Automatic Deception Detection: Methods for Finding Fake News

Abstract

This research surveys the current state-of-the-art technologies that are instrumental in the adoption and development of fake news detection. " Fake news detection " is defined as the task of categorizing news along a continuum of veracity, with an associated measure of certainty. Veracity is compromised by the occurrence of intentional deceptions. The nature of online news publication has changed, such that traditional fact checking and vetting from potential deception is impossible against the flood arising from content generators, as well as various formats and genres. The paper provides a typology of several varieties of veracity assessment methods emerging from two major categories – linguistic cue approaches (with machine learning), and network analysis approaches. We see promise in an innovative hybrid approach that combines linguistic cue and machine learning, with network-based behavioral data. Although designing a fake news detector is not a straightforward problem, we propose operational guidelines for a feasible fake news detecting system.

Full text

Automatic Deception Detection: Methods for Finding Fake
News
Niall J. Conroy, Victoria L. Rubin, and Yimin Chen
Language and Information Technology Research Lab (LIT.RL)
Faculty of Information and Media Studies
University of Western Ontario, London, Ontario, CANADA
nconroy1@uwo.ca, vrubin@uwo.ca, ychen582@uwo.ca
!
ABSTRACT
This research surveys the current state-of-the-art
technologies that are instrumental in the adoption and
development of fake news detection. “Fake news detection”
is defined as the task of categorizing news along a
continuum of veracity, with an associated measure of
certainty. Veracity is compromised by the occurrence of
intentional deceptions. The nature of online news
publication has changed, such that traditional fact checking
and vetting from potential deception is impossible against
the flood arising from content generators, as well as various
formats and genres.
The paper provides a typology of several varieties of
veracity assessment methods emerging from two major
categories – linguistic cue approaches (with machine
learning), and network analysis approaches. We see promise
in an innovative hybrid approach that combines linguistic
cue and machine learning, with network-based behavioral
data. Although designing a fake news detector is not a
straightforward problem, we propose operational guidelines
for a feasible fake news detecting system.
Keywords
Deception detection, fake news detection, veracity
assessment, news verification, methods, automation, SVM,
knowledge networks, predictive modelling, fraud
INTRODUCTION
News verification aims to employ technology to identify
intentionally deceptive news content online, and is an
important issue within certain streams of library and
information science (LIS). Fake news detection is defined
as the prediction of the chances of a particular news article
(news report, editorial, expose, etc.) being intentionally
deceptive (Rubin, Conroy & Chen, 2015). Tools aim to
mimic certain filtering tasks which have, to this point, been
the purview of journalists and other publishers of traditional
news content. The proliferation of user-generated content,
and Com put er Media ted Com mun i cat ion (CM C )
technologies such as blogs, Twitter, and other social media
have the capacity of news delivery mechanisms on a mass
scale— yet much of the information is of questionable
veracity (Ciampaglia, Shiralkar, Rocha, Bollen, Menczer &
Flammini, 2015). Establishing the reliability of information
online is a daunting but critical challenge. Four decades of
deception detection research has helped us learn about how
well humans are able detect lies in text. The findings show
we are not so good at it. In fact, just 4% better than chance,
based on a meta-analysis of more than 200 experiments
(Bond & DePaulo, 2006). This problem has led researchers
and technical developers to look at several automated ways
of assessing the truth value of potentially deceptive text
based on the properties of the content and the patterns of
computer-mediated communication .
Structured datasets are easier to verify than non-structured
(or semi-structured) data such as texts. When we know the
language domain (e.g., insurance claims or health-related
news) we can make better guesses about the nature and use
of deception. Semi-structured non-domain specific web
data come in many formats and demand flexible methods
for veracity verification. For some time, however, the
development and evaluation of different methods have
remained in isolated corners, relatively unknown in LIS.
More recently, efforts of methodological cross-pollination
and hybrid approaches have produced promising results
(Rubin et al., 2015A). The range of journalistic practices
and available news sources (see Rubin et al. (2015B) for an
overview) demand consideration of multiple methods since
one approach often addresses known weaknesses in another.
How then is it possible to gauge the veracity of online
news?
