10.1145/3132847.3132877.
Salem, Fatima Abu, Roaa Al Feel, Shady Elbassuoni, Mohamad Jaber, and May
Farah (Jan. 2019). “Dataset for fake news and articles detection.” In: doi:
10.
5281/ZENODO.2532642.url:https://zenodo.org/record/2532642.
Schuster, Tal, R Schuster, Darsh J Shah, and Regina Barzilay (June 2020). “The
Limitations of Stylometry for Detecting Machine-Generated Fake News.” In:
Computational Linguistics 46 (2), pages 499–510. do i:
10.1162/coli_a_
00380.
Shang, Jingbo, Jiaming Shen, Tianhang Sun, Xingbang Liu, Anja Gruenheid, Flip
Korn, Adam D. Lelkes, Cong Yu, and Jiawei Han (2018). “Investigating Rumor
News Using Agreement-Aware Search.” In: ACM Press, pages 2117–2125. d oi:
10.1145/3269206.3272020.
Shi, Hanyu, Martin Gerlach, Isabel Diersen, Doug Downey, and Luis A. N. Amaral
(Jan. 2019). “A new evaluation framework for topic modeling algorithms based
on synthetic corpora.” In: url:http://arxiv.org/abs/1901.09848.
Shu, Kai and Huan Liu (2019b). Detecting Fake News on Social Media. Springer
Cham. isb n: 978-3-031-01915-9. doi:10.1007/978-3-031- 01915-9.
Shu, Kai, Deepak Mahudeswaran, Suhang Wang, Dongwon Lee, and Huan Liu
(Sept. 2018). “FakeNewsNet: A Data Repository with News Content, Social
references 90
Context and Dynamic Information for Studying Fake News on Social Media.”
In: url:http://arxiv.org/abs/1809.01286.
Shu, Kai, Deepak Mahudeswaran, Suhang Wang, Dongwon Lee, and Huan Liu
(2020). “FakeNewsNet: A Data Repository with News Content, Social Context,
and Spatiotemporal Information for Studying Fake News on Social Media.”
In: Big Data 8 (3), pages 171–188. d o i:
10.1089 / big . 2020.0062
.url:
https://doi.org/10.1089/big.2020.0062.
Shu, Kai, Deepak Mahudeswaran, Suhang Wang, and Huan Liu (May 2020c).
“Hierarchical Propagation Networks for Fake News Detection: Investigation
and Exploitation.” In: Proceedings of the International AAAI Conference on Web
and Social Media 14 (1), pages 626–637. do i:
10.1609/icwsm.v14i1.7329
.
url:https://ojs.aaai.org/index.php/ICWSM/article/view/7329.
Shu, Kai, Amy Sliva, Suhang Wang, Jiliang Tang, and Huan Liu (2017). “Fake News
Detection on Social Media.” In: ACM SIGKDD Explorations Newsletter 19 (1),
pages 22–36. d oi:10.1145/3137597.3137600.
Shu, Kai, Suhang Wang, and Huan Liu (2019). “Beyond News Contents: The Role
of Social Context for Fake News Detection.” In: Proceedings of the Twelfth ACM
International Conference on Web Search and Data Mining 9, pages 312–320.
doi:
10.1145/3289600
.url:
https://doi.org/10.1145/3289600.
3290994.
Shu, Kai, Xinyi Zhou, Suhang Wang, Reza Zafarani, and Huan Liu (2020b). “The
Role of User Profiles for Fake News Detection.” In: Proceedings of the 2019
IEEE/ACM International Conference on Advances in Social Networks Analysis
and Mining. ASONAM ’19. Vancouver, British Columbia, Canada: Associa-
tion for Computing Machinery, 436–439. is b n: 9781450368681. do i:
10.
1145/ 3341161 .3342927
.url:
https: / /doi.org /10.1145 / 3341161.
3342927.
Sisodia, Dilip Singh (2019). “Ensemble learning approach for clickbait detection
using article headline features.” In: Informing Science 22 (2019), pages 31–44.
doi:10.28945/4279.
Socher, Richard, Alex Perelygin, Jean Y.Wu, Jason Chuang, Christopher D. Man-
ning, Andrew Y. Ng, and Christopher Potts (2013). “Recursive Deep Models
for Semantic Compositionality Over a Sentiment Treebank.” In: PLoS ONE.
doi:10.1371/journal.pone.0073791.
Soll, Jacob (2016). The Long and Brutal History of Fake News.url:
https: //
www.politico.com/magazine/story/2016/12/fake-news-history-
long-violent-214535.
references 91
Stevens, Keith, Philip Kegelmeyer, David Andrzejewski, and David Buttler (2012).
