This article has been accepted for publication in IEEE Access.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3457860

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2024.DOI

RoBERTa-GCN: A Novel Approach for

Combating Fake News in Bangla Using
Advanced Language Processing and
Graph Convolutional Networks
MEJBAH AHAMMAD1§ , AL SANI2§ , KHALILUR RAHMAN3§ , MD TANVIR ISLAM4§ , MD
MOSTAFIZUR RAHMAN MASUD5§ , MD. MEHEDI HASSAN6 , (Member, IEEE), MOHAMMAD
ABU TAREQ RONY7 , SHAH MD NAZMUL ALAM8 , MD. SADDAM HOSSAIN MUKTA 9*
1
Department of Computer Science and Engineering, American International University Bangladesh, Dhaka, Bangladesh
2
Department of Industrial and Production Engineering, Jashore University of Science and Technology, Jashore, Bangladesh
3
Department of Computer Science and Engineering, Sylhet Engineering College, Sylhet, Bangladesh
4
Department of Software, Sungkyunkwan University, Gyeonggi-do, South Korea
5
School of Computing, Universiti Teknologi Malaysia, Johor Bahru, Malaysia
6
Computer Science and Engineering Discipline, Khulna University, Khulna, Bangladesh
7
Department of Statistics, Noakhali Science and Technology University, Noakhali, Bangladesh
8
Department of Computer Science and Engineering, Arizona State University
9
LUT School of Engineering Science, LUT University, Lappeenranta, Finland
Corresponding author: Md. Saddam Hossain Mukta ([email protected])
Authors with § symbol have equal contributions.

ABSTRACT In this era of widespread information, combating fake news in less commonly represented
languages like Bengali is a significant challenge. Fake news is a critical issue in Bangla, a language that
a vast population uses but lacks adequate natural language processing tools. To address this, our research
introduces RoBERTa-GCN, a cutting-edge model combining RoBERTa with a graph neural network (GCN)
to accurately identify fake news in Bangladesh. The dataset we utilized comprises articles from 22 prominent
Bangladeshi news portals covering diverse subjects such as politics, sports, economy, and entertainment.
This comprehensive dataset enables the model to learn and adapt to the intricacies of the Bangla language
and its news ecosystem, facilitating effective fake news detection across various content categories. Our
approach integrates the RoBERTa model, adapted for Bangla, with GCN’s expertise in processing relational
data, forming an effective means to differentiate between authentic and fake news. This study’s key
achievement is the creation and application of the RoBERTa-GCN model to the Bangla language, an area
not thoroughly explored in previous research. The findings show that RoBERTa-GCN surpasses existing
methods, achieving impressive accuracy rates of 98.60%, highlighting its capability as a robust model for
preserving news integrity in the digital era, especially for the Bangla-speaking population.

INDEX TERMS Fake news detection, Graph neural network, NLP, Bangla language, Deep learning,
Machine learning, Encoder

I. INTRODUCTION formation, including fake news, within the digital realm [2].

Recent years, the emergence of online news plat- “Fake news” refers to articles that might delude or deceive
IN forms, including social media, news blogs, and on-
line newspapers, has prompted individuals to actively seek
readers by disseminating fabricated information [3]. For the
sake of traffic, some self-media and internet users spread a
and consume news due to their advantages, such as rapid in- significant amount of unverified news that was subsequently
formation dissemination, convenient accessibility, and cost- reprinted and unthinkingly followed to expand and under-
effectiveness [1]. Meanwhile, easy access to social media mine the credibility and authority of mainstream media.
platforms enables the rapid and widespread spread of false in- However, it additionally resulted in economic, political, and

VOLUME 4, 2016 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3457860

M. Ahammad et al.: RoBERTa-GCN: A Novel Approach for Combating Fake News in Bangla Using Advanced Language Processing

other hidden risks [4], [5]. Approximately 265 million people speak Bangla as their na-
The propagation of false news serves to create fear, pro- tive language, making it the 7th most widely spoken language
mote racist ideology, and provoke acts of bullying and vio- globally [23]. The majority of the readers are unable to recog-
lence against innocent individuals [6]. According to intelli- nize fake news with their own knowledge. There are several
gence studies, social media rumors have a particularly long- studies available in English. As far as we know, limited
lasting influence on less clever people, preventing them from resources or computational methods are currently available to
making optimal judgments [7]. address the menace of fake news generated in Bangla, which
Fake news is widely recognized as a significant threat to poses a potential threat to this sizable community. Therefore,
global trade, media, and democracy, carrying severe ramifi- to address these challenges, we propose a custom method of
cations [8]. Around 25% of Americans visited a fraudulent detecting Bangla fake news utilizing the power of RoBERTa
news website six weeks before the 2016 US election. This and GCN. In addition, our research through extensive data
occurrence has been hypothesized as a potential factor that analysis and exploring various methods for fake news detec-
impacted the election’s outcome [9]. A fake news report tion marked by the following key contributions:
claimed the explosion resulted in injuries to US President 1) We propose a custom model named RoBERTa-GCN
Barack Obama, leading to a significant loss of $130 billion for Bangla fake news detection. To the best of our
in the stock market [10]. knowledge, we are the first to integrate RoBERTa and
In addition to having a higher rate of social media penetra- GCN for Bangla fake news detection.
tion, Bangladesh is a South Asian nation plagued by poverty 2) The combination of RoBERTa’s contextual embed-
and rumors. As of the beginning of 2023, Bangladesh’s dings with GCN’s structural learning is an innovative
internet user population stood at 66.94 million, as reported approach that enhances Bangla fake news detection by
by Datareportal [11], [12]. Bangladesh has encountered many leveraging the relationships within news content, an
quantifiable incidences of misinformation in recent years. aspect often overlooked by other models.
Ten individuals were injured, and rioters fatally beat five in 3) We benchmark several SOTA models against Ban-
July 2019 due to widespread rumors regarding the antici- FakeNews and compare them with our proposed
pated human sacrifice during the Padma Bridge construction RoBERTa-GCN, which outperforms all existing mod-
project [13]. An additional case of misinformation resulting els by achieving 98.60% accuracy.
in violence occurred in 2012 when a local Buddhist youth’s The structure of the subsequent sections of this document
Facebook post featuring an image of a burned Quran, the is meticulously designed to facilitate a coherent presentation
holy book of Islam, incited an enraged mob to torch 12 of our research findings. Section II is dedicated to a crit-
Buddhist temples, pillage, and burn over 50 houses in a ical examination of pertinent literature, situating our work
Buddhist community situated in Ramu, Cox’s Bazar [14]. within the broader scholarly context. Section III, we explain
During the COVID-19 pandemic, vaccination skepticism in overall data collection and analysis procedures. In Section
Bangladesh was encouraged by allegations that coronavirus IV, we delineate our proposed methodology, elaborating on
vaccines "contain a microchip" that makes it possible for the innovative approaches and techniques employed in our
Western governments to spy on people. Such unjustified investigation. Section V presents a rigorous analysis of the
conspiracy concepts pose a huge obstacle to the country’s data obtained through the application of advanced method-
public health programs [15]. Another false news stated that ological procedures, offering a comparative evaluation of
just 200,000 vaccinations would be imported for the first the results. Finally, Section VI synthesizes the key insights
installment, with all of this being delivered to the government and discoveries emanating from our study, underscoring their
leaders and members of the ruling party in Bangladesh [16]. significance and implications for the field.
To counter the spread of misinformation, dedicated plat-
forms like PolitiFact [17], FactCheck [18], and JaChai [19] II. RELATED WORKS
manually review and update potential fake news articles The dissemination of false information on online media
in online media, providing detailed explanations and citing platforms has led to a need for clarification among various
logical and factual reasons to debunk the false claims. De- individuals, giving rise to numerous challenges. Several re-
spite their efforts, computational tools have recently been searchers have employed diverse techniques to control the
employed to fight the threat of false news [3]. One effort to spread of fake news.
identify fake news involves using multi-perspective speaker Rai et al. [35] used Long Short-Term Memory (LSTM)
profiles [20]. Conversely, another approach introduces an and bidirectional encoder representation from transformers
automated fact-checking methodology leveraging third-party (BERT) to classify fake news. Their results showed higher
sources to verify the precision of news articles [21]. Ad- accuracy compared to the baseline models on the PolitiFact
ditionally, a methodology has been established to detect and GossipCop datasets. Additionally, Rai et al. employed a
incongruities between Bangla news headlines and body con- vanilla BERT model combined with an LSTM layer to in-
tent [22]. vestigate performance improvements further. Nasir et al. [24]
Distinguishing fake news in Bangla is significantly more showed how to use a hybrid deep learning (DL) model
challenging than in other languages because of its structure. that combines convolutional and recurrent neural networks
2 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3457860

