'HHS/HDUQLQJ$OJRULWKPVIRU'HWHFWLQJ)DNH

1HZVLQ2QOLQH7H[W

Sherry Girgis Eslam Amer Mahmoud Gadallah

Faculty of Computer Science Faculty of Computer Science Faculty of Computer Science
Modern Academy for Computer Science Misr International University Modern Academy for Computer Science
and Management Technology Cairo, Egypt and Management Technology
Cairo, Egypt [Link]@[Link] Cairo, Egypt
sherrygirgis91@[Link] mgadallah1956@[Link]

Abstract— Spreading of fake news is a social phenomenon at the beginning of events with details and updates at the
that is pervasive at the social level between individuals, and also moment unlike newspapers or magazines in the old days, on
through social media such as Facebook and Twitter. Fake news the other hand, it enables the widespread of 'fake news'
that we are interested in is one of many kinds of deception in because of its accessibility and lack of cost and control of the
social media, but it’s more important one as it is created with
Internet. Recent reports suggest that the outcome of the U.S.
dishonest intention to mislead people. We are concerned about
this issue because we have noticed that this phenomenon has Presidential Elections is due to the rise of online fake news.
recently caused through the means of social communication to Reports indicate that the human ability to detect deception
change the course of society and peoples and also their views, for without special assistance is only 54%. So we need to use
example, during revolutions in some Arab countries have
emerged some false news that led to the absence of truth and stirs
machine learning for classifying texts automatically [2].
up public opinion and also fake of news is one of the factors Fake news detection is considered one of the most
Trump successes in the presidential election. So we decided to dangerous types of deception because it recently caused
face and reduce this phenomenon, which is still the main factor to deceiving many people, Fake news defined as the prediction of
choose most of our decisions. Techniques of fake news detection
varied, ingenious, and often exciting. In this paper our objective
the chances of a particular news article (news report, editorial,
is to build a classifier that can predict whether a piece of news is expose, etc.) being intentionally deceptive (Rubin, Conroy &
fake or not based only its content, thereby approaching the Chen, 2015) [3].
problem from a purely deep learning perspective by RNN
We concerned about the fake news because of the problem
technique models (vanilla, GRU) and LSTMs. We will show the
difference and analysis of results by applying them to the dataset of fake news detection is more challenging than detecting
that we used called LAIR. We found that the results are close, deceptive reviews. A recent report by the Jump-shot Tech
but the GRU is the best of our results that reached (0.217) Blog1found that Facebook referrals accounted for 50% of the
followed by LSTM (0.2166) and finally comes vanilla (0.215). Due total traffic to fake news sites and 20% of the total traffic to
to these results, we will seek to increase accuracy by applying a reputable websites [4]. Since the majority of U.S. Adults –
hybrid model between the GRU and CNN techniques on the same 62%– gets news on social media (Jeffrey and Elisa, 2016) [5]
data set. so that the ability to identify fake content in online sources is,
therefore, an urgent need.
Keywords— Deception detection; Deep Learning; Artificial
Intelligence; RNN (Recurrent Neural Network); LSTM (long Our system establishes for detecting fake news by using
short-term memories) ; Vanilla ; GRU (Gated Recurrent Unit) ; deep learning technique that shows an improvement than
CNN(Convolutional Neural Networks ). linguistic cues recently, we use RNN models (vanilla, and
GRU) and LSTM technique with the LAIR dataset (discuss it
Keywords— Deception detection; Deep Learning; Artificial
Intelligence; RNN (Recurrent Neural Network); LSTM (long
in details in next sections).
short-term memories); Vanilla; GRU (Gated Recurrent Unit); The organization of the paper is as follows: The next
CNN(Convolutional Neural Networks ). section we will talk about the fake news problem in general
and changelings that faced us and the researchers when trying
I. INTRODUCTION to reach the best result, then will presenting the previous
Social media for news consumption, such as Facebook and findings of the researchers in this field to classify news
Twitter is a double-edged sword [1]. On the one hand, low passages as fake or not, next we will mention our experiments
cost and easy access to information and dissemination quickly in details that we reached with the dataset that we use, and it's
to push people to search for news and know what is happening preparation, next we will mention our view of the steps of the
system and proposed our model, Then will clarify our results