This paper provides researchers with a map of the current
landscape of veracity (or deception) assessment methods,
their major classes and goals, all with the aim of proposing
a hybrid approach to system design. These methods have
emerged from separate development streams, utilizing
disparate techniques. In this survey, two major categories of
methods emerge: 1. Linguistic Approaches in which the
content of deceptive messages is extracted and analyzed to
associate language patterns with deception; and 2. Network
Approaches in which network information, such as message
metadata or structured knowledge network queries can be
harnessed to provide aggregate deception measures. Both
forms typically incorporate machine learning techniques for
training classifiers to suit the analysis. It is incumbent upon
researchers to understand these different areas, yet no
known typology of methods exists in the current literature.
The goal is to provide a survey of the existing research
while proposing a hybrid approach, which utilizes the most
eff e c t i v e dec e p t i o n dete c t i o n meth o d s for th e
implementation of a fake news detection tool.
!
ASIST 2015, November 6-10, 2015, St. Louis, MO, USA.
Copyright © 2015 Niall J. Conroy, Victoria L. Rubin & Yimin Chen

LINGUISTIC APPROACHES
Most liars use their language strategically to avoid being
caught. In spite of the attempt to control what they are
saying, language “leakage” occurs with certain verbal
aspects that are hard to monitor such as frequencies and
patterns of pronoun, conjunction, and negative emotion
word usage (Feng & Hirst, 2013). The goal in the linguistic
approach is to look for such instances of leakage or, so
called “predictive deception cues” found in the content of a
message.
Data Representation
Perhaps the simplest method of representing texts is the
“bag of words” approach, which regards each word as a
single, equally significant unit. In the bag of words
approach, individual words or “n-grams” (multiword)
frequencies are aggregated and analyzed to reveal cues of
deception. Further tagging of words into respective lexical
cu es for e xam ple, pa rts of s pee ch or “s hal low
syntax” (Hancock & Markowitz, 2013), affec ti ve
dimensions (Vrij, 2006), or location-based words (Hancock,
et al, 2013) are all ways of providing frequency sets to
reveal linguistic cues of deception.
The simplicity of this representation also leads to its biggest
shortcoming. In addition to relying exclusively on
language, the method relies on isolated n-grams, often
divorced from useful context information. In this method,
any resolution of ambiguous word sense remains non-
existent (Larcker & Zakolyukina 2012). Many deception
detection researchers have found this method useful in
tandem with different, complementary analysis (Zhang,
Fan, Zeng & Liu, 2012; Lary, Nikitov & Stone, 2010; Ott,
Cardi, & Hancock, 2013), several of which are discussed in
the remainder of this proposal.
Deep Syntax
Analysis of word use is often not enough in predicting
deception. Deeper language structures (syntax) have been
analyzed to predict instances of deception. Deep syntax
analysis is implemented through Probability Context Free
Grammars (PCFG). Sentences are transformed to a set of
rewrite rules (a parse tree) to describe syntax structure, for
example noun and verb phrases, which are in turn rewritten
by their syntactic constituent parts (Feng, Banerjee & Choi,
2012). The final set of rewrites produces a parse tree with a
certain probability assigned. This method is used to
distinguish rule categories (lexicalized, unlexicalized,
parent nodes, etc.) for deception detection with 85-91%
accuracy (depending on the rule category used) (Feng et al.,
2012).
Third-party tools, such as the Stanford Parser (de Marneffe,
MacCartney, Manning, 2006; Rahangdale & Agrawa,
2014), AutoSlog-TS syntax analyzer (Oraby, Reed,
Compton, Riloff, Walker, & Whittaker, 2015) and others
assist in the automation. Alone, syntax analysis might not
be sufficiently capable of identifying deception, and studies
often combine this approach with other linguistic or
network analysis techniques (e.g., Feng et al., 2012; Feng &
Hirst, 2013).
Semantic Analysis
As an alternative to deception cues, signals of truthfulness
have also been analyzed and achieved by characterizing the
degree of compatibility between a personal experience (e.g.,
a hotel review) as compared to a content “profile” derived
from a collection of analogous data. This approach extends
the n-gram plus syntax model by incorporating profile
compatibility features, showing the addition significantly
improves classification performance. (Feng & Hirst, 2013).