M. Ahammad et al.: RoBERTa-GCN: A Novel Approach for Combating Fake News in Bangla Using Advanced Language Processing

TABLE 1. Summarizing various methods in the literature for detecting fake news, detailing the datasets used, implemented techniques, and their overall
performances.

Ref. Year Dataset Technique Accuracy (%)

2021 FA-KES(804 news articles) Hybrid CNN-RNN 60.00
[24]
2021 300 articles Random Forest 82.00
[25]
2020 726 news articles Random Forest 85.00
[26]
2023 English articles, English tweets, Urdu articles, Urdu BERT 90.00
[27] tweets, Arabic Articles, Arabic Tweets
2020 22 most popular and mainstream trusted news portals Support Vector Machine 91.00
[3] (50,000 news articles)
2023 English ISOT(44,493 news articles), Dravidian-language XLM-RoBERT 93.31
[28] fake news(25,347 news articles)
2022 Facebook (676 individual posts) DistilBERT, XLM- 97.00
[29] RoBERTa, BERT
2023 IBFND BERT 97.22
[30]
2021 George McIntire(6,335 news articles), Kaggle(20,761 BerConvoNet 97.45
[31] news articles), Gossipcop(21,641 news articles), Politi-
fact(948 news articles)
2022 Bengali newsletter-based research database (49000 news Support Vector Machine 98.08
[32] articles gathered from various Bangaledesi news sources
2023 Twitter and YouTube (10,700 posts and news articles) CT-BERT+BiGRU 98.46
[33]
2022 Buzzfeed(1700 news), GossipCop(17,520 news), LSTM-LF 98.50
[34] ISOT(44,900 news), Politifact(700 news)

(CNN-RNN) to sort fake news. This model did better than datasets for social media.
non-hybrid methods on the ISOT and FA-KES datasets and
To notice fake news, Xia et al. [39] utilized a multi-head
shows promise for future tests that will use it on other
attention model to find major emergencies in a think tank
datasets as well. Alghamdi et al. [33] explored the impact
and public opinion stage and a hybrid CNN, Bi-LSTM, and
of freezing and unfreezing parameters in a neural network
attention mechanism model, which helped to increase metrics
architecture used for BERT and CT-BERT. They found that
values like loss, accuracy, f1-score, and recall. To distin-
models like BiGRU and CT-BERT achieved superior perfor-
guish fake news from SOTA models by adding variational
mance on the COVID-19 fake news dataset. Sudhakar and
Bi-LSTM, autoencoder, and semantic topic-related features,
Kaliyamurthie [36] employed advanced machine learning
Hosseini et al. [40] used LDAVAE, an integrated LDA,
classifiers, specifically Logistic Regression(LR) and Naive
supervised Bi-LSTM VAE probabilistic model where the
Bayes(NB), where the G-power experiment ensures accuracy
model ablation measured the accuracy and performance of
and T-test analysis displays that LR works better than NB
neural embeddings and topic features. Palani and Elango [41]
for independent samples. Choudhary and Arora [37] applied
proposed a hybrid BERT-BiLSTM-CNN (BBC-FND) model
a sequential neural model incorporating an LSTM-based
structure consisting of three primary layers: an embedding
word embedding model. The feature-based sequential model
layer, a feature representation layer, and a classification
performed better than ML-based models and LSTM-based
layer that then captures applicable information, patterns,
word embeddings in less time. Choudhary et al. [31] showed
and global meaning to detect the forecast news validity.
that BerConvoNet, using BERT in the news embedding block
Okunoye and Ibor [42] used genetic search to select neural
with various kernel sizes, effectively distinguishes real from
architecture, utilizing neural architecture and DL techniques
fake news across benchmark datasets, outperforming other
to classify bogus news. Malhotra and Malik [43] evaluated
state-of-the-art (SOTA) models. Khullar and Singh [38] em-
SVM, LR, CatBoost, XGBoost, multinomial, NB, and RF
phasized the need for a distributed client-server architecture
ensemble models, where the deep auto-ViML model and
to protect data, proposing a model named F-FNC, utilizing
passive-aggressive classifier worked well. SVM was the
LSTM, CNN-LSTM, Bi-LSTM, and CNN-BiLSTM algo-
fastest at 0.245 ms. Ghamdi et al. [27] utilized natural lan-
rithms, the model outperformed other distributed DL clas-
guage processing (NLP), BERT, and human programming
sifiers based on different metrics. Raja et al. (raja2023fake)
to detect misinformation in Twitter and website content. A
used transfer learning with mBERT and XLM-R models to
labeled dataset in English, Arabic, and Urdu was used to
effectively detect fake news in the Dravidian Fake dataset,
highlight the efficiency of their models in detecting false
achieving high accuracy in multilingual settings. Similarly,
material. Mohawesh et al. [44] proposed a semantic tech-
Kaliyar et al. (kaliyar2020fndnet) developed FNDNet, a
nique combining pre-trained word2vec vectors with a capsule
CNN-based model that identifies biased features to classify
neural network to enhance multilingual fake news detection.
fake news, showing improved performance on benchmark
This approach surpassed existing methods by incorporating
VOLUME 4, 2016 3