978-1-5386-5111-7/18/$31.00 ©2018 IEEE 93

and analysis of each result, Finally will presenting conclusions deception. Used features are token unigrams and linguistic
of all sections. cues derived from the classes of words from the Linguistic
Inquiry and Word Count (LIWC) [10]. This was
II. PROBLEM DEFINITION a psychological experiment with the analysis of the
Fake news becomes a major issue for the public and participants’ writings, focusing on the connection between
government. Fake news can take advantage of multimedia deception and fantasy proneness. The task of detection that
content to mislead readers and get published, which can lead was performing is thus opinion spam detection or fake news
to negative effects or even manipulation of public events. One detection, which is a more variant of deception detection.
of the unique challenges of detecting fake news on social
In the past in deception detection mostly relied on manual
media is how to identify fake news about recent events.
feature selection based on, for example, psycholinguistic
The task of detecting fake news has tested a variety of theories of deception and/or computational linguistics,
labels, from misinformation to rumors; to spam. There has followed by supervised machine learning to build a classifier
been a large body of work surrounding text analysis of fake [11], but recent NLP researches are now increasingly focusing
news and similar topics such as rumors or spam. We have tried on the use of new deep learning methods shown in Fig.1 [12].
to mention some papers that interested in this subject also and
each paper will recall it’s experience and results and is based
on its own concept of these words deceptive [6].
There are two directions to detect text deception. The first of
which is based on separate handheld features, which can
capture linguistic and psychological causes, however, these
features failed to classify text well, which limits performance.
The second approach based on a neural network model learns
document-level representation to discover deception text. Fig. 1. Percentage of deep learning papers in ACL, EMNLP, EACL, NAACL
over the last 6 years (long papers)
Neural network models have been used to learn semantic
representations for NLP tasks and it reaches the highest Here we mention some of the previous works used many
competitive result According to (Le and Mikolov, 2014; Tang techniques for fake news detection.
et al., 2015) [7] and also in NLP, fake data have been collected
by crawling the web or crowdsourcing: fake product reviews (Niall J. Conroy, 2015) [4] Using linguistic cues
(Mukherjee, Venkataraman, Liu, & Glance, 2013) [8]. approaches and network analysis approaches to design a basic
fake news detector which provides high accuracy in terms of
Most researchers used many deep learning popular classification tasks. They propose a hybrid system whose
algorithms such as CNN, Bidirectional-LSTM, and RNN features like multi-layer linguistic processing, the addition of
that we will mention in detail later, but datasets was a reason network behavior, they propose a method to detect online
that why the fake news detection was not successful in the past deceptive test by using a logistic regression classifier which is
because it was small and included unrealistic news, In 2017 based on POS tags extracted from a corpus deceptive and
William created a new benchmark dataset called truthful texts and achieves an accuracy of 72% which could be
LAIR which Collected 12.8K short data is labeled in different further improved by performing cross-corpus analysis of
Contacts from [Link], which Provides detailed classification models and reducing the size of the input feature
analysis report and links to source documents for each case. vector ,it’s one of the best results that these features can
William works on LAIR dataset with many techniques such as reach so as to improve the results the next researchers have to
logistic regression, support vector machines, and use neural networks.
(Bidirectional-LSTM and CNN) models for deep learning and
CNN results is the best [9]. (Samir Bajaj, 2017)[13] used many techniques for neural
networks and machine learning to determine which algorithm
Our work in this paper will implement RNN reach the best result among of them, he apply these algorithms
models (Vanilla RNN, GRU) and LSTMs on LAIR dataset to to his dataset which collect it two different sources; from an
determine if the news is truthful or deceptive and will show open Kaggle dataset sold 13,000 Fake news articles and
the results compared to William's results with analysis. 50,000 authentic news articles (negative examples for the
The following part, we represent some of the previous classifier) were extracted from the Signal Media News dataset.
practical neural network models and their datasets to detect Split all of these data (63000 articles) into 60% training, 20%
deception on a text. dev/validation, and 20% test sets. He used many techniques
from machine learning and neural network such as Logistic
III. RELATED WORK Regression, Feedforward Network, RNN (Vanilla, GRU),
Deception detection (in the framework of computational LSTMs, Bi- LSTMs, CNN with Max-Pooling and CNN with
linguistics) is a text classification problem where our system Max-Pooling and Attention, The results proved that GRU gets
should classify an unseen document as either truthful or the better F1 score and best results overall.
deceptive. Such a system is first trained on known instances of