The intuition is that a deceptive writer with no experience
with an event or object (e.g., never visited the hotel in
question) may include contradictions or omission of facts
present in profiles on similar topics. For product reviews, a
writer of a truthful review is more likely to make similar
comments about aspects of the product as other truthful
reviewers. Extracted content from key words consists of
attribute:descriptor pair. By aligning profiles and the
description of the writer’s personal experience, veracity
assessment is a function of the compatibility scores: 1.
Compatibility with the existence of some distinct aspect
(eg. an art museum near the hotel); 2. Compatibility with
the description of some general aspect, such as location or
serv ic e. Pr ed iction of f al se ho od is shown to be
approximately 91% accurate with this method.
Although demonstrated useful in the above context of
reviews, this method has so far been restricted to the
domain of application. There are two potential limitations
in this method: the ability to determine alignment between
Figure 1: Fact-checking statements. (a) Structured information about President
Obama contained in the “infoboxes” of Wikipedia articles. (b) Shortest knowl-
edge graph path returned for the false statement “Barack Obama is a Muslim”.
The path traverses high-degree nodes representing generic entities, such as
Canada, and is assigned a low truth value. (Ciampiaglia et al., 2015)

attributes and descriptors depends on a sufficient amount of
mined content for profiles, and the challenge of correctly
associating descriptors with extracted attributes.
Rhetorical Structure and Discourse Analysis
At the discourse level, deception cues present themselves
both in CMC communication and in news content. A
description of discourse can be achieved through the
Rhetorical Structure Theory (RST) analytic framework, that
identifies instances of rhetoric relations between linguistic
elements. Systematic differences between deceptive and
truthful messages in terms of their coherence and structure
has been combined with a Vector Space Model (VSM) that
assesses each message’s position in multi-dimensional RST
space with respect to its distance to truth and deceptive
centers (Rubin & Lukoianova, 2014). At this level of
linguistic analysis, the prominent use of certain rhetorical
relations can be indicative of deception. Tools to automate
rhetorical classification are becoming available, although
not yet employed in the context of veracity assessment.
Classifiers
Sets of word and category frequencies are useful for
subsequent automated numerical analysis. One common use
is for the training of “classifiers” as in Support Vector
Machines (SVM) (Zhang et al., 2012) and Naïve Bayesian
models (Oraby et al., 2015). Simply put, when a
mathematical model is sufficiently trained from pre-coded
examples in one of two categories, it can predict instances
of future deception on the basis of numeric clustering and
distances. The use of different clustering methods and
distance functions between data points shape the accuracy
of SVM (Strehl, Ghosh & Mooney, 2000), which invites
new experimentation on the net effect of these variables.
Naïve Bayes algorithms make classifications based on
accumulated evidence of the correlation between a given
variable (e.g., syntax) and the other variables present in the
model (Mihalcea & Strapparava, 2009).
The classification of sentiment (Pang & Lee, 2008; Ott et
al., 2013) is based on the underlying intuition that deceivers
use unintended emotional communication, judgment or
evaluation of affective state (Hancock, Woodworth, &
Porter, 2011). Likewise, syntactic patterns may be used in
distinguishing feeling from fact-based arguments by
associating learned patterns of argumentation style classes.
In studies of business communication, performance is
significantly better than a random guess by 16%, and the
language of deceptive executives exhibits fewer non-
extreme positive emotions (Larcker & Zakolyukina, 2012).
Compari son b etween hum an judgement and SVM
classifiers showed 86% performance accuracy on negative
deceptive opinion spam (Ott et al., 2013). Fake negative
reviewers over-produced negative emotion terms relative to
the truthful reviews. These were deemed not the result of
“leakage cues” from the emotional distress of lying, but
exaggerations of the sentiment deceivers are trying to
convey.
These linguistic approaches all rely on language usage and
its analysis, and are promising when used in hybrid
approaches. However, findings emerging from topic-
specific studies (product reviews, business) may have
limited generalizability towards real-time veracity detection
of news.
