IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS, VOL. 11, NO.

4, AUGUST 2024 4975

AI-Assisted Deep NLP-Based Approach for

Prediction of Fake News From Social Media Users
Ganesh Gopal Devarajan , Senior Member, IEEE, Senthil Murugan Nagarajan , Member, IEEE,
Sardar Irfanullah Amanullah , S. A. Sahaaya Arul Mary , and Ali Kashif Bashir , Senior Member, IEEE

Abstract— Social networking websites are now considered to be misleading actions. Traditionally, newspapers are the promi-
the best platforms for the dissemination of news articles. However, nent source to obtain daily and worldwide news. However,
information sharing in social media platforms leads to explosion after the rise of modern development, which made online
of fake news. Traditional detection methods were focusing on
content analysis, while the current researchers examining social platforms to tackle with up to date and instant news spreading
features of the news. In this work, we proposed a novel artificial source that uses various online and social media platforms,
intelligence (AI)-assisted fake news detection with deep natural such as Google, Twitter, YouTube, Line, and Instagram. This
language processing (NLP) model. The proposed work is charac- mode of information sharing with others made possible for
terized in four layers: publisher layer, social media networking researchers and data scientist to collect and analyze for the
layer, enabled edge layer, and cloud layer. In this work, four steps
were carried out: 1) data acquisition; 2) information retrieval use of various insights [2], [3].
(IR); 3) NLP-based data processing and feature extraction; and At early stage, researchers focused mainly the design of
4) deep learning-based classification model that classifies news efficient features for fake news detection from the collected
articles as fake or real using credibility score of publishers, data by various sources, such as news diffusion patterns,
users, messages, headlines, and so on. Three datasets, such as user profiling data, and textual contents [4]. Separation of
Buzzface, FakeNewsNet, and Twitter, were used for evaluation of
the proposed model, and simulation results were computed. This fake news from the real is explored using various methods,
proposed model obtained an average accuracy of 99.72% and an such as sensational headlines, lexical and syntactic analysis,
F1 score of 98.33%, which outperforms other existing methods. writing styles, and linguistic features. Also, temporal features
Index Terms— Artificial intelligence (AI), deep learning, fake and user-based features were proposed by researchers for the
news, information retrieval (IR), natural language processing study of fake news detection. Various measures were taken
(NLP). to deal with fake news by big companies, such as YouTube,
Facebook, Google, and Twitter, to obtain the important insights
I. I NTRODUCTION
from it [5], [6].
I N RECENT years, internet and social media users have
increased rapidly, as the technology has reached to var-
ious extent and remote places. This foreseen development
Recent works have considered learning of word embed-
ding for tackling feature-based problems in text and from
news contents. Various lexical associations have considered
has increased the evolution of communication and informa-
by the word embedding model, as it is generated by the
tion technology. The development of these technologies also
neural language model as a dense real-valued vector. This
made convenience for the internet users to create lot of fake
is the reason that natural language processing (NLP) works
news content [1]. These irrelevant and false information have
commonly with these word embedding that taken as input to
become a major concern for the governments in order to avoid
the neural network models [7]. Also, several researchers have
Manuscript received 29 December 2022; revised shown attention toward the deep neural network (DNN) model
7 February 2023; accepted 15 March 2023. Date of publication 29 March in various domains, such as finance, multimedia, computer
2023; date of current version 2 August 2024. (Corresponding author:
Senthil Murugan Nagarajan.) vision, sentiment analysis (SA), and medical informatics [8].
Ganesh Gopal Devarajan is with the Department of Computer Science and Researchers also have suggested DNN models to be the best
Engineering, SRM Institute of Science and Technology, Delhi NCR Campus, method for analyzing the text data for improving the perfor-
Ghaziabad, Uttar Pradesh 201204, India (e-mail: [email protected]).
Senthil Murugan Nagarajan is with the Department of Mathemat- mance of learning word embedding or machine learning tasks.
ics, Faculty of Science, Technology and Medicine, University of Some deep learning techniques, such as convolutional neural
Luxembourg, Belval Campus, 1855 Esch-Sur-Alzette, Luxembourg (e-mail: network (CNN) and recurrent neural networks (RNNs), were
[email protected]).
Sardar Irfanullah Amanullah is with the Department of Computer Science more popular when comes to text processing [9], [10]. The
and Engineering, Royal Commission for Yanbu Colleges and Institutes, Yanbu main reason for using such techniques is that the local patterns
46452, Saudi Arabia (e-mail: [email protected]). can be learned easily and can also be worked with sequential
S. A. Sahaaya Arul Mary is with the School of Computer Science and
Engineering, Vellore Institute of Technology, Vellore, Tamil Nadu 632014, modeling. Handling of exploding and vanishing gradients is
India (e-mail: [email protected]). more complex part when comes for RNNs particularly when
Ali Kashif Bashir is with the Department of Computing and Mathematics, we have long dependencies as input for the model and is most
Manchester Metropolitan University, M15 6BH Manchester, U.K., also with
the Woxsen School of Business, Woxsen University, Hyderabad 502 345, popular in NLP approaches or domain of SA [11].
India, and also with the Department of Computer Science and Mathemat- Some researchers have developed NLP-based information
ics, Lebanese American University, Beirut 1102 2801, Lebanon (e-mail: extraction or data acquisition system that can populate knowl-
[email protected]).
Digital Object Identifier 10.1109/TCSS.2023.3259480 edge automatically. This has been used as an elicitation tool
2329-924X © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Seshadri Rao Gudlavalleru Engineering College. Downloaded on August 24,2024 at 14:39:12 UTC from IEEE Xplore. Restrictions apply.
 4976 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS, VOL. 11, NO. 4, AUGUST 2024