94
 Other research by (Natali Ruchansky et. Al, 2017) [14], for IV. EXPERMINENTS
getting better result more than previous one; she proposed a
model called CSI, it is built up from deep neural networks A. Dataset
which can extract information from different domains and There are very useful datasets to study fake detection but
capture temporal dependencies in user engagement with the positive training data are collected from a tested out (in a
articles, and also select important features. CSI (which is way that was close to the real thing) (surrounding
composed of three modules: Capture, Score, and Integrate) conditions). More importantly, these datasets are not good for
evades the cost of manual features election by incorporating fake statements detection; since the fake
neural networks. The features they use to capture the temporal news on TVs and social media are much shorter than customer
behavior and textual content in a general way that does not reviews.
depend on the data context nor requires distributional
assumptions. They created this technique to solve three main According to (William Yang [Link], 2017) presents a new
problems in fake text, first of them evaluate the matching benchmark dataset called LIAR: it’s a new, publicly available
score for the headlines and body of an article, second the dataset for fake news detection. Collected a decade-long,
emotion that reach to readers from an article and how make 12.8K manually labeled short statements in various contexts
them feel, the last characteristic is knowing the source of the from [Link], Which provides a detailed
article by checking the structure of the URL, to the credibility analytical report and a link to its source level for each case.
of the media source. They used two datasets from Twitter This dataset can be used for fact-checking research as well.
and Weibo (real-world social media datasets), CSI gives the They investigate the automatic detection of fake news based
best performance overall comparison models and versions. We on surface-level linguistic patterns [12].
see that integrating user features boosts the overall numbers up The LIAR dataset includes 12,836 short statements labeled
to 4.3% from GRU-2. Put together, these results demonstrate for truthfulness, subject, context/venue, speaker, state, party,
that CSI successfully captures and leverages all three and prior history. With this size and time span of ten years,
characteristics of text, response, and source, for accurately LIAR cases are collected in more natural context, such as
classifying fake news (Shown in table 1). political debate, TV ads, Facebook posts, tweets, interview,
Other Works by (William Yang [Link], 2017) [12] Due to news release, etc. In each case, the labeler provides a lengthy
lack of fake news datasets and lack of their efficiency, he analysis report to the ground each judgment [12].
decided to present a new benchmark dataset called LIAR (will They have evaluated several popular learning based
mention it in details in next sections): it’s a new, publicly methods on this dataset. The baselines include logistic
available dataset for fake news detection. Use it with many regression, support vector machines, LSTM and the CNN
techniques such as logistic regression, support vector model. We will present some examples of LIAR dataset in
machines, and (Bidirectional-LSTM and CNN) models for Fig.2 [12].
deep learning. The results proved that CNN
(Convolutional Neural Networks) models are the best. Statement: “Newly Elected Statement: "Under the
Republican Senators Sign health care law, everybody
As we showed that William’s works did not recover using Pledge to Eliminate Food will have lower rates, better
RNN technique, although GRU offered the best results, Stamp Program in 2015.” quality care, and better
therefore we use this technique in LAIR dataset and will Speaker: Facebook posts access."
compare our works with his results and give a brief analysis. Context: social media Speaker: Nancy Pelosi
posting Context: on ’Meet the
TABLE I. COMPARISON OF DETECTION ACCURACY ON TWO DATASETS
Label: Pants on Fire Press’
Twitter WEIBO Justification: More than Label: False
Accuracy F-score Accuracy F-score 115,000 social media users Justification: Even the
DT-RANK
0.624 0.636 0.732 0.726 passed along a story study that Pelosi’s staff cited
0.711 0.702 0.831 0.831 headlined, "Newly Elected as the source of that
DTC Republican Senators Sign statement suggested that
SVM-TS
0.767 0.773 0.857 0.861 Pledge to Eliminate Food some people would pay
0.814 0.808 0.896 0.913 Stamp Program in 2015." more for health insurance.
LSTM-1
But they failed to do due Analysis at the state level
0.835 0.830 0.910 0.914
GRU-2 diligence and were found the same thing. The
CI 0.847 0.846 0.928 0.927 snookered since the story general understanding of the
0.854 0.848 0.939 0.940 came from a publication that word “everybody” is every
CI-T
bills itself (quietly) as a person. The predictions
CSI 0.892 8.894 0.953 0.954
"satirical, parody website." don’t back that up. We rule
We rate the claim Pants on this statement False.
Fire.
Fig.2. depicts some example from LIAR dataset.