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3457860

M. Ahammad et al.: RoBERTa-GCN: A Novel Approach for Combating Fake News in Bangla Using Advanced Language Processing

relational variables extracted from the text. Dixit et al. [34] III. DATASET: BANFAKENEWS
showed the four-step false news detection models: data pre- In this paper, we utilized a dataset named "BanFakeNews",
processing, feature reduction, feature extraction, and LSTM- which was specially introduced for developing models for
LF classification. It worked better than previous ways of classifying fake and real news. "BanFakeNews" was collected
finding fake news on Buzzfeed, GossipCop, ISOT, and Poli- from 22 most prominent online news sites in Bangladesh,
tifact datasets. Jain et al. [45] utilized a resilient DL model which was used to acquire false news data for our research.
to detect false information assertions using associated em- Ensuring the dataset’s diversity involved selecting news sto-
bedding, attention methods, and pertinent metadata to show ries from various fields, such as sports, politics, economics,
that it worked effectively across real datasets. In contrast and the environment. Here is a synopsis of all the news stories
to CNN, Palani et al. [46] used the BERT model to pull that were part of our data collection:
out written parts that kept their meaning and came up with
the Capsule network (CapsNet) algorithm, which helped • kalerkantho.com • bd24live.com
combine useful data to make a lot of data representations • jagonews24.com • risingbd.com
for accurate fake news detection. Ravish et al. [47] suggested • banglanews24.com • dailyjanakantha.com
using a multi-layered Principal Component Analysis (PCA) • banglatribune.com • bd-pratidin.com
feature selection method along with Multiclass Support Vec- • jugantor.com • channelionline.com
tor Machines (MSVMs) for classification. This improved • dhakatimes24.com • samakal.com
model accuracy across ten datasets, especially when more • ittefaq.com.bd • independent24.com
characteristics were needed to validate the feature extrac- • somoynews.tv • rtnn.net
tion methods. Hossain et al. [3] studied and tested cutting- • dailynayadiganta.com • bangla.thereport24.com
edge models such as BERT, CNN, Bi-LSTM, LR, and RF. • bangla.bdnews24.com • mzamin.com
These models used common language variables and neural • prothomalo.com • bhorerkagoj.net
networks to make low-resource language research more ad-
vanced. To detect fake news classification, Hasib et al. [48] Table 2 presents the framework of a dataset specifically
used DT, SGD, BERT, CNN, and ANN models, where BERT created for identifying fake news in the Bengali language.
worked better than any other models and achieved maximum The dataset consists of columns representing the article ID,
accuracy in balanced datasets and good performance through domain, publication date, category, title, text, and a label
cross-validation with varied K values. Pranto et al. [29] indicating the authenticity of the piece. Every column has
experimented with ML models, among which BERT was a distinct function for analysis. The bar chart in Figure 1
the most prominent automatic detection method for Bengali represents a dataset breakdown for noticing fake news in
false news classification. When BERT was applied to Bengali Bangla. Each bar corresponds to a different news domain,
Facebook posts related to the COVID-19 dataset, 10 topics and the length of the bar reflects the number of articles from
were identified and grouped into three groups: system, belief, that domain included in the dataset. Here’s an analysis based
and social. Roman et al. [30] enhanced the Bangla Fake News on the visualization:
Dataset by addressing the imbalance, using web scraping TABLE 2. Features of the fake news dataset used in this study.
and Google Translate, achieving the highest F1-score via
Columns Descriptions
the BERT model, surpassing the BanFakeNews dataset’s articleID Unique identifier for each article
performance. Ali et al. [32] conducted a feasibility study on domain The domain in which the publication was made
detecting Bangla false news from social media, employing date Date of publication of the article
diverse feature extraction methods and ML algorithms, ulti- category Category or genre of the article
headline Headline or title of the article
mately achieving optimal accuracy using LSTM in their pro- content Main content or body of the article
posed system. Anjum et al. [25] used a Random Forest(RF) label Label indicating the authenticity of the article
classifier to distinguish bogus news from a combination of
false and legitimate sources, with an accuracy of 82%. It
included more than 300 articles. Islam et al. [26] employed a Domain Diversity:The dataset comprises articles from di-
data mining approach to classify Bengali fake news in South verse domains. This diversity is beneficial for building a
Asia, analyzing 726 articles, and found RF to archive an robust fake news detection mechanism, allowing the model
accuracy rate of 85%. Bhattacharjee et al. [49] introduced to learn from various writing styles and content types.
BanglaBERT and BanglishBERT and achieved results in Volume of Articles: The domains contribute a varying num-
NLU for Bangla, along with new datasets. ber of articles, with kalerkantho.com and jagonews24.com
providing the largest amount. In contrast, bhorerkagoj.net
As highlighted in the literature review and Table 1, numer- and mzamin.com contribute the least.
ous models have shown potential for fake news detection, but Data Representation: The dataset reflects the diversity
most are tailored for English. This paper addresses this gap of domains and the proportionality of fake and real news
by proposing a specialized model for detecting fake news in within each domain to prevent overfitting to specific domain
the Bangla language. characteristics that are not related to news authenticity.
4 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3457860

M. Ahammad et al.: RoBERTa-GCN: A Novel Approach for Combating Fake News in Bangla Using Advanced Language Processing

FIGURE 1. Distribution of Bangla fake news by domain, showing the count

of fake news articles from various news websites.