based on user preference. Several e-business contexts have sentence levels. Authors used graphical recurrent unit (GRU)-
implemented data acquisition or data-controlled personaliza- based sequence encoders in which the gated recurrent unit
tion applications for extracting information. Individual prefer- is stacked with attention models. Wen and Li [21] discussed
ences were stored as a profile based on their services or goods the combination of GRU and CNN with attention mechanism
that created as data acquisition [12]. called the activity recognition and classification (ARC) model
In order to enhance the performance in analyzing the fake for the classification of tweets and reviews. Three various
news detection, we have proposed the NLP-based model that CNN models were employed for extracting bidirectional GRU
uses integrated deep CNN Bi-LSTM (N-DCBL) attention units, global patterns, and local n-gram. Various degrees of
network. Furthermore, we used three datasets mainly, such importance are not determined for backward and forward
as Buzzface, FakeNewsNet, and Twitter, for obtaining the directions using this model.
performance of the proposed work and to predict the fake news Guo et al. [22] proposed the mixed graph neural net-
and social media account users. Simulation of the proposed works (GNNs) model for the detection of fake content from
model is conducted on these datasets, and the performance sustainable vehicular social networks (VSNs). Authors inte-
outcome is compared with other existing methods. The main grated both RNN and convolution neural network with GNN
contributions of this research is summarized below. as bottom architecture for capturing two aspects of seman-
1) We proposed NLP-based-integrated deep CNN tics, which increase in detection effect by this combination.
Bi-LSTM (N-DCBL) attention network for Davoudi et al. [23] proposed stance network and propagation
classification. tree as hybrid model for the detection of fake news. Their
2) We used deep learning-based multiheaded attention net- proposed model consists of three major components, such as
work for checking the credibility of social account users. structural analysis, dynamic analysis, and static analysis. The
3) Four layered architecture is presented for the analysis veracity of news articles was determined by aggregating the
and processing of fake news data. components outputs.
4) We used three datasets, namely, Buzzface, FakeNews- Garg and Sharma [24] focused on different categories
Net, and Twitter, for analyzing the performance of the of linguistic features, such as readability index, stylometric
proposed model. features, complexity features, and psycho-linguistic features
5) Finally, we compared the simulation results of the pro- for the identification of fake news. Different machine learning
posed method with other existing models. algorithms were used in this study and used various datasets,
such as Reuter dataset, in which the model obtained 90.8%,
Buzzface dataset obtained 90% accuracy, and 93.8% for the
II. R ELATED W ORKS
random political dataset. Song et al. [25] introduced fake news
Various supervised machine learning techniques were prediction framework using novel temporal propagation, which
used traditionally for the SA as their main classification or could fuse content semantics, temporal information, and struc-
clustering module. Artificial neural networks (ANNs), Naive ture. Freire et al. [26] proposed a hybrid crowd signals (HCS)
Bayes, and support vector machine (SVM) are among the approach based on crowd signals for considering implicit used
most common techniques that used to analyze the meanings opinions by omitting explicit ones.
of users texts for achieving good results [13], [14]. Two steps
entity level for tweets is proposed by Zhang et al. [15] in III. M ETHODOLOGY
which supervised method is used based on high recall and In this section, we presented the detailed view of the
lexicon-based method for high precision. Mudinas et al. [16] proposed work for detecting and identifying the fake news
proposed a combination model using lexicon-based and and rumors that spread across social media network. Two
machine learning for the analysis of concept-based sentiment main objectives were considered in this proposed model.
tweets. First, detecting fake news article via NLP-integrated deep
Hameed and Garcia-Zapirain [17] presented a study based CNN Bi-LSTM attention (N-DCBL) network to prevent it
on benefited attention mechanism with the combination of from further propagation. Second, verify the credibility of
CNN and Bi-LSTM. In their study, authors did not consider social media user account and publisher via deep learning
co-occurrence of long and short dependencies. Improved word enabled structure-based multiheaded deep attention network.
embedding and CNN are considered by Rezaeinia et al. [18] in The layered architecture of the proposed work is shown in
which different importance values of long and word dependen- Fig. 1. It consists of four layers from which news articles are
cies are not taken into account. Liu and Gao [19] proposed a created by news publishing channel at publication layer. The
hybrid model for question answering and text classification original news article is then applied to propagating network via
using the combination of CNN, attention mechanism, and social media network layer. Social media users take the news
Bi-LSTM (ac-BiLSTM). Authors used 1-D layer in CNN for from publication layer and shared on social media platform
extracting local features, long dependencies were extracted via their social media account. First, the shared news articles
using BiLSTM, and significant text domains were focused were sent to edge layer where we apply NLP-integrated deep
using attention mechanism. CNN Bi-LSTM attention network to detect whether shared
Yang et al. [20] proposed hierarchical attention network news article is fake or real. If news is detected as real, the
(HAN) based on attention model for the text classification. news article is shared on his or her social media timeline;
Two attention models were utilized in HANs for word and otherwise, alert message is sent to the user for not sharing

Authorized licensed use limited to: Seshadri Rao Gudlavalleru Engineering College. Downloaded on August 24,2024 at 14:39:12 UTC from IEEE Xplore. Restrictions apply.
 DEVARAJAN et al.: AI-ASSISTED DEEP NLP-BASED APPROACH FOR PREDICTION OF FAKE NEWS 4977

Fig. 1. Proposed layered architecture using deep learning.