95
B. Data Preparation As is common in data mining problems, once the models are
After we have LIAR data set as we mention in the previous built, the process might be repeated with new data and new
section, we must preprocess it to be suitable for our system features.
that can work with. Preprocessing means that data set is
VI. RESULTS AND ANALYSIS
clearer to our algorithm by removing dummy characters,
string, and Impurities. Preprocessing Works on three steps : We constructed three different experiments.
x Splitting: Separate each sentence from the next sentences x Vanilla: Is the first model of RNN that used from 1980,
to deal with them individually. and it’s just Single Layer Network (with feedback).
x Stop words removing: remove UN-important words from x GRU (Gated Recurrent Unit). researchers used to use it
each sentence. (from 2014) because of avoiding vanilla issues which
x Stemming: Returns each word to its origin. filter the information flow to enable the modeling of
long-term dependencies
C. Deception System x LSTM (long short-term memories) Behave like RNNs,
Due to the deception phenomenon that spread nowadays but LSTMs have a different function of computing the
through the traditional media and the social media platform hidden state by introducing input, forget, and output
especially the Facebook and Twitter, which controls the users' gate mechanism and an additional memory cell state
selections in relation to the economic part such as (buying (can store information for a longer time).
products, business, etc…) also in relation to the political life,
In the first experiment we used Vanilla RNN, Then we
so we decided to create s system which solves this problem
used GRU model in the second experiment, finally, in the last
not only fake reviews but we concerned about fake news by
experiment we used the LSTM technique. The following table
using preprocessing LAIR data which has been prepared based
illustrates the accuracy of each model.
on what we have cleared in details in the previous part we
apply it to Word embedding (or word vector) that gives each
word a vector and each vector represents a latent feature of a TABLE II. COMPARISON OF OUR RESULTS AND WILLIAM'S ACCURACY
word , then the result of the word vector apply to RNN models Model Test Accuracy
(Vanilla, GRU) and LSTM , then we will get the results that
determine that a piece of news is deceptive or not. SVMs 0.255

We will explain in details each step of our system in the Logistic Regress0ion 0.247
next section. Bi-LSTMs 0.233

V. PROPOSED MODEL CNN 0.270

Our work clarified in the following steps as follows: Vanilla 0.215

First step: Preparing LIAR dataset in four levels: GRU 0.217
¾ The first level is splitting each sentence to deal with
LSTM 0.2166
separately.
¾ The second level is removing stop words and that We compared our results with (William Yang [Link], 2017)
includes identifying the useless words in each statement in the previous table II.
like (the, a, an, etc). We found the worst result of our experiments gets from
¾ The third level is stemming which every word return to vanilla because of its failure to solve complex tasks that have a
its infinitive. practical application, It also changed the format of the original
Second step: Output of stemming will be the input to word information, which meant that it was unable to hold the
embedding which played an important role in deep learning important memory content for more than a few time steps,
based on deception analysis that includes representing each And also Gradient vanishing Is one of its disadvantages.
LSTM also showed inefficiency compared to GRU and
single word in each sentence by dimensional vector and get
CNN because its two main drawbacks, first it is more
the relation between two words not only syntactic but also the
same (as ‘see’ and ‘watch’ are very different in syntactic, but expensive to calculate the network output and apply back
their meaning is somewhat related) [15]. Another benefit is propagation. we simply have more maths to do because of the
that the algorithm detects the words that appear mostly complex activation. However this is not as important as the
together (like ‘wear’ and ‘clothes’) and it shows their second point, second, the explicit memory adds several more
weights to each node, all of which must be trained. This
relationship and then this is able to predict the next word [16].
increases the dimensionality of the problem and potentially
Third step: Results of word embedding level will be the makes it harder to find an optimal solution.
input to the RNN models (vanilla, GRU ) and LSTMs
technique. The best result of our experiments is GRU, we have seen a
slight improvement in its results than vanilla and LSTM
Fourth step: the output of step four will Getting final because of solving gradient vanishing problem which is a
result determining if the piece of news is truthful or deceptive. problem in vanilla and it is easy to modify and doesn't need