FIGURE 2. Category Wise distribution of News in Donut Chart for Bangla

The utilization of the chart guarantees that the dataset em- fake news detection.
ployed to train a model for detecting false news is equitably
distributed and accurately reflects the actual dissemination of
news across various Bangla domains. Furthermore, Figure 2 pendable identification and filtration of misinformation. In
illustrates the category-wise distribution of the dataset, show- this paper, we used the BanFakeNews [3] dataset, which we
ing an even spread across five categories: politics, sports, preprocess as shown in Figure 5 and further described as:
economy, entertainment, and technology. Each category rep-
resents roughly one-fifth of the total, with entertainment 1) Text Cleaning
slightly leading. This balanced distribution suggests that fake The Bangla Fake News Text Cleaning Process is a meticulous
news is not concentrated in a single category but is spread and crucial step in data preparation, especially tailored for
across various topics. Such a balance in the dataset is valuable the unique characteristics of the Bengali language. This pro-
for developing a fake news detection model that remains cedure involves several key activities: removing non-textual
unbiased and effective across multiple categories. Similarly, elements, correcting typographical errors, and standardizing
Figure 3 presents a polar bar chart highlighting the frequency various dialectal forms prevalent in the Bangla script. Distin-
of top words associated with fake news in Bengali. The chart guishing characters, punctuation, and numerical information
features several spikes, each representing the frequency of a that fail to enhance the semantic value of the text is a
different word. Notably, one word stands out with a signifi- particular emphasis. The process also includes normalizing
cantly higher frequency, indicating its common usage in the colloquial expressions and idioms, ensuring the text is ana-
context of fake news. This visual tool helps in identifying key lytically relevant and linguistically accurate. This thorough
terms prevalent in fake news articles, which can be critical for cleaning is essential for effectively analyzing and detecting
enhancing the performance of machine learning algorithms fake news in Bangla text datasets.
in text classification. Figure 4 provides a comparative visual-
ization of word clouds of the dataset. These clouds showcase • Text Normalization: Converting text to a uniform for-
the most frequently used words within each news type, with mat by lowercasing characters, standardizing dates and
the size of the words reflecting their frequency. numbers, and resolving inconsistencies.
• Noise Removal: Eliminating irrelevant characters and
symbols like HTML tags, URLs, emojis, special char-
A. DATA PREPROCESSING acters, and extra whitespaces.
The processing of Bangla fake news data involves specialized • Dealing with Punctuation: Removing or utilizing punc-
steps tailored to the nuances of the Bengali language. It tuation marks to improve sentence structure parsing.
includes text normalization, the removal of irrelevant charac- • Handling Numerals: Convert numbers to words or re-
ters, and contextual analysis. Additionally, algorithms adapt move them based on their analysis and relevance.
to linguistic patterns for accurate classification. Significantly • Tokenization: Breaking the text into tokens such as
contributing to preserving information integrity within the words, numbers, or symbols, considering the unique
Bengali-speaking community, this procedure guarantees de- script and grammar of the Bangla language.
VOLUME 4, 2016 5

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3457860

M. Ahammad et al.: RoBERTa-GCN: A Novel Approach for Combating Fake News in Bangla Using Advanced Language Processing

FIGURE 3. Polar Bar chart visualizing frequency of top words used in FIGURE 4. Word cloud of the dataset we used for RoBERTa-GCN.
Bangla language fake news articles

recognition to pinpoint and categorize names, places, and

• Removing Stop Words: Discard frequently occurring
organizations. These steps are tailored to comprehend the
words that have little meaning, such as prepositions,
complexities of Bangla syntax and semantics, allowing for
conjunctions, and certain verbs.
an in-depth examination of news content to identify and flag
• Stemming and Lemmatization: Reducing words to their
misinformation accurately.
root form through stemming or using vocabulary and
morphological analysis for lemmatization, which is
complex due to Bangla’s morphology. 4) Content Authentication
• Handling Slang and Abbreviations: Standardize or In the fight against fake news within the Bangla-speaking
translate slang terms and abbreviations to maintain clar- community, content authentication stands as a pivotal line
ity in the dataset. of defense. This intricate process employs sophisticated
fact-checking algorithms tailored to the Bangla language,
2) Feature Extraction scrutinizing the veracity of claims and narratives. Source
The cleaning procedure is crucial for producing Bangla credibility analysis evaluates the trustworthiness of informa-
text data to identify false news. The procedure involves tion origins, while cross-referencing mechanisms compare
making essential modifications to the text to enhance the data across multiple databases and records. These measures,
efficiency of feature extraction methods, including Bag of essential in the digital age, are designed to safeguard the
Words (BOW) [50], TF-IDF [51], Word Embedding [52], and truth and combat the spread of misinformation in Bangla
N-Grams [53]. This stage rigorously removes discrepancies, news, thereby preserving the integrity of the informational
standardizes textual deviations, and eradicates non-linguistic ecosystem for millions of Bengali speakers.
components unique to the Bengali script. The thoroughness
of the cleaning process is crucial, as it directly impacts the 5) Syntactic Analysis
quality of the recovered features. Consequently, the accuracy Syntactic analysis in Bangali fake news detection is a struc-
of recognizing fake news within the linguistically varied tured approach to dissecting the grammatical intricacies of
Bangla dataset is affected. text. It involves constructing parsing tree structures that map
the grammatical hierarchy and relationships between words.
3) Semantic Analysis Grammar analysis examines the rules specific to Bangla syn-
Semantic analysis encompasses a multifaceted procedure tax, ensuring sentences are constructed correctly. Sentence
that interprets the latent meanings, sentiments, and veracity structure analysis then delves deeper, evaluating the arrange-
present in Bengali text to identify false news. This phase ment of words and phrases to discern patterns that might
involves sentiment analysis to gauge the emotional tone, indicate misinformation. This rigorous syntactic scrutiny is
theme identification to understand the central subjects, con- vital for automated systems to effectively identify and flag
textual analysis to grasp situational meanings, and entity fake news in the richly complex Bangla language.
6 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3457860

M. Ahammad et al.: RoBERTa-GCN: A Novel Approach for Combating Fake News in Bangla Using Advanced Language Processing

FIGURE 5. Workflow for preparing bangla text data for analysis, featuring cleaning, feature extraction, and various analytical methods.