news article. Moreover, the result of detection is fetched to

cloud layer to verify the credibility of user and publisher
via deep learning enabled structure-based multiheaded deep Fig. 2. Creation and propagation of news on social media network.
attention (DL-SMHDA) network. If the user account found to
be suspicious, then it is blocked.
IV. S YSTEM M ODEL
Fig. 3 represents the system model of the proposed
A. Problem Overview work that addresses the issue presented in Section III-A.
Here, first, we check species entities that are involved in the In this system model, we present network of publication
propagation of news article in social media network. Fig. 2 platform ( p1 , p2 , . . . , p P ), user enabled social media network
represents the creation and propagation of news article. There (u 1 , u 2 , . . . , u ( U )), edge computing network, and cloud server
are P number of publishers p1 , p2 , . . . , p P , U number of users network. We performed the computation of fake news detec-
u 1 , u 2 , . . . , u U on social media network, and N number of tion at edge network nearby user device. Edge network consist
news articles a1 , a2 , . . . , a N generated by publisher. Suppose, of nodes e1 , e2 , . . . , e E . At each edge node, the N-DCBL
n i (i = 1, 2, . . . , N ) news article is generated by p j ( j = network model is deployed to detect the fake news. Once
1, 2, . . . , P) publisher and u k (k = 1, 2, . . . , U ) users shared the news detected as fake, then alert message sent to user
n i news created by p j publisher on his timeline. However, and prevent propagation of fake news. Also, result from edge
some u ′k user takes the news from u k user, who force news network is sent to cloud network where further analysis is
article to be shared on u ′k social media timeline, and this forced carried out. Cloud network consists of nodes c1 , c2 , . . . , cC .
news will further propagate in social media network by his At each cloud node, the DL-SMHDA network model is
friends and followers. deployed that verifies the result sent by edge node and also
To better understand the problem statement, we explained checks the credibility of user and publishers. If credibility
the whole process with an example. For simplicity, we use The score of user or publisher is below the threshold, then cloud
Times of India (TOI) as our news publication source, social network sends an alert message to user and notifies him/her
media network platform, and users on social media network. about report rate, which is set as 1. If the report rate is greater
Initially, news article is created by TOI and published on its than 3, then user account will be blocked immediately.
websites. Bob obtains the news article from TOI official web-
site and shared on his social media timeline. From the timeline
of Bob, John takes the story, modifies the news content, and A. Mathematical Model
forged news to be shared on his timeline. Mike connected with Let A = a1 , a2 , . . . , a A be the number of news articles
John via social media network, and he shares the same fake published by different publication platforms ( p1 , p2 , . . . , p P );
news article unknowingly on his timelines. The other user who each news article is labeled as L i ∈ 0, 1 where “0” represents
are connected with Mike and John shared the fake news on fake news and “1” illustrates real news. Each news article
their timeline also. This leads to the propagation of fake news contains headlines (punch line), body (information based on
across the social media network. However, some users who are the headline), source (from where it has been created), and
connected with both John and Bob can access the information so on. When the news article is posted and shared on social
from timeline of both. Also, if they share the news from media, various social media network users (u 1 , u 2 , . . . , u U )
John’s timeline, then fake news is propagated at high speed. respond with different social media contexts (SMCs). The
Propagation of fake news can create mishaps and chaos among SMC includes users’ interaction on social media platforms in
the societies and leads to unfavorable events against nations, the form of likes/dislikes, comments on posts, replies, and fol-
which create mishaps. Therefore, this article proposes to first lowers. For each news article, SMC is defined as SMC(ai ) =
detect fake news article spreading across social media network {(u 1 , smc1 , t1 ), (u 1 , smc1 , t1 ), . . . , (u smc , smcsmc , tsmc )}, where
and prevent it from further spreading using the N-DCBL the tuple (u i , smci , ti ) represents the ith SMC (likes/dislike,
network model and, second, verifies the credibility of user comments of posts, replies, and follower) on the posted news
and publisher using the DL-SMHDA network model. article ai by user u i during time ti . Therefore, the fake news

Authorized licensed use limited to: Seshadri Rao Gudlavalleru Engineering College. Downloaded on August 24,2024 at 14:39:12 UTC from IEEE Xplore. Restrictions apply.
 4978 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS, VOL. 11, NO. 4, AUGUST 2024

Fig. 3. Proposed system model.

Fig. 4. Workflow architecture of NLP-integrated deep CNN Bi-LSTM
attention (N-DCBL) network.
detection task is defined in the following equation:
sharing on social media. Share button contains code of news
Ô(ai ) = N-DCBL(NC(ai ), SMC(ai )) (1)
shared on user’s timeline.
where N-DCBL represents the deep learning model for detect-
ing fake news article, NC(ai ) represents the news content of B. Information Retrieval
the article of ai , and SMC(ai ) represents SMC (likes/dislike, Once the news article is sent to the edge network, then the
comments of posts, replies, and follower) posted news article web crawler retrieves the shared news article information using
ai by a user. social media or www network. First of all, news from the main
web page is extracted and retrieved from the user/place where
V. NLP-I NTEGRATED D EEP CNN B I -LSTM ATTENTION it is posted. Algorithm 1 represents the pseudocode for DT
(N-DCBL) N ETWORK and IR.
The proposed NLP-integrated deep CNN Bi-LSTM atten- Algorithm 1 Data Transfer and Information Retrieval
tion (N-DCBL) network model for fake news detection con- Input: News=a1 , a2 , . . . ., a A , news article website
sists of three modules: data transfer (DT), information retrieval (s1 , s2 , . . . ., s N )
(IR) module, NLP-based preprocessing data module, and
deep learning-based deep CNN Bi-LSTM attention (N-DCBL) 1: procedure DT_IR(shared_news)
network module. A news article is sent to the edge node 2: Begin:
for detecting fake news when the user clicks on the share 3: for each news article i.e., ai ∈ news do
button. The information is retrieved from the internet using the 4: if ai .URL is True() then
IR module; the retrieved model is then processed with NLP 5: News_Article→ ai
preprocessing tasks, such as segmentation, cleaning, parts of 6: Send to N-DCBL at edge node
speech (PoS) tagging, stop word removal, word embedding, 7: Web crawler fetches websites si
and so on. Finally, the N-DCBL module classifies the news 8: retrieve ai from si
article as fake or real. Fig. 4 represents the working archi- 9: Return ai
tecture of the NLP-integrated deep CNN Bi-LSTM attention 10: else
(N-DCBL) network. 11: user event → ai
12: Shared on user’s timeline
A. Data Transfer
News article published by the publication channel is shared
on social media network via social media user. In this proposed C. NLP-Based Data Preprocessing
work, we embed code on the share button that transfers the Social media language is informal and unstructured, includ-
news data to edge network for fake news detection before ing acronyms, grammatical errors, spelling mistakes, and

Authorized licensed use limited to: Seshadri Rao Gudlavalleru Engineering College. Downloaded on August 24,2024 at 14:39:12 UTC from IEEE Xplore. Restrictions apply.
 DEVARAJAN et al.: AI-ASSISTED DEEP NLP-BASED APPROACH FOR PREDICTION OF FAKE NEWS 4979

nonstandard punctuation. Therefore, ant news shared on is expressed as a(t, i) = [a1,i , a2,i , . . . , aT,i ]. Hence, R d
social media network needs extensive processing. For this, represents the dimensional space for every word, and R T ×d
we designed NLP-based data preprocessing module that represents the dimensional space of input news article at,i =
removes all the unnecessary details from the news article. [a1,i , a2,i , . . . , aT,i ]. Finally, we compute the feature
L vector Fv
Following steps are followed for data preprocessing. for the news article, which is concatenated ( ) with word
embedding. The concatenation each word in news article is
D. Data Cleaning calculated using the following equation:
M M M
In data cleaning step, removal of number, special characters, F1:t = w1 w2 ··· wt . (6)
punctuation marks, @username, ♯hashtag, emoji’s, lemmatiza-
tion, and lowercasing is performed. Word vector matrix Mi is obtained by integrating Glove
embedding with TF-IDF using the following equation:

E. Word Segmentation Mi = W ta,t × F1:t . (7)

For word segmentation in news article, we use blank space By integrating TF-IDF with Glove embedding dimen-
and special symbols as delimiters between two words in arti- sionality problem of high-dimensional sparse matrix, pseu-
cle. Thus, each word a j,i in news article ai is mathematically docode for NLP-based data preprocessing (NLP-D) is given
expressed in the following equation: in Algorithm 2.
at,i = a1,i , a2,i , . . . , aT,i t = 1, 2, . . . , T term/word (2)
 
Algorithm 2 NLP-Based Data Preprocessing
where at,i represents the tth term or tth word in news Input: News article = a1 , a2 , . . . ., a N , news article website
article ai . (s1 , s2 , . . . sW )
Output: Weighted word vector matrix Initialize: stopword[]
F. Stop Word Removal = ‘articles’, ‘pronoun’, ‘preposition’, ‘conjunction’, ‘be verb’,
In general, articles, prepositions, conjunction, and pronouns ‘modal verb’
are considered stop words. So, in this step, all the stop words 1: procedure NLP-D ((news article ai ))
are removed from the news article. 2: Begin:
3: for Each ai from i=1 to N do
4: data cleaning
G. Word Embedding
5: at,i = [a1,i , a2,i , . . . , aT,i ]
Word embedding is performed to improve efficiency and 6: t=1,2,. . . ,T term/word
accuracy of deep leaning classifier. To perform word embed- 7: Segmentation
ding, we integrated two techniques: term frequency-inverse 8: if at,i =stopword[i] then
document frequency (TF-IDF) and Glove. 9: Remove stopword from article
10: TF=(♯ of times the term t appeared in news article
H. TF-IDF ai /(♯ o f time ter ms in news ar ticle ai )
TF-IDF is weighting scheme for text mining. Term or 11: I D F = log(N /n f )
word frequency in article is determined using the following 12: W tai ,t = TL F × IL DF L
equations: 13: F1:t = w1 w2 . . . wt
14: Mi = W ta,t × F1:t
♯ of times the term t appeared in news article ai
TF = 15: Output: Weight Word Vector Matrix Mi
♯ of time terms in news article ai
(3)
N
IDF = log (4)
nf J. Deep CNN Bi-LSTM Attention Network
W tai ,t = TF × IDF (5) Deep CNN Bi-LSTM attention consists of three CNN
where N represents the total number of news article in corpse, layers, attention layer, Bi-LSTM layer, dense layer, and
n f represents the frequency of occurrence of term t in N , and concatenation layer. Weighted word vector representation
W tai ,t represents the weight value of word or term in article a. obtained from TF-IDF-integrated Glove word embedding
is fed as input to the convolutional layer based on each
newsLarticle.LLet L Mi:i+ j denote the word concatenation
I. Glove (Mi Mi+1 · · · Mi+ j ). Mi:i+ j represents the local vec-
The Glove is an effective word embedding method that tor matrix from the line ith to line i + jth line of sentence
presents word2vec word representation. Gloved-based word vector matrix. The sentence vector matrix in the news article
embedding techniques generated a 300-D vector by pretraining is represented with a matrix dimension of t × k. To perform
1.2 million treasures, 27 billion tokens, and two billion tweets. convolution operation, we use K ∈ R vk convolution filter,
With Glove-based TF-IDF, we convert each word into a d- which is then applied to the window of v word in t × k
dimensional word vector. Each word a j,i in news article ai sentence vector matrix. After the convolution operation, word

Authorized licensed use limited to: Seshadri Rao Gudlavalleru Engineering College. Downloaded on August 24,2024 at 14:39:12 UTC from IEEE Xplore. Restrictions apply.
 4980 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS, VOL. 11, NO. 4, AUGUST 2024

window (Mi:i+v−1 ) generates xi feature using the following

J
where tanh represents the tangent function and represents
mathematical expression: the hyperbolic tangent. Based on the cell state, output gate
defines what sort of information should be outputted using
xi = f (Mi:i+v−1 × K [i] + Bi ) (8) the following equations:
where f represents the nonlinear activation function-like opτ = sigmoid ωop h τ −1 + u op xτ + Bop


(16)
hyperbolic tangent function or sigmoid function, Bi represents K
the bias, and K represents the weighted kernel filter. Both K h τ = opτ tanh(mcτ ). (17)
and B learned during training period. For every word window The LSTM model captures sequential contextual informa-
M1:v , M2:v+1 , . . . , Mn:n−v+1 , convolution filter K is applied tion but does not capture the information from the previous
that generate feature map using the following equation: sequential LSTM model. Therefore, we employ a Bi-LSTM
x = x1 , x2 , . . . , xn−v+1 where, x ∈ R n−v+1 . (9) network to capture contextual information sequentially from
−−−→
the preceding and succeeding models. Forward ( f wτ ) cap-
The feature map generated from convolutional layer is then tures preceding context-based sequential information, while
−−−→
applied to max-pooling layer. This layer reduces the dimension backward (bwτ ) captures succeeding context-based sequential
of feature map and generates only important feature map using information. The forward and backward LSTM layer output
the following equation: is obtained using the following equations:
Pooli = max(xi ) (10) −−−→ − →
( f wτ ) = L (mcτ −1 , h τ −1 , Asi ) (18)
−−−→ ← −
where Pooli ∈ R n−v+1/2 represents the feature map pro- (bwτ ) = L (mcτ −1 , h τ −1 , Asi ) (19)
duced after the max-pooling operation, and the max-pooling
operation can extract important features from feature maps where mcτ −1 represents the previous state of memory cell
generated via convolution operation. Still, the feature map unit and h τ −1 represents the previous hidden state. Finally,
generated from the max-pooling operation must emphasize we generate the output of Bi-LSTM layer by concatenating
the importance of polarity and semantics. We use an atten- context output information of forward and backward LSTM
tion layer that emphasizes this important characteristic while layer using the following equation:
−−→ ←−−
extracting features to achieve the importance of semantics xBi-LSTM = L f wτ , bwτ . (20)
and polarity in the feature map. The attention layer computes
the attention score for the feature map using the following Generated output provides annotation for each input
equation: weighted word vector. Bi-LSTM network is then fed into
fully connected dense layer that converts bidirectional network
epooli
Asi = P pool (11) output into advanced sentimental representation that helps in
i e
i
predicting sentimental polarity of each feature map. The output
where e represents the exponential function. of fully connected layer is represented using the following
Feature context and long-term dependencies are achieved equation:
by the Bi-LSTM network layer. Attention score-based feature
xi = ReLu(ωi xBi-LSTM + Bi ) (21)
map is then fed as input to Bi-LSTM layer. Bi-LSTM network
←−−− −−−→
consists of forward ( f wτ ) and backward (bwτ ) LSTM layers. where xi represents feature map generated from fully con-
The LSTM network consists of three gates and a memory cell nected layer and xBi-LSTM represents feature map generated
unit. At time τ , i pτ represents the input gate, f gτ represents from Bi-LSTM layer. To predict falsification of news article,
the forget gate, opτ represents the output gate, and mcτ rep- we fed output of fully connected layer into SoftMax layer that
resents the memory cell unit that allow data flow in network. produce the following output:
Forget gate is used to define which information should ignored
Ô i = softmax(xi ωi + Bi ). (22)
from the cell state, and its mathematical expression is defined
using the following equation: Here, sigmoidal classifier is used for classification in Soft-
Max layer
f gτ = sigmoid ω f g h τ −1 + u f g xτ + B f g