NETWORK APPROACHES
Innovative and varied, using network properties and
behavior are ways to complement content-based approaches
that rely on deceptive language and leakage cues to predict
deception. As real-time content on current events is
in cre asi ngl y pr oli fer ate d thro ugh mic ro- blo ggi ng
applications such as Twitter, deception analysis tools are all
the more important.
Linked data
The use of knowledge networks may represent a significant
step towards scalable computational fact-checking methods.
For certain data, false “factual statements” can represent a
form of deception since they can be extracted and examined
alongside findable statements about the known world. This
approach leverages an existing body of collective human
knowledge to assess the truth of new statements. The
method depends on querying existing knowledge networks,
or publicly available structured data, such as DBpedia
ontology, or the Google Relation Extraction Corpus
(GREC).
The inherently structured data network of entities is
connected through a predicate relationship. Fact checking
can be effectively reduced to a simple network analysis
problem: the computation of the simple shortest path (see
Figure 1). Queries based on extracted fact statements are
assigned semantic proximity as a function of the transitive
relationship between subject and predicate via other nodes.
The closer the nodes, the higher the likelihood that a
particular subject-predicate-object statement is true.
There are several so-called ‘network effect’ variables that
are exploited to derive truth probabilities (Ciampaglia et al.,
2015), so the outlook for exploiting structured data
repositories for fact-checking remains promising. From the
short list of existing published work in this area, results
using sample facts from four different subject areas range
from 61% to 95%. Success was measured based on whether
the machine was able to assign higher true values to true
statements than to false ones (Ciampaglia, et al., 2015). A
problem with this method, however, rests in the fact that
statements must reside in a pre-existing knowledge base.
Social Network Behavior
Authentication of identity on social media is paramount to
the notion of trust. The proliferation of news in the form of
current events through mass technologies like micro-blogs
invites ways of ascertaining the difference between fake
and genuine content. Outside of the analysis of content
comes the use of metadata and telltale behavior of
questionable sources (Chu, Gianvecchio, Wang & Jajodia,
2010). The recent use of twitter in influencing political
perceptions (Cook et al., 2013) is one scenario where
certain data, namely the inclusion of hyperlinks or
associated metadata, can be compiled to establish veracity
assessments. Centering resonance analysis (CRA), a mode
of network-based text analysis, represents the content of

large sets of texts by identifying the most important words
that link other words in the network. This was employed by
Papacharissi & Oliviera to identify content patterns in posts
about Egypt’s elections (2012). Combining sentiment and
behaviour studies have demonstrated the contention that
sentiment-focused reviews from singleton contributors
significantly affects online ranking (Wu, Greene, Smyth &
Cunningham, 2010), and that this is an indicator of
“shilling” or contributing fake reviews to artificially distort
a ranking.
CONCLUSION
Linguistic and network-based approaches have shown high
accuracy results in classification tasks within limited
domains. This discussion drafts a basic typology of
methods available for further refinement and evaluation,
and provides a basis for the design of a comprehensive fake
news detection tool. Techniques arising from disparate
approaches may be utilized together in a hybrid system,
whose features are summarized:
• Linguistic processing should be built on multiple layers
from word/lexical analysis to highest discourse-level
analysis for maximum performance.
• As a viable alternative to strictly content-based
approaches, network behavior should be combined to
incorporate the ‘trust’ dimension by identifying credible
sources.
• Tools should be designed to augment human judgement,
not replace it. Relations between machine output and
methods should be transparent.
• Contributions in the form of publicly available gold
standard datasets should be in linked data format to assist
in up-to-date fact checking.
ACKNOWLEDGMENTS
This research has been funded by the Government of
Canada Social Sciences and Humanities Research Council
(SSHRC) Insight Grant (#435-2015-0065) awarded to Dr.
Rubin for the project entitled Digital Deception Detection:
Identifying Deliberate Misinformation in Online News.
REFERENCES
Ciampaglia, G., Shiralkar, P., Rocha, L., Bollen, J. Menczer, F., &
Flammini, A. (2015). Computational fact checking from
knowledge networks.