6) Pesudo Algorithm for Bangla Fake News Data fake. If an article is predicted to be fake, it is flagged, and
Preprossessing relevant stakeholders are notified. This automated process
streamlines the task of identifying misinformation in Bangla
The Algorithm 1 describes a process for preprocessing the news articles. Table 3 outlines the symbols and their mean-
dataset for training our proposed RoBERTa-GCN. It begins ings used in Algorithm 1.
by fetching articles from Bangla news sources. These articles
are preprocessed, which includes normalization, tokeniza-
tion, noise removal, etc., to clean the text. Next, features are IV. PROPOSED: ROBERTA-GCN
extracted using NLP techniques like BOW and TF-IDF. A In NLP and misinformation detection, we present
pre-trained model is then loaded to analyze these features. ‘RoBERTa-GCN,’ an innovative Convolutional (CNN)
Furthermore the pseudocode shows the pipeline for assessing model to detect false information in the Bangla language.
each article and determines whether the content is real or This model is a fusion of RoBERTa’s [54] enhanced lan-
VOLUME 4, 2016 7

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3457860

M. Ahammad et al.: RoBERTa-GCN: A Novel Approach for Combating Fake News in Bangla Using Advanced Language Processing

TABLE 3. Defination of the symbols used in Algorithm 1. TABLE 4. Defination of the symbols used in Algorithm 2.

Symbol Defination Symbol Defination

S Set of news articles x Bangla news article
L Labeling function FRoBERTa pre-trained RoBERTa model
P Articles preprocessing FGCN Graph Convolutional Network model
F Feature extraction P Article preprocessing
M Instantiate a pre-trained model E Article embedding via RoBERTa
ψ Classification function G Graph construction via embedding
F ↕⊣} Flag articles procedure C Classification function
Eval Evaluation function

guage processing abilities, particularly optimized for dy-

namic contexts and the intricate relational data handling representation, which is then processed through several fully
capabilities of Graph Convolutional Network (GCN) [55]. connected neural network layers. These layers reduce the
Tailored to classify news articles into ‘Authentic’ or ‘Fake’ information to a binary classification that, using a sigmoid
categories, RoBERTa-GCN is meticulously trained on a function, distinguishes between real news and propaganda or
diverse dataset of Bangla news, capturing the language’s fake news. This sophisticated architecture Figure 6 ensures
unique characteristics and cultural nuances. It features a nuanced detection tailored for the Bangla language.
dynamic GCN component, crucial for adapting to the ever- As shown in Algorithm 2, our proposed RoBERTa-GCN
evolving language use and patterns in fake news distribution, begins by preprocessing the news article and then generates
ensuring accuracy and robustness. This schematic outlines a embeddings using a pre-trained RoBERTa [54] model. These
Bangla fake news detection model that ingests news articles, embeddings are then used to construct a graph, where nodes
segregating headlines and contents. It employs encoders represent different aspects of the article, and edges define
like RoBERTa-GCN for contextual understanding and topic their relationships. The GCN [55] is then employed to per-
encoders for thematic insights. The model constructs a form graph convolution on this structure, enhancing the node
dynamic graph to encapsulate inter-feature relationships. features by aggregating information from neighboring nodes.
Features are pooled to synthesize the data into a compact The resulting graph is then classified using a classification
function, and the final output is evaluated based on accuracy,
Algorithm 1 Pseudocode for preprocessing the "BanFake- precision, recall, and F1-score. The notation used in Algo-
News" dataset for training our proposed ROBERTa-GCN. rithm 2 is described in Table 4.
Require: S
To represent both the headline and the content as a single
Ensure: L : S → {0, 1}, where 1 corresponds to FAKE and 0 to REAL
1: Procedure Preprocess_and_Classify(S):
2: S ′ ← {s′ | ∀s ∈ S, s′ = P(s)} {Preprocess each article} Algorithm 2 RoBERTa-GCN for news classification
3: V ← {v | ∀s′ ∈ S ′ , v = F (s′ )} {Extract features for each article}
4: M ← M() {Load the machine learning model for prediction} Require: x , FRoBERTa , FGCN
5: L ← {(s, ψ(M, v)) | ∀s ∈ S, ∀v ∈ V, s′ = P(s)} {Label each Ensure: y - classification label (0 for Real, 1 for Fake)
article as FAKE or REAL} 1: Procedure RoBERTa-GCN(x, FRoBERTa , FGCN ) :
6: for each s ∈ S do 2: x′ ← P(x)
7: if L(s) = 1 then 3: E ← E(x′ , FRoBERTa )
8: F ↕⊣}(s, 1) {Flag the article as FAKE} 4: G(V, E) ← G(E)
9: end if 5: G′ (V ′ , E ′ ) ← GraphConvolution(G, FGCN )
10: end for 6: ypred ← C(G′ )
11: return L 7: Eval(ypred , ytrue )
12: Function P(s): 8: return Eval
13: P : S → S′ 9: end Procedure
14: return Normalize(Tokenize(RemoveNoise(s))) {Return a prepro- 10: Function P(x) :
cessed article} 11: P : x → x′
15: Function F (s′ ): 12: return Preprocess(x)
16: F : S ′ → Rn {Mapping to an n-dimensional feature space} 13: Function E(x′ , FRoBERTa ) :
17: return ExtractFeatures(s′ ) {Return the n-dimensional feature vec- 14: E : x′ → E
tor} 15: return FRoBERTa (x′ )
18: Function M(): 16: Function G(E) :
19: M : ∅ → Model 17: V ← CreateNodes(E)
20: return RetrievePretrainedModel() {Return the pre-trained fake news 18: E ← CreateEdges(V )
detection model} 19: return G(V, E)
21: Function ψ(M, v): 20: Function GraphConvolution(G, FGCN ) :
22: ψ : Model × Rn → {0, 1} 21: for each v ∈ V do
23: return M(v) {Apply model M to feature vector v and return label} 22: V ′ ← FGCN (v, G)
24: Procedure F ↕⊣}(s, label): 23: return G′ (V ′ , E)
25: F ↕⊣} : S × {0, 1} → Void 24: Function C(G′ ) :
26: if label = 1 then 25: C : G′ → {0, 1}
27: MarkArticleAsFake(s) {Mark the article s as FAKE} 26: return Classify(G′ )
28: NotifyStakeholders(s) {Notify stakeholders of the FAKE article} 27: Function Eval(ypred , ytrue ) :
29: end if 28: return (accuracy, precision, recall, F1 )

8 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3457860

M. Ahammad et al.: RoBERTa-GCN: A Novel Approach for Combating Fake News in Bangla Using Advanced Language Processing

FIGURE 6. A convoluted working flow demonstrates a Bangla fake news detection model.