(12)
ex j
softmax(.) = sigmoid − →

where xτ represents the input feature map at time τ , h τ −1 is x i = Pc xj
(23)
j=1 e
the hidden state at time τ − 1, and ω f g and u f g represent the
weight for input feature and hidden state, while B f g represents where c represents the number of classes (fake or real). The
the bias for forget gate. After calculating mc ˜ τ and i pτ , what proposed N-DCBL model is evaluated using cross-entropy
must be store in current memory cell unit is defined by input model with the following mathematical expression:
gate i pτ . Mathematical model for input and memory cell unit X
Loss = − Oi log Ô i (24)
are expressed in the following equations:
i

i pτ = sigmoid ωi p h τ −1 + u i p xτ + Bi p


(13) where Oi represents the real output and Ô i represents the
˜ τ = tanh(ωmc h τ −1 + u mc xτ + Bmc )
mc (14) predicted output.
K K Pseudocode for deep CNN Bi-LSTM attention network is
mcτ = f gτ mcτ −1 + i pτ mc
˜ τ (15)
shown in Algorithm 3.
Authorized licensed use limited to: Seshadri Rao Gudlavalleru Engineering College. Downloaded on August 24,2024 at 14:39:12 UTC from IEEE Xplore. Restrictions apply.
 DEVARAJAN et al.: AI-ASSISTED DEEP NLP-BASED APPROACH FOR PREDICTION OF FAKE NEWS 4981

publisher. Credibility is classified into two classes: reliable and

unreliable. CNN and Bi-LSTM layer operation is the same as
in section DCBL network.
Multihead Attention Mechanism: To determine the credibil-
ity of user and publisher, we fed max-pooled feature Pooli
obtained from max-pooling operation for the news article ai
generated by publisher pi and shared by the user u i into
multihead attention layer. Multihead attention is the extension
of attention mechanism that improves the performance. In the
multihead attention technique, first, we compute the attention
score of user and publisher based on the article shared on
social media platform. Attention score A is computed as
SoftMax function for three tuples: Q ∈ R nq ×k that represents
query value, K ∈ R n k ×k that represents key value, and V ∈
R n v ×k that represents value item. Mathematical equation to
p
compute attention score Asi j and publisher p j is as follows:
Fig. 5. Architecture of DCBMHA network. p
Asi j (Q, K , V )
QWhd K T K
 
Algorithm 3 Pseudocode for Deep CNN Bi-LSTM Attention p −1/2 pa

a −1/2


= softmax √ Dg P Dg V (25)
Network k
Input: Weight Word Vector Matrix Mi
Output: News article: real or fake where Q, K , and V represent the linear transformation matrix,
P pa ∈ R |P|×|A| represents the adjacency matrix of pooled
1: Initialize = Learning factor weight ω and bias B pa
feature map pooli whose element P ji represents publisher
2: for each word Mi from i = 1 to n-v+1 do P pa
j publish news article i, (Dg ) j j = ( i P ji and (Dg a )ii =
p
3: feature (xi ) = f (Mi:i+v−1 × K [i] + Bi ) P pa
( i PJ ji ) represent the diagonal matrix for adjacency matrix
4: Pooli = max(xi )
pooli P pa , represents the elementwise matrix multiplication, and
5: Asi = Pe e pooli
−−→ − →i W represents the weight in multiheaded attention layer.
6: f w = L (mcτ −1 , h τ −1 , Asi ) u
Similarly, attention score Asi j for user u j that shares news
←−−τ ← −
7: bwτ = L (mcτ −1 , h τ −1 , Asi ) article ai is defined as follows:
−−→ ←−−
8: X Bi−L ST M = L( f wτ , bwτ ) u
9: xi = ReLu(ωi x Bi−L ST M + Bi ) Asi j (Q, K , V )
i ωi + Bi ) QWhd K T K
 
10: Ô i = sigmoid(x u −1/2 ua

a −1/2


P = softmax √ Dg U Dg (26)
11: Loss = − i Oi log Ô i k
12: Output: News article: real or fake
where Q, K , and V represent the linear transformation matrix,
U ua ∈ R |U |×|A| represents the adjacency matrix of pooled
pa
feature map pooli whose element UP ji represents user j share
VI. D EEP CNN B I -LSTM W ITH M ULTIHEADED news
P article i, and (Dg ) j j = ( i U ua
u
ji ) and (Dg )ii =
a

ATTENTION (DCBMHA) N ETWORK ( i U ji ) represent the diagonal matrix for adjacency matrix
ua

U ua
In this section, we predict publisher credibility and user
a
credibility and verify the fake news prediction result obtained Asi j (Q, K , V )
from the DCBL network. Fig. 5 represents the architecture 
QWhd K T K

t −1/2 ua a −1/2



of the DCBMHA Network. For prediction, we deployed deep = softmax √ Dg U Dg (27)
CNN Bi-LSTM with multiheaded attention (DCB) network k
at each cloud node in a cloud network. In this network, the where Q, K , and V represent the linear transformation matrix,
first news article is preprocessed based on NLP tasks, such Ata ∈ R |T |×|A| represents the adjacency matrix of pooled fea-
as data cleaning, stop word removal, word segmentation, and a
ture map pooli whose element P Ataji representsa word j P in news
word embedding explained in Section V. The preprocessed article i, and (Dg ) j j = ( i A ji ) and (Dg )ii = ( i Ataji )
t

data are then fed into the convolutional layer that produces represent the diagonal matrix for adjacency matrix Ata .
a feature map. Feature map obtained from CNN layer is Equations (22) and (23) are considered as operation of one
then fed into Bi-LSTM layer. The Bi-LSTM layer creates a head, whose expression is given as follows:
contextual-based sequential feature map, which is considered  
p p Q
as input to the multihead attention layer. In DCBMHA, we use hdi j = Asi j QWhd i
, K W K
hd i
, V W V
hd i
(28)
a global average pooling layer instead of a fully connected u u