“Exploring topic coherence over many models and many topics.” In: pages 952–
961.
Syed, Shaheen and Marco Spruit (Oct. 2018). “Full-Text or abstract? Examining
topic coherence scores using latent dirichlet allocation.” In: volume 2018-
January. IEEE, pages 165–174. d o i:
10.1109/DSAA.2017.61
.url:
http:
//ieeexplore.ieee.org/document/8259775.
Tandoc, Edson C., Zheng Wei Lim, and Richard Ling (Feb. 2018). “Defining
“Fake News”: A typology of scholarly definitions.” In: Digital Journalism 6 (2),
pages 137–153. d oi:
10.1080/21670811.2017.1360143
.url:
https://
www.tandfonline.com/doi/full/10.1080/21670811.2017.1360143.
Torres, Russell, Natalie Gerhart, and Arash Negahban (July 2018). “Epistemology
in the Era of Fake News.” In: ACM SIGMIS Database: the DATABASE for
Advances in Information Systems 49 (3), pages 78–97. do i:
10.1145/3242734.
3242740
.url:
http:/ / dl . acm . org/citation. cfm ? doid = 3242734 .
3242740.
U.S. Department of Homeland Security (2018). Countering False Information on
Social Media in Disasters and Emergencies. U.S. Department of Homeland Secu-
rity. url:
https://www.dhs.gov/publication/st-frg- countering-
false-information-social-media-disasters-and- emergencies.
U.S. Department of State (Jan. 2022). Russia’s Top Five Persistent Disinforma-
tion Narratives.url:
https :/ / www. state .gov /russias - top - five-
persistent-disinformation-narratives.
Vijjali, Rutvik, Prathyush Potluri, Siddharth Kumar, and Sundeep Teki (Dec. 2020).
“Two Stage Transformer Model for COVID-19 Fake News Detection and Fact
Checking.” In: International Committee on Computational Linguistics (ICCL),
pages 1–10. url:https://aclanthology.org/2020.nlp4if-1.1.
Vinck, Patrick, Phuong N. Pham, Kenedy K. Bindu, Juliet Bedford, and Eric J.
Nilles (May 2019). “Institutional trust and misinformation in the response to
the 2018–19 Ebola outbreak in North Kivu, DR Congo: a population-based
survey.” In: The Lancet Infectious Diseases 19 (5), pages 529–536. issn: 14744457.
doi:
10.1016/S1473-3099(19)30063-5
.url:
http://www.thelancet.
com/article/S1473309919300635/fulltext.
Vosoughi, Soroush, Deb Roy, and Sinan Aral (Mar. 2018). “The spread of true and
false news online.” In: Science (New York, N.Y.) 359 (6380), pages 1146–1151.
doi:
10.1126/science.aap9559
.url:
http://www.sciencemag.org/
lookup/doi/10.1126/science.aap9559.
Waldman, Ari Ezra (2018). “The Marketplace of Fake News.” In: University of
Pennsylvania Journal of Constitutional Law 20, pages 846 –869. doi:
10.4135/
references 92
9781604265774 . n911
.url:
http : / / sk . sagepub . com / cqpress /
encyclopedia-of-the-first-amendment/n911.xml.
Wang, Shihan, Izabela Moise, Dirk Helbing, and Takao Terano (July 2017). “Early
Signals of Trending Rumor Event in Streaming Social Media.” In: IEEE, pages 654–
659. doi:
10.1109/COMPSAC.2017.115
.url:
http://ieeexplore.ieee.
org/document/8030007.
Wang, Tai-Li (2012). “Presentation and impact of market-driven journalism on
sensationalism in global TV news.” In: International Communication Gazette
74 (8), pages 711–727. d oi:
10 . 1177 / 1748048512459143
.url:
https :
//doi.org/10.1177/1748048512459143.
Wang, William Yang (2017b). “"Liar, Liar Pants on Fire": A New Benchmark
Dataset for Fake News Detection.” In: Proceedings of the 55th Annual Meet-
ing of the Association for Computational Linguistics (Volume 2: Short Papers) 2,
pages 422–426. d oi:
10.18653/v1/P17-2067
.url:
http://aclweb.org/
anthology/P17-2067http://arxiv.org/abs/1705.00648.