vector, embeddings from RoBERTa [54] for each word or encapsulating the relationships between them. The GCN [55]
subword token are first obtained. Then, a pooling operation layer utilizes graph convolution to gather data from neighbor-
(e.g., mean pooling or max pooling) is applied across these ing nodes. The operation for the l-th layer of GCN is defined
embeddings. This pooling process reduces the embeddings to as follows:  
a single vector that captures the essence of the entire article, H (l+1) = σ ÃH (l) W (l) (2)
providing a holistic representation that serves as input for
the GCN. The graph is constructed by treating the unified
In the given context, Ã denotes the normalized adjacency
representation vector as a feature of a single node or multiple
matrix, H (l) represents the node feature matrix at layer l, and
nodes (if further segmenting the article). Edges are defined
W (l) signifies the weight matrix for layer l
based on the relationships between these nodes, which could
be determined through similarity measures, contextual re-
lationships, or predefined rules capturing the structural and 3) Dynamic Adaptation Mechanism (D)
semantic connections within the text. This model is designed The dynamic component adjusts the GCN weights over time
specifically to tackle the challenges of detecting fake news in or other changing factors, represented by:
Bangla, leveraging both language understanding and graph-
based learning. W (l) (t) = g(W (l) , t) (3)

where g is a function that modifies the weights W (l) at each

A. MATHERMATICAL REPRESENTATION OF
layer l in response to the temporal dimension t.
ROBERTA-GCN
As discussed earlier, RoBERTa-GCN is the final result of
4) Classification Layer (C)
integrating the RoBERTa [54] and GCN [55] models. In
this section, we present the mathematical representations that By passing the output of the final GCN layer H (L) through a
detail the integration of these models, forming our proposed classification layer, the probability that an individual article
RoBERTa-GCN for Bangla fake news classification. is of the ‘Authentic’ or ‘Fake’ class is determined:
  
1) RoBERTa Embedding Layer (E) P (y|x) = sigmoid C H (L) (4)
Given a Bangla news article represented as a sequence of
tokens x = {x1 , x2 , ..., xn }, the RoBERTa [54] embedding A sigmoid activation function is used in the RoBERTa-
layer transforms these tokens into dense vector representa- GCN architecture’s classification layer to turn the output
tions: of the last graph convolutional layer into a probability dis-
E(x) = {E(x1 ), E(x2 ), ..., E(xn )} (1) tribution over binary class labels. Mathematically, this is
articulated as P (y|x) = σ(Wc h + bc ), where σ denotes the
where E(xi ) ∈ Rd is the d-dimensional embedding of token sigmoid function, Wc represents the weights of the classifi-
xi obtained through RoBERTa. cation layer, h is the output vector from the last GCN layer,
and bc is the bias term. The output of the sigmoid function,
2) GCN Layer denoted by σ(z) = 1+e1−z , is a scalar between 0 and 1, which
Consider a graph G = (V, E) constructed from the dataset, interprets the likelihood of the input article x being classified
with vertices V representing news articles and edges E as ’Authentic’ (label close to 1) or ’Fake’ (label close to 0).
VOLUME 4, 2016 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3457860

M. Ahammad et al.: RoBERTa-GCN: A Novel Approach for Combating Fake News in Bangla Using Advanced Language Processing

5) Performance Optimization and Regularization (O) 8) Feature Extraction

The training objective includes a regularization component to a: Topic Encoder
mitigate overfitting. The topic encoder maps topics to one-hot encoded vectors in
Rk :
L
(
X 1 if i = j
Ltotal = Loriginal + λ ∥W (l) ∥2F (5) TE(ti ) = ei where eij = (9)
l=1
0 otherwise

Where Ltotal represents the total loss function, Loriginal is b: Sentence Encoder (RoBERTa)
the original loss function, such as cross-entropy The sentence encoding function by RoBERTa transforms the
PL for classi-
fication. λ is the regularization coefficient. l=1 ∥W (l) ∥2F sentence into embedding vectors:
denotes the sum of the squared Frobenius norms of the
SE(c) = (ec1 , ec2 , . . . , ecn ) (10)
weight matrices W (l) at each layer l, which is a common
regularization term to prevent overfitting.
c: Word Representation
Each word from the vocabulary is transformed into a dense
6) Integration of RoBERTa-GCN vector via an embedding matrix:
The complete "RoBERTa-GCN" model, incorporating all the
components, can be formally represented as: ewi = Evwi (11)

RoBERTa-GCN(x, t) = P(y|x) 9) Dynamic s (GCN)

  At each time step t, the GCN updates node features based on


the graph structure:
= softmax C GCN E(x), G(V, E), t (6)   
1 1
H (l+1) (t) = σ I − D(t)− 2 A(t)D(t)− 2 H (l) (t)W (l)
This equation shows the whole process that the model (12)
goes through, from using RoBERTa to embed Bangla news
articles to classifying them using a dynamically adapted 10) Fake News Detection
GCN. The training process also includes regularization and a: Pooling Layer
performance optimization. The evaluation metrics offer a A pooling function reduces the dimensionality of the feature
quantifiable measure of the model’s effectiveness in accu- matrix: !
T
rately categorizing news reports as ‘Authentic’ or ‘Fake.’ M
(L)
hpool = pool H (t) (13)
The problem at hand is to construct a predictive model, de- t=1
noted as RoBERTa-GCN(x, t), which discerns the veracity of
Bangla news articles. The input x represents the textual data b: Fully Connected Layers
of an article, and t encapsulates temporal features. The model The fully connected layers are defined recursively:
leverages the RoBERTa embeddings E(x) to transform x
into a vector space, subsequently processed by a (GCN). The zl = σ (Wl zl−1 + bl ) , z0 = hpool (14)
GCN captures the inter-article relational dynamics over time,
formalized as GCN (E(x), G(V, E), t). The final output y ∈ c: Sigmoid Output
{0, 1}, where 0 denotes ‘Authentic’ and 1 signifies ‘Fake’, is The final classification is determined by the sigmoid func-
generated when a softmax function is applied to the output tion:
⊤


of the GCN, resulting in the probability distribution P (y|x). P (Fake|zLfc ) = σ wout zLfc + bout (15)
The effectiveness of the model is assessed using conventional
binary classification metrics.
B. EXPERIMENTAL SETUP
The dataset used in our study is first described, with an
7) Input News emphasis on its background and the methods used to compile
The input consists of a headline and content represented as it. After that, we’ll go over the data pretreatment processes
sequences of words: that were used to better feed our analytical models. After that,
we go into depth about our research to find the optimal num-
Headline = (h1 , h2 , . . . , hm ) (7) ber of subjects and the process of hyperparameter tweaking
in the modeling technique. Finally, we detail the measures
used to evaluate our all-encompassing method for identi-
Content = (c1 , c2 , . . . , cn ) (8) fying Bangla-language fake news. This research constructs
machine learning models using widely-used Python modules
where hi and cj are tokens indexed by a function ϕ. such as sklearn, keras, and tensorflow. The research tests
10 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3457860