Q

layer to avoid overfitting and increase speed. The output hdi j = Asi j QWhd i
, K WhdK
i
, V Whd
V
i
(29)
of the global intermediate pooling layer is then fed into
 
a a Q
hdi j = Asi j QWhd , K WhdK
, V Whd
V
(30)
the SoftMax layer to predict the credibility of the user and i i i

Authorized licensed use limited to: Seshadri Rao Gudlavalleru Engineering College. Downloaded on August 24,2024 at 14:39:12 UTC from IEEE Xplore. Restrictions apply.
 4982 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS, VOL. 11, NO. 4, AUGUST 2024

p a u
where hdi j , hdi j , and hdi j represent one head operation for A. Datasets
publisher p j , news article ai , and user u j . Multihead operation For performance evaluation of proposed model, we consider
for publisher p j , news article ai , and user u j is defined using four different datasets form different domains.
the following equations: 1) Buzzface: This dataset is the extension of Buzzface News
p p

M M pj  dataset that consists different features, including images, text,
Mhdi j (Q, K , V ) = hdi j ··· hd A W o (31)
 Facebook SMC, such as comments, and links of news posted
u u u
M M
Mhdi j (Q, K , V ) = hdi j hd A j W o on the Facebook page. This dataset comprised of 2263 pub-


··· (32)
 lished news articles and comment of about 1.6 million [27].
a a
a
M M
Mhdi j (Q, K , V ) = hdi j ··· hd Aj W o . (33) 2) FakeNewsNet: This dataset consists of multidimensional
information from news domains, such as entertainment and
Deep CNN Bi-LSTM-based multiheaded attention network political sources. That dataset is gathered from two well-
Ej
generated feature map xDCBMH i
for user, publisher, and news known fact-checking websites, Gossip-Cop and Political-fact.
article and is represented using the following equation: A total of 783 news items were extracted from Politics-fact,
j E
xDCBMH
E
= Mhdi j (Q, K , V ) (34) among which 336 are fake news, and 447 are real news, while
i
18 417 news are collected from Gossip-Cop, which includes
E p u a
where xDCBMH j
i
j
= {xDCBMH i
, xDCBMH
j
i
, xDCBMH
j
i
}. Equation (21) 1650 fake news and 16 767 true [28].
is used to computed feature map for each entity E j = 3) Twitter: This dataset contains tweets that are related to
{ p j , u j , a j }. 17 different events from different domains, including political,
The output feature map obtained from multihead operation entertainment, sport, and so on. This dataset includes 6225 real
is then fed into fully connected layer that transform it into tweet posted by 5895 Twitter users having 197 images,
final output using the following equation: while 9404 fake tweets posted by 9025 Twitter users having

Ej
 191 images [29].
xiE = ReLu ωiE xDCBMH i
+ Bi (35)

where xi represents the feature map generated from fully con- B. Evaluation Metrics
Ej
nected layer and xDCBMH i
represents the feature map generated To evaluate the performance of the N-DCBL model, we use
from multiheaded attention layer. To predict falsification of five metric measure, including accuracy (AY), recall (RL),
news article, and credibility of user and publisher, we fed precision (PN), F1 measure (FM), and loss. This metric
output of fully connected layer into SoftMax layer that produce measure can be computed using the following equations:
the following output: tp
Precision(pn) = (39)
Ô i = softmax(xi ωi + Bi ). (36) (tp + fp)
tp
Here, sigmoidal classifier is used for classification in SoftMax Recall(RL) = (40)
(tp + fn)
layer
2PN ∗ RL
ex j F1-measure(FM) = (41)
softmax(.) = sigmoid ← x− i = P pn + RL


 (37)
c
e xj tp + tn
j=1 Accuracy(AY) = (42)
tp + tn + fp + fn
where c represents the number of classes (fake or real) to
where tp, tn, fp, and fn represent true positive, true negative,
classify news article. Also, for credibility prediction, c is
false positive, and false negative values, respectively. Loss is
categorized as “reliable” and “unreliable.” The loss for each
computed using (21).
entity is computed using cross-entropy loss. The final loss
outcome of the proposed DCBMH model is obtained using
the following equation: VIII. P ERFORMANCE E VALUATION
We evaluated the performance of the proposed N-DCBL
Loss = Lossu i + Loss pi + Lossai (38) model for three different datasets on five performance metrics.
where Lossu i , Loss pi , and Lossai represent the cross-entropy We also demonstrate a comparative study of the proposed
loss for user credibility prediction, publisher credibil- N-DCBL model with the existing deep learning models (CNN,
ity prediction, and fake news prediction, which are ResNet, and Bi-LSTM). We have divided the dataset into two
computed using (24). parts, the training dataset (70%) and the test dataset (30%).
The model was first trained on the training dataset and then
VII. R ESULT A NALYSIS tested on the testing dataset.
In this section, we represent the performance evaluation of
our proposed N-DCBL and DCBMHA models for fake news A. Comparative Analysis for Buzzface Dataset
detection on a different dataset with different metric measures Fig. 6 shows the performance of the different models for the
under different experimental settings. Also, we represent the Buzzface dataset [27] in the form of training time, accuracy,
comparative analysis of our proposed model with the existing recall, precision, F1 measure, and error loss computed using
model. cross entropy. The confusion matrix of each deep learning

Authorized licensed use limited to: Seshadri Rao Gudlavalleru Engineering College. Downloaded on August 24,2024 at 14:39:12 UTC from IEEE Xplore. Restrictions apply.
 DEVARAJAN et al.: AI-ASSISTED DEEP NLP-BASED APPROACH FOR PREDICTION OF FAKE NEWS 4983

TABLE I
C OMPARATIVE A NALYSIS BASED ON THE B UZZFACE DATASET

Fig. 7. Comparative analysis of error loss based on the Buzzface dataset.