Wang, Yaqing, Fenglong Ma, Zhiwei Jin, Ye Yuan, Guangxu Xun, Kishlay Jha,
Lu Su, and Jing Gao (2018). “EANN: Event Adversarial Neural Networks
for Multi-Modal Fake News Detection.” In: ACM Press, pages 849–857. d oi:
10.1145/3219819.3219903.
Wardle, Claire (2017). “Fake News. It’s Complicated.” In: First Draft.url:
https:
//firstdraftnews.org/fake-news-complicated.
Wardle, Claire (2018). “Information Disorder: The Essential Glossary.” In: First
Draft.url:
https : / / firstdraftnews . org / wp - content / uploads /
2018/07/infoDisorder_glossary.pdf.
Wardle, Claire (2020). “Understanding Information Disorder.” In: First Draft.url:
https://firstdraftnews.org/long-form-article/understanding-
information-disorder.
Wardle, Claire and Hossein Derakhshan (2017b). “Information Disorder: Toward
an interdisciplinary framework for research and policy making.” In: Council of
Europe report, DGI (2017).
Winick, Erin (July 2018). Facebook’s latest acquisition is all about fighting fake news.
url:
https: / /www . technologyreview. com /2018 / 07/ 02 /141613 /
facebooks - latest- acquisition - is - all- about - fighting - fake-
news.
Wolfram Research (2021a). Classify. Wolfram Research. url:
https://reference.
wolfram.com/language/ref/Classify.html.
references 93
Wolfram Research (2021b). “Linear” (Machine Learning Method). Wolfram Re-
search. url:
https : / / reference . wolfram . com / language / ref /
method/Linear.html.
Wright, Susan (Nov. 2017). Collins 2017 Word of the Year Shortlist.url:
https:
//blog.collinsdictionary.com/language-lovers/collins- 2017-
word-of-the-year-shortlist.
Wu, Ke, Song Yang, and Kenny Q. Zhu (2015). “False rumors detection on Sina
Weibo by propagation structures.” In: volume 2015-May, pages 651–662. d oi:
10.1109/ICDE.2015.7113322.
Wu, Liang and Huan Liu (2018). “Tracing fake-news footprints: Characteriz-
ing social media messages by how they propagate.” In: volume 2018-Febua,
pages 637–645. d oi:10.1145/3159652.3159677.
Yang, Yang, Lei Zheng, Jiawei Zhang, Qingcai Cui, Xiaoming Zhang, Zhoujun Li,
and Philip S Yu (June 2018). “TI-CNN: Convolutional Neural Networks for
Fake News Detection.” In: doi:10.48550/arxiv.1806.00749.
Yang, Yuting, Juan Cao, Mingyan Lu, Jintao Li, and Chia-Wen Lin (Feb. 2019).
“How to Write High-quality News on Social Network? Predicting News Quality
by Mining Writing Style.” In: d o i:10.48550/arxiv.1902.00750.
Yoon, Seunghyun, Kunwoo Park, Joongbo Shin, Hongjun Lim, Seungpil Won,
Meeyoung Cha, and Kyomin Jung (2019). “Detecting Incongruity between
News Headline and Body Text via a Deep Hierarchical Encoder.” In: Proceedings
of the AAAI Conference on Artificial Intelligence 33, pages 791–800. do i:
10.
1609/aaai.v33i01.3301791.
Zafarani, Reza, Xinyi Zhou, Kai Shu, and Huan Liu (2019). “Fake News Research:
Theories, Detection Strategies, and Open Problems.” In: Proceedings of the 25th
ACM SIGKDD International Conference on Knowledge Discovery Data Mining.
doi:10.1145/3292500.3332287.
Zannettou, Savvas, Michael Sirivianos, Jeremy Blackburn, and Nicolas Kourtellis
(Apr. 2019). “The web of false information: Rumors, fake news, hoaxes, clickbait,
and various other shenanigans.” In: Journal of Data and Information Quality
11 (3). d oi:
10.1145/3309699
.url:
http://dx.doi.org/10. 1145 /
3309699.
Zhang, Amy X., Martin Robbins, Ed Bice, Sandro Hawke, David Karger, An Xiao
Mina, Aditya Ranganathan, Sarah Emlen Metz, Scott Appling, Connie Moon
Sehat, Norman Gilmore, Nick B. Adams, Emmanuel Vincent, and Jennifer Lee
(2018). “A Structured Response to Misinformation: Defining and Annotating
Credibility Indicators in News articles.” In: WWW-2018, pages 603–612. d o i:
10.1145/3184558.3188731.
references 94
Zhang, Jiawei, Bowen Dong, and Philip S. Yu (Apr. 2020). “FakeDetector: Effective
fake news detection with deep diffusive neural network.” In: Proceedings -
International Conference on Data Engineering 2020-April, pages 1826–1829.
doi:10.1109/ICDE48307.2020.00180.