96
memory units, therefore, faster to train than LSTM and give as [3] Conroy, Niall J., Victoria L. Rubin, and Yimin Chen. "Automatic
deception detection: Methods for finding fake news." Proceedings of
per performance. the 78th ASIS&T Annual Meeting: Information Science with Impact:
As we compared our results with William Yang's result we [4] Pérez-Rosas, Verónica, et al. "Automatic Detection of Fake News."
found that CNN is the best among all the results as CNN tend arXiv preprint arXiv:1708.07104 ,2017.
to be much faster (~5 times faster) than RNN and more [5] Allcott, Hunt, and Matthew Gentzkow. "Social media and fake news
in the 2016 election." Journal of Economic Perspectives 31.2 ,2017,
efficient depends on our implementation and because of 211-36.
Nvidia has historically focused much more on CNN than [6] Kumar S, Shah N.” False Information on Web and Social Media: A
RNN, as computer vision mostly employs CNN. Survey”. arXiv preprint arXiv: 1804.08559, 2018 Apr 23.
[7] Lopez MM, Kalita J.” Deep Learning applied to NLP”. arXiv preprint
VII. CONCLUSION arXiv:1703.03091,2017 Mar 9.
In recent years, deception detection in online reviews & [8] Rubin, Victoria L., Yimin Chen, and Niall J. Conroy. "Deception
detection for news: three types of fakes." Proceedings of the 78th
fake news has an important role in business, law enforcement, ASIS&T Annual Meeting: Information Science with Impact:
national security, political due to the potential impact fake Research in and for the Community. American Society for
reviews can have on consumer behavior and purchasing Information Science, 2015.
decisions. Researchers used deep learning with the large [9] Wang WY.” " liar, liar pants on fire": A new benchmark dataset for
dataset to increase in learning and thus get the best results by fake news detection”. arXiv preprint arXiv:1705.00648. 2017, May 1
using word embedding for extract features or cues that [10] Verhoeven, Ben, and Walter Daelemans. "CLiPS Stylometry
Investigation (CSI) corpus: a Dutch corpus for the detection of age,
distinguish relations between words in syntactic and semantic. gender, personality, sentiment and deception in text." In LREC 2014-
In this paper we cover implementation of RNN technique NINTH INTERNATIONAL CONFERENCE ON LANGUAGE
models (Vanilla, GRU ) and LSTMs that have been proposed RESOURCES AND EVALUATION, pp. 3081-3085,2014.
for the detection of online fake news after we prepare our [11] Hagiwara, Masato. "A supervised learning approach to automatic
synonym identification based on distributional features." In
LAIR dataset applying to prepare data to word embedding to Proceedings of the 46th Annual Meeting of the Association for
get vectors of words then entering this vectors to our deep Computational Linguistics on Human Language Technologies:
learning technique, we found that the results of our Student Research Workshop, pp. 1-6. Association for Computational
experiments are close but GRU(Gated Recurrent Unit) is the Linguistics, 2008
best because it's solving the problems of Vanilla that popular [12] K Young T, Hazarika D, Poria S, Cambria E. “Recent trends in deep
learning based natural language processing.” arXiv preprint
of gradient vanishing problem and LSTMs (long short-term arXiv:1708.02709. 2017 Aug 9.
memories) which GRU is easy to modify and doesn't need [13] Bajaj S. “The Pope Has a New Baby! Fake News Detection Using
memory units , so, we gain faster training than LSTM that Deep Learning”.2017 .
effect in our performance ,but for comparing our results with [14] Ruchansky N, Seo S, Liu Y. Csi.” A hybrid deep model for fake news
William's results we found that CNN’s(Convolutional Neural detection”. InProceedings of the 2017 ACM on Conference on
Information and Knowledge Management ,2017 .Nov 6. pp. 797-806,
Networks ) is the best from other models due to its speed and ACM.
its best results and performance for windows .our future [15] Lepping, Joachim. "Wiley Interdisciplinary Reviews: Data Mining
works will increase this accuracy by merging GRU and and Knowledge Discovery.",2018.
CNN’s to get the best result . [16] Mikolov T, Sutskever I, Chen K, Corrado GS, Dean J. “Distributed
representations of words and phrases and their compositionality.
REFERENCES InAdvances in neural information processing systems”, 2013 ,pp.
3111-3119.
[1] Erlich, Aaron, et al. "The double-edged sword of mobilizing citizens
via mobile phone in developing countries." Development Engineering [17] [Link].
3, 2018, 34-46. [18] [Link] .com
[2] Krishnamurthy, Gangeshwar, et al. "A deep learning approach for
multimodal deception detection." arXiv preprint arXiv: 1803.00344,
2018.

97