Fig. 6. Comparative results for the Buzzface dataset.

model is shown in Fig. 6, while their associated performance

parameter is demonstrated in Fig. 7 and Table I. The metric
measures for each model are computed from the confusion
matrix. Simulation results were demonstrated in Fig. 6. The
N-DCBL model outperforms other deep learning models. The
proposed N-DCBL gives the highest accuracy (99.39%), pre-
cision (96.96%), recall (95.69%), and F1 measure (98.78%).
To compute the cross-entropy loss, we set the learning rate
and decay rate at 0.001 and 0.9. The hidden state size of
each model is set as 130, 110, 100, and 510 for the proposed
N-DCBL, LSTM, ResNet, and CNN, respectively, to match the
total number of hidden parameters in the hidden layer. Despite
dissimilar size of the neuron layer, each model has the same
hidden parameter. The task of fake news detection requires that
each model will proficiently override the problem of gradient
descent. From Fig. 7, it was seen that the CNN and ResNet are
fixed at baseline, while Bi-LSTM converges to baseline with Fig. 8. Comparative analysis for the fake news detection dataset.
only hundreds of iterations. However, the proposed N-DCBL
converges to baseline efficiently for each training iteration with
minimum loss. of accuracy, precision, recall, and F1 measure. The results
shows that the proposed N-DCBL model outperforms CNN,
ResNet, and Bi-LSTM with the highest accuracy, precision,
B. Comparative Analysis for FakeNewsNet recall, and F1-measure score.
The second dataset is used to test the performance of our For the hidden state size of each model, which is set as
proposed model for fake news detection, which is FakeNews- 130, 110, 100, and 510 for the proposed N-DCBL, LSTM,
Net [28]. The confusion matrix that helps to evaluate the ResNet, and CNN, respectively, the learning rate and decay
performance of the proposed N-DCBL model and the existing rate of 0.001 and 0.9 are set to match with the total number
deep learning model is shown in Fig. 9. The accuracy, preci- of hidden parameters in the hidden layer. Fig. 9 shows the
sion, recall, and F1 measure are computed from true and false results of the Bi-LSTM model, which it cannot converge to
predicted values as indicated in the confusion matrix. Fig. 8 baseline, while ResNet converges to baseline with maximum
and Table II represent the performance of each model in terms error loss. CNN and proposed N-DCBL converge to baseline

Authorized licensed use limited to: Seshadri Rao Gudlavalleru Engineering College. Downloaded on August 24,2024 at 14:39:12 UTC from IEEE Xplore. Restrictions apply.
 4984 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS, VOL. 11, NO. 4, AUGUST 2024

TABLE II
C OMPARATIVE A NALYSIS BASED ON THE FAKE N EWS D ETECTION DATASET

TABLE III
C OMPARATIVE A NALYSIS BASED ON THE T WITTER DATASET

Fig. 11. Error loss comparison for the Twitter dataset.

Fig. 9. Error loss comparison for the fake news detection dataset.
C. Comparative Analysis for Twitter
Finally, we used the Twitter dataset to evaluate the perfor-
mance of the proposed N-DCBL model and compared with the
existing deep learning models. Table III shows the comparative
results based on the Twitter dataset. Based on the simulation
result, Fig. 10 is plotted in which it clearly shows that the
proposed N-DCBL model leads to highest rank among other
deep learning model, as it achieved high value for each metric,
which includes accuracy (99.97%), precision (98.96%), recall
(97.69%), and F1 measure (98.08%).
For each deep learning model N-DCBL, LSTM, ResNet,
and CNN, we set hidden state size of 130, 110, 100, and
510 with a learning rate and a decay rate of 0.001 and
0.9 so as to match with total number of hidden parameters in
hidden layer. Fig. 11 shows that the proposed N-DCBL model
outperforms CNN, Bi-LSTM, and ResNet, as it converges
to baseline with each iteration with minimum loss and high
learning speed.
Fig. 10. Comparative analysis for the Twitter dataset.
IX. C ONCLUSION
Propagation of fake news is constantly surging with
for each iteration, but in comparison with CNN, the proposed a growing social media network. Fake news propagation
N-DCBL model shows minimum error loss. has severe consequences for the country and its citizens.

Authorized licensed use limited to: Seshadri Rao Gudlavalleru Engineering College. Downloaded on August 24,2024 at 14:39:12 UTC from IEEE Xplore. Restrictions apply.
 DEVARAJAN et al.: AI-ASSISTED DEEP NLP-BASED APPROACH FOR PREDICTION OF FAKE NEWS 4985