Chen, Y., Conroy, N. J., & Rubin, V. L. (2015). News in an Online
World: The Need for an “Automatic Crap Detector”. In The
Proceedings of the Association for Information Science and
Technology Annual Meeting (ASIST2015), Nov. 6-10, St. Louis.
Chu, Z., Gianvecchio, S., Wang, H., & Jajodia, S. (2010). Who is
tweeting on Twitter: Human, Bot, or Cyborg? in the
Proceedings of the 26th Annual Computer Security Applications
Conference, ACSAC ’10, pp. 21-30.
Cook, D., Waugh, B., Abdipanab, M, Hashemi, O., Rahman, S.
(2013). Twitter Deception and Influence: Issues of Identity,
Slacktivism and Puppetry
de Marneffe, M., MacCartney, B. & Manning, C. (2006).
Generating typed dependency parses from phrase structure
parses. In Proceedings of the 5th International Conference on
Language Resources and Evaluation (LREC 2006).
Hancock, J., Woodworth, M. & Porter, S. (2011). Hungry like a
wolf: A word pattern analysis of the language of psychopaths.
Legal and Criminological Psychology. 113.
Hancock, J. & Markowitz, D. (2014). Linguistic Traces of a
Scientific Fraud: The Case of Diederik Stapel. PLoS ONE,9(8)
Feng, V. & Hirst, G. (2013) Detecting deceptive opinion with
profile compatibility.
Feng, S., Banerjee, R. & Choi, Y. (2012). Syntactic Stylometry for
Deception Detection. 50th Annual Meeting of the Association
for Computational Linguistics. Association for Computational
Linguistics, 171–175.
Larcker, D., Zakolyukina, A. (2012). Detecting Deceptive
Discussions in Conference Calls. Journal of Accounting
Research, 50(2), 495540.
Mihalcea, R. & Strapparava, C. (2009). The Lie Detector:
Explorations in the Automatic Recognition of Deceptive
Language. Proceedings of the ACL-IJCNLP Conference Short
Papers, pp. 309–312,
Oraby, S., Reed, L., Compton, R., Riloff, E., Walker, M. &
Whittaker, S. (2015). And That’s A Fact: Distinguishing Factual
and Emotional Argumentation in Online Dialogue
Ott, M., Cardie, C. & Hancock, J. (2013). Negative Deceptive
Opinion Spam. Proceedings of NAACLHLT. pp. 497–501,
Pang, B. & Lee, L. (2008). Opinion mining and sentiment
analysis. Foundations and Trends in Information Retrieval,
2(1-2), pp. 1–135.
Papacharissi, Z. & Oliveira, M. (2012).The Rhythms of News
Storytelling on #Egypt. Journal of Communication. 62. pp.
266–282.
Rahangdale, A. & Agrawa, A. (2014). Information extraction using
discourse analysis from newswires. International Journal of
Information Technology Convergence and Services. 4(3), pp.
21-30.
Rubin, V., Conroy, N. & Chen, Y. (2015)A. Towards News
Verification: Deception Detection Methods for News Discourse.
Hawaii International Conference on System Sciences.
Rubin, V. L. Chen, Y.,& Conroy, N. J. (2015)B. Deception
Detection for News: Three Types of Fakes. In The Proceedings
of the Association for Information Science and Technology
Annual Meeting (ASIST2015), Nov. 6-10, St. Louis.
Rubin, V. & Lukoianova, T. (2014). Truth and deception at the
rhetorical structure level. Journal of the American Society for
Information Science and Technology, 66(5).DOI: 10.1002/asi.
23216 ·
Strehl, A. Ghosh, J. & Mooney, R. (2000). Impact of Similarity
Measures on Web-page Clustering. AAAI Technical Report
WS-00-01.
Wu, G., Greene, D. Smyth, B. & Cunningham P. (2010).
Distortion as a Validation Criterion in the Identification of
Suspicious Reviews. 1st Workshop on Social Media Analytics.
Zhang, H., Fan, Z., Zeng, J. & Liu, Q. (2012). An Improving
De cep ti on De tec ti on Me thod in Co mpute r- Media ted
Communication. Journal of Networks, 7 (11),