M. Ahammad et al.: RoBERTa-GCN: A Novel Approach for Combating Fake News in Bangla Using Advanced Language Processing

FIGURE 7. Overview of a comprehensive Bangla news detection framework. The framework includes preprocessing of Bangla text datasets, fact-checking
entities, detecting text patterns, various machine learning models, and evaluation metrics to ensure robust fake news detection.

are carried out in a Google Colab setting using a powerful ogy to ensure the credibility of information. It initiates source
GPU backend with 52 GB of system RAM. The experimental verification to authenticate the origin. Claim extraction iso-
setup utilizes an Intel processor as the system model. The lates the main assertions for scrutiny. Evidence retrieval
investigations included using the Python 3 programming then gathers data supporting or refuting the claims. Stance
language to train and test the efficacy of the models. The detection assesses the evidence’s position relative to the
evaluation measures included in the experiments include claim, followed by claim validation against established facts.
accuracy, precision, recall, F1-score, and time complexity, Cross-referencing with trusted sources further corroborates
which are utilized to evaluate the performance of machine findings, while logical inconsistency identification spots con-
learning models. tradictions. Data and statistical verification quantitatively an-
alyze the plausibility. Finally, machine learning-based truth
1) Modeling Approach prediction algorithms synthesize these steps, predicting the
Figure 7 outlines a comprehensive approach for detecting likelihood of the information being factual.
fake news in Bangla. It begins with preprocessing a dataset Detect Text Pattern: The Detect Text Pattern step in Bangla
of Bangla news articles, and then a fact-checking entity runs fake news analysis is a sophisticated process designed to
various checks like source verification, evidence retrieval, identify distinctive linguistic patterns that are commonly as-
and stance detection. Both traditional ML models and Neural sociated with misinformation. This method involves compu-
Network (NN) models are used for analysis. The process tational analysis of the text structure, seeking out anomalies
also includes checking for sensationalism and propaganda or irregularities in the use of language. It encompasses the
techniques. Finally, an evaluation matrix with metrics like identification of stylistic features, narrative frameworks, and
recall, accuracy, and the F1-score assesses the performance discourse constructs that diverge from standard journalistic
of the models. This structured framework is essential for de- practices. By leveraging natural language processing tailored
veloping effective tools to combat misinformation in Bangla to the Bangla language, this step is pivotal in automat-
news. ing the recognition of fabricated content, streamlining the
fact-checking process in the expansive and diverse Bangla-
2) Hyperparameter Tuning speaking media landscape.
Hyperparameter tuning is a critical phase in ML where the Model Tuning. In addressing Bangla fake news, a suite of
ideal hyperparameters for a specific model are determined. traditional machine learning models is deployed to discern
This process entails experimenting with various hyperparam- patterns indicative of misinformation. These models include
eter combinations and assessing the model’s effectiveness K-Nearest Neighbors(KNN) [56], RF [57], AdaBoost [58],
using a test set. The goal of hyperparameter tuning is to Support Vector Machines [59], Stochastic Gradient De-
identify the hyperparameters that yield the highest perfor- scent [60], XGBoost [61], Decision Trees [62], Bagging [63],
mance on the test set. This step is vital as it can enhance Gradient Boosting(GB) [64], and pipelines integrating with
the effectiveness of the ML methods. Table 5 shows the set . Each algorithm brings unique strengths in handling the
of hyperparameters we utilized in our experiments. complexities of Bangla text data, from probabilistic output
Fact-Checking Entity. The Fact-Checking Entity for to ensemble methods like RF and GB that capitalize on
Bangla fake news operates through a systematic methodol- collective decision-making to improve prediction accuracy.
VOLUME 4, 2016 11

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3457860

M. Ahammad et al.: RoBERTa-GCN: A Novel Approach for Combating Fake News in Bangla Using Advanced Language Processing

TABLE 5. The set of Hyperparameter settings used across different models for optimization.

Technique Hyperparameters Values

Logistic Regression solver=’liblinear’, penalty=’l2’, C=1.0, max_iter=100, fit_intercept=True, class_weight=None, random_state=None
Random Forest n_estimators=100, max_depth=None, criterion=’gini’, min_samples_split=2, min_samples_leaf=1, max_features=’auto’
SVC C=1.0, kernel=’rbf’, degree=3, gamma=’scale’, coef0=0.0, shrinking=True, probability=False, tol=1e-3, cache_size=200
XGBC n_estimators=100, max_depth=3, learning_rate=0.1, verbosity=1, objective=’binary:logistic’, booster=’gbtree’
Decision Tree criterion=’gini’, splitter=’best’, max_depth=None, min_samples_split=2, min_samples_leaf=1, min_impurity_decrease=0.0
GBC loss=’deviance’, learning_rate=0.1, n_estimators=100, subsample=1.0, criterion=’friedman_mse’
KNC n_neighbors=5, weights=’uniform’, algorithm=’auto’, leaf_size=30, p=2, metric=’minkowski’, metric_params=None
AdaBoost base_estimator=None, n_estimators=50, learning_rate=1.0, algorithm=’SAMME.R’
SGD loss=’hinge’, penalty=’l2’, alpha=0.0001, l1_ratio=0.15, fit_intercept=True, max_iter=1000, tol=1e-3, shuffle=True, verbose=0
LinearSVC penalty=’l2’, loss=’squared_hinge’, dual=True, tol=1e-4, C=1.0, multi_class=’ovr’, fit_intercept=True, max_iter=1000