It deteriorates the reputation of well-known personalities and [10] F. Huang, X. Li, C. Yuan, S. Zhang, J. Zhang, and S. Qiao, “Attention-
degrades the country’s fame. Early detection and prevention of emotion-enhanced convolutional LSTM for sentiment analysis,” IEEE
Trans. Neural Netw. Learn. Syst., vol. 33, no. 9, pp. 4332–4345,
fake news with the highest accuracy are the most challenging Sep. 2022.
tasks that need to work out. We proposed a natural language [11] L. Wang, J. Niu, and S. Yu, “SentiDiff: Combining textual information
processing-integrated deep learning model N-DCBL deployed and sentiment diffusion patterns for Twitter sentiment analysis,” IEEE
Trans. Knowl. Data Eng., vol. 32, no. 10, pp. 2026–2039, Oct. 2020.
at an edge network to address this challenge. The proposed [12] T. Shanahan, T. P. Tran, and E. C. Taylor, “Getting to know you:
N-DCBL consists of the TF-IDF-integrated Glove mechanism Social media personalization as a means of enhancing brand loyalty and
for word embedding while an attention mechanism-based perceived quality,” J. Retailing Consum. Services, vol. 47, pp. 57–65,
Mar. 2019.
CNN-Bi-LSTM model. We claim early detection of fake [13] I. Chaturvedi, Y.-S. Ong, I. W. Tsang, R. E. Welsch, and E. Cambria,
news with our proposed model as before sharing the post “Learning word dependencies in text by means of a deep recurrent belief
on the user timeline, the shared post is first fetched to the network,” Knowl.-Based Syst., vol. 108, pp. 144–154, Sep. 2016.
edge cloud network for fake news detection via the proposed [14] M. E. Basiri and A. Kabiri, “Words are important: Improving sentiment
analysis in the Persian language by lexicon refining,” ACM Trans. Asian
N-DCBL model. Simulation results sent to the cloud network Low-Resource Lang. Inf. Process., vol. 17, no. 4, pp. 1–18, Dec. 2018.
for verification and prediction of the credibility of the [15] L. Zhang, R. Ghosh, M. Dekhil, M. Hsu, and B. Liu, “Combining
user and publisher. If the credibility of user or publisher lexicon-based and learning-based methods for Twitter sentiment anal-
ysis,” HP Lab., Palo Alto, CA, USA, Tech. Rep., HPL-2011-89, 2011.
predicted to be unreliable, then their account from the [16] A. Mudinas, D. Zhang, and M. Levene, “Combining lexicon and learning
social media network is blocked. The performance of the based approaches for concept-level sentiment analysis,” in Proc. 1st
proposed N-DCBL model is evaluated based on accuracy, Int. Workshop Issues Sentiment Discovery Opinion Mining, Aug. 2012,
pp. 1–8.
recall, precision, F1 measure, and loss for three different [17] Z. Hameed and B. Garcia-Zapirain, “Sentiment classification using a
datasets, including Buzzface, FakeNewsNet, and Twitter. The single-layered BiLSTM model,” IEEE Access, vol. 8, pp. 73992–74001,
result showed that the proposed N-DCBL outperformed the 2020.
[18] S. M. Rezaeinia, R. Rahmani, H. Veisi, and A. Ghodsi, “Sentiment
existing deep learning models, including CNN, ResNet, and analysis based on improved pre-trained word embeddings,” Exp. Syst.
Bi-LSTM. This work has some shortcomings that can extend Appl., vol. 117, pp. 139–147, Mar. 2019.
to its future work. In the future, this work can further extend [19] G. Liu and J. Guo, “Bidirectional LSTM with attention mechanism and
by adding some effective feature selection methods over a convolutional layer for text classification,” Neurocomputing, vol. 337,
pp. 325–338, Apr. 2019.
massive feature to avoid overfitting issues. [20] Z. Yang, D. Yang, C. Dyer, X. He, A. Smola, and E. Hovy, “Hierarchical
attention networks for document classification,” in Proc. Conf. North
Amer. Chapter Assoc. Comput. Linguistics, Human Lang. Technol., 2016,
R EFERENCES pp. 1480–1489.
[21] S. Wen and J. Li, “Recurrent convolutional neural network with atten-
[1] K. Shu, S. Dumais, A. H. Awadallah, and H. Liu, “Detecting fake tion for Twitter and yelp sentiment classification,” in Proc. Int. Conf.
news with weak social supervision,” IEEE Intell. Syst., vol. 36, no. 4, Algorithms, Comput. Artif. Intell., Dec. 2018, pp. 1–7.
pp. 96–103, Jul. 2021. [22] Z. Guo, K. Yu, A. Jolfaei, G. Li, F. Ding, and A. Beheshti, “Mixed
[2] M. Kamyab, G. Liu, and M. Adjeisah, “Attention-based CNN and graph neural network-based fake news detection for sustainable vehicular
bi-LSTM model based on TF-IDF and GloVe word embedding for social networks,” IEEE Trans. Intell. Transp. Syst., early access, Jul. 22,
sentiment analysis,” Appl. Sci., vol. 11, no. 23, p. 11255, Nov. 2021. 2022, doi: 10.1109/TITS.2022.3185013.
[3] S. Rastogi and D. Bansal, “A review on fake news detection 3T’s: [23] M. Davoudi, M. R. Moosavi, and M. H. Sadreddini, “DSS: A hybrid
Typology, time of detection, taxonomies,” Int. J. Inf. Secur., vol. 22, deep model for fake news detection using propagation tree and stance
pp. 177–212, Nov. 2022. network,” Exp. Syst. Appl., vol. 198, Jul. 2022, Art. no. 116635.
[4] N. Ebadi, M. Jozani, K.-K.-R. Choo, and P. Rad, “A memory network [24] S. Garg and D. K. Sharma, “Linguistic features based framework for
information retrieval model for identification of news misinformation,” automatic fake news detection,” Comput. Ind. Eng., vol. 172, Oct. 2022,
IEEE Trans. Big Data, vol. 8, no. 5, pp. 1358–1370, Oct. 2022. Art. no. 108432.
[5] H. S. S. Al-deen, Z. Zeng, R. Al-Sabri, and A. Hekmat, “An improved [25] C. Song, K. Shu, and B. Wu, “Temporally evolving graph neural network
model for analyzing textual sentiment based on a deep neural network for fake news detection,” Inf. Process. Manag., vol. 58, no. 6, Nov. 2021,
using multi-head attention mechanism,” Appl. Syst. Innov., vol. 4, no. 4, Art. no. 102712.
p. 85, Oct. 2021. [26] P. M. S. Freire, F. R. M. da Silva, and R. R. Goldschmidt, “Fake news
[6] P. Li, X. Sun, H. Yu, Y. Tian, F. Yao, and G. Xu, “Entity-oriented multi- detection based on explicit and implicit signals of a hybrid crowd:
modal alignment and fusion network for fake news detection,” IEEE An approach inspired in meta-learning,” Exp. Syst. Appl., vol. 183,
Trans. Multimedia, vol. 24, pp. 3455–3468, 2022. Nov. 2021, Art. no. 115414.
[7] P. K. Verma, P. Agrawal, I. Amorim, and R. Prodan, “WELFake: Word [27] G. Santia and J. Williams, “BuzzFace: A news veracity dataset with
embedding over linguistic features for fake news detection,” IEEE Trans. Facebook user commentary and egos,” in Proc. Int. AAAI Conf. Web
Computat. Social Syst., vol. 8, no. 4, pp. 881–893, Aug. 2021. Social Media, Jun. 2018, vol. 12, no. 1, pp. 531–540.
[8] X. Dong, U. Victor, and L. Qian, “Two-path deep semisupervised [28] K. Shu, D. Mahudeswaran, S. Wang, D. Lee, and H. Liu, “FakeNewsNet:
learning for timely fake news detection,” IEEE Trans. Computat. Social A data repository with news content, social context, and spatiotemporal
Syst., vol. 7, no. 6, pp. 1386–1398, Dec. 2020. information for studying fake news on social media,” Big data, vol. 8,
[9] K. Schouten and F. Frasincar, “Survey on aspect-level sentiment anal- no. 3, pp. 171–188, 2020.
ysis,” IEEE Trans. Knowl. Data Eng., vol. 28, no. 3, pp. 813–830, [29] C. Boididou et al., “Verifying multimedia use at MediaEval 2015,”
Oct. 2015. MediaEval, vol. 3, no. 3, p. 7, 2015.

Authorized licensed use limited to: Seshadri Rao Gudlavalleru Engineering College. Downloaded on August 24,2024 at 14:39:12 UTC from IEEE Xplore. Restrictions apply.