Zhang, Qiang, Shangsong Liang, Aldo Lipani, and Emine Yilmaz (May 2019).
“Reply-aided detection of misinformation via Bayesian deep learning.” In: The
Web Conference 2019 - Proceedings of the World Wide Web Conference, WWW
2019, pages 2333–2343. d oi:10.1145/3308558.3313718.
Zhang, Yan, Weiling Chen, Chai Kiat Yeo, Chiew Tong Lau, and Bu Sung Lee
(Oct. 2016). “A distance-based outlier detection method for rumor detection
exploiting user behaviorial differences.” In: IEEE, pages 1–6. doi:
10.1109/
ICODSE.2016.7936102.
Zhang, Yan, Weiling Chen, Chai Kiat Yeo, Chiew Tong Lau, and Bu Sung Lee
(July 2017). “Detecting rumors on Online Social Networks using multi-layer
autoencoder.” In: 2017 IEEE Technology and Engineering Management Society
Conference, TEMSCON 2017, pages 437–441. d o i:
10.1109/TEMSCON.2017.
7998415.
Zhou, Xinyi, Atishay Jain, Vir V. Phoha, and Reza Zafarani (June 2020b). “Fake
News Early Detection: A Theory-driven Model.” In: Digital Threats: Research
and Practice 1 (2). issn: 25765337. d oi:
10. 1145 /3377478
.url:
https:
//dl.acm.org/doi/10.1145/3377478.
Zhou, Xinyi and Reza Zafarani (2018). “Fake News: A Survey of Research, Detec-
tion Methods, and Opportunities.” In: url:
http://arxiv.org/abs/1812.
00315.
Zhou, Xinyi and Reza Zafarani (Nov. 2019). “Network-Based Fake News De-
tection: A Pattern-Driven Approach.” In: SIGKDD Explor. Newsl. 21.2, 48–60.
issn: 1931-0145. doi:
10.1145/3373464.3373473
.url:
https://doi.
org/10.1145/3373464.3373473.
Zhou, Xinyi and Reza Zafarani (2020). “A Survey of Fake News: Fundamental
Theories, Detection Methods, and Opportunities.” In: ACM Computing Surveys
(CSUR) 53 (5), pages 1–40. issn: 15577341. d o i:10.1145/3395046.
Zubiaga, Arkaitz, Ahmet Aker, Kalina Bontcheva, Maria Liakata, and Rob Procter
(Apr. 2018). “Detection and resolution of rumours in social media: A survey.”
In: ACM Computing Surveys 51 (2). do i:
10.1145/3161603
.url:
http://
arxiv.org/abs/1704.00656http://dx.doi.org/10.1145/3161603.
Zubiaga, Arkaitz, Geraldine Wong Sak Hoi, Maria Liakata, and Rob Procter
(2016b). “PHEME dataset of rumours and non-rumours.” In: doi:
10.6084/
M9.FIGSHARE.4010619.V1
.url:
https://figshare.com/articles/
dataset/PHEME_dataset_of_rumours_and_non-rumours/4010619/1.
references 95
Zubiaga, Arkaitz, Maria Liakata, and Rob Procter (Oct. 2016). “Learning Reporting
Dynamics during Breaking News for Rumour Detection in Social Media.” In:
url:http://arxiv.org/abs/1610.07363.
Zubiaga, Arkaitz, Maria Liakata, Rob Procter, Geraldine Wong Sak Hoi, and Peter
Tolmie (Mar. 2016c). “Analysing How People Orient to and Spread Rumours
in Social Media by Looking at Conversational Threads.” In: PLOS ONE 11
(3). Edited by Naoki Masuda, e0150989. do i:
10 . 1371 / journal . pone .
0150989
.url:
https : / / dx . plos . org / 10 . 1371 / journal . pone .
0150989.
van Dijk, Teun A. (1983). “Discourse Analysis: Its Development and Application
to the Structure of News.” In: Journal of Communication 33.2, pages 20–43. d oi:
https: / /doi .org /10 . 1111/ j. 1460 - 2466. 1983. tb02386. x
. eprint:
https: / / onlinelibrary . wiley .com / doi /pdf / 10 . 1111 / j. 1460 -
2466.1983.tb02386.x
.url:
https://onlinelibrary.wiley . com /
doi/abs/10.1111/j.1460-2466.1983.tb02386.x.