These techniques form a robust analytical framework essen- 1) Coherence UCI

tial for automating the detection of fake news in Bangla. The Coherence Uniform Coherence Index (UCI) evaluation
Advanced NN models provide a sophisticated approach to for Bangla fake news detection is an advanced metric assess-
text analysis. Utilizing architectures like BERT with Bi- ing the logical consistency of text within the dataset. By eval-
LSTM or BiGRU layers enhances contextual understand- uating how well the narrative within the news articles holds
ing. XLM-RoBERTa [65] offers cross-lingual capabilities, together, it specifically addresses the particular challenge that
which are crucial for diverse datasets. LSTM [66] networks Bangla’s rich linguistic structure poses. This step is pivotal
capture long-term dependencies within text, while CNNs in discerning the authenticity of news by examining the flow
excel in pattern recognition. Combining CNN with RNN and connection of ideas, an aspect where fabricated stories
layers leverages both local feature extraction and sequence often falter. The evaluation ensures that the Bangla text data
prediction. These NN models are fine-tuned on Bangla text maintains a coherent and believable storyline, which is a
data, capturing the language’s unique syntactic and semantic hallmark of genuine news, thereby enhancing the detection
nuances, thus elevating the accuracy and reliability of fake process’s effectiveness against the spread of misinformation.
news classification. The UCI is formulated as follows:
Model Performance Evaluation. In the combat against
Bangla fake news, evaluating the performance of detection
N  
models is paramount. Key metrics such as precision and 1 X Number of Coherent Segments
recall, F1-score, and overall accuracy serve as the bench- U CI = (16)
N i=1 Total Number of Segments i
marks. Precision measures the model’s ability to correctly
identify fake news, while recall assesses its capability to U CI represents the Uniform Coherence Index for
detect as many fake instances as possible. The F1-score the dataset. N is the number of documents in the
provides a harmonic balance between precision and recall, dataset. i refers to an individual document in the
which is crucial for models where equilibrium is needed. dataset. Number of Coherent Segments in document i is the
Ensuring model robustness involves testing with diverse data, count of segments that are determined to be coherent.
including verified real and deceptive news articles, to validate Total Number of Segments in document i is the count of all
the model’s resilience and its effectiveness in the nuanced segments in the document.
Bangla linguistic context. The Pointwise Mutual Information (PMI) evaluation in
Bangla fake news detection quantifies the association be-
C. METRICS EVALUATION tween words within the text, which is critical for under-
We adhere to a systematic approach that rigorously evalu- standing contextual nuances in the Bengali language. This
ates the model’s performance using a comprehensive set of statistical measure helps identify unusual word pairings often
statistical metrics, including accuracy, precision, recall, and found in fabricated news, enhancing the accuracy of detect-
the F1-score. These metrics are meticulously calibrated to ing misinformation in Bangla text datasets. The mathematical
address the syntactic and semantic intricacies of the Bengali formula for PMI is as follows:
language. A comprehensive evaluation assesses the model’s
ability to distinguish between factual and fraudulent content 
P (x, y)

accurately, ensuring its effectiveness in real-world scenarios PMI(x, y) = log (17)
P (x) × P (y)
where the accuracy of information is critical. The ultimate
objective is to refine these models to achieve near-perfect PMI(x, y) is the PMI between events x and y. P (x, y) is
precision, thereby enhancing the integrity of the Bangla the joint probability of x and y occurring together. P (x) and
digital news landscape. P (y) are the individual probabilities of x and y occurring.
log is the logarithm function.
12 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3457860

M. Ahammad et al.: RoBERTa-GCN: A Novel Approach for Combating Fake News in Bangla Using Advanced Language Processing

2) Coherence NPMI 5) Evaluation Metrics

In the Bangla fake news analysis, the coherence Normalized For the binary classification task in Bangla Fake News Detec-
NPMI evaluation is a pivotal step. It assesses the contextual tion, which discerns between "Fake News" and "Real News",
congruence of words within Bangla text, which is crucial the total number of samples T evaluated is calculated using
for detecting fabricated stories. This metric finely gauges the the following metric:
naturalness and logical flow of news content, enhancing the
M M
robustness of misinformation identification. The mathemati- X X
T = (T Pi + F Ni ) + (F Pi + T Ni ) (21)
cal formula for NPMI is as follows:
i=1 i=1
PMI(x, y) • Clean Accuracy (Cln): Clean Accuracy, often used to
NPMI(x, y) = (18)
− log P (x, y) assess the performance of a model on a dataset free from
NPMI(x, y) is the NPMI between events x and y. adversarial examples or noise, is particularly crucial in
contexts such as fake news detection, where precision is
 y) is thepointwise mutual information, calculated as
PMI(x,
paramount. For a Bangla Fake News Detection system
P (x,y)
log P (x)×P (y) . P (x, y) is the joint probability of x and with two categories (usually "Fake News" and "Real
y occurring together. log represents the natural logarithm News"). Given weights w1 and w2 for ’Real News’ and
function. ’Fake News’ respectively, and a confidence threshold θ,
the Clean Accuracy (Cln) is defined as:
3) Coherence CV
The COHERENCE CV (Coherence Vector) evaluation in w1 · T P1 + w2 · T N2
Cln = (22)
Bangla fake news analysis is a sophisticated tool that lever- W
ages linguistic pattern recognition to ascertain thematic con- where W is the weighted sum of all true positive and
sistency. This method critically appraises the congruity of true negative classifications that exceed the confidence
textual elements, effectively discerning anomalies indicative threshold θ, computed as:
of misinformation, thus bolstering the integrity of news con-
tent in the Bengali language. The Coherence Vector (CV) M
X
involves several steps in its formulation: W = wi ·(T Pi +T Ni ) for all i where P (ŷi |xi ) ≥ θ
1) Representing words as vectors using techniques like i=1
Word2Vec [67], GloVe [68], or FastText [69]. (23)
2) Forming sentence or document vectors by aggregating Here, T P1 is the number of True Positives for ’Real
word vectors. News,’ T N2 is the number of True Negatives for ’Fake
3) Measuring coherence through cosine similarity be- News’, and P (ŷi |xi ) denotes the predicted probability
tween an adjacent sentence or segment vectors. that sample xi is classified as label ŷi . The sum W nor-
malizes the accuracy score, considering the confidence
The cosine similarity is calculated as:
of the predictions and the assigned weights to each class,
⃗ ⃗
⃗ = A·B
⃗ B)
Cosine Similarity(A, (19)
which corrects for class imbalance.
|A||B| where M = 2 represents the two categories of news.
⃗ and B ⃗ are vector representations of sentences or text • Robust Accuracy (Boa): Robust Accuracy (often de-
A noted as Boa for "Best Overall Accuracy") is a metric
segments.· denotes the dot product of the vectors. |A| and designed to measure the performance of a classifica-
|B| are the magnitudes of the vectors. tion model in the presence of adversarial examples or
when the model is subjected to certain perturbations or
4) Topic Diversity
noise. In the context of Bangla Fake News Detection,
The topic diversity evaluation in Bangla fake news analysis where there are two categories (fake news and Authentic
employs advanced text analytics to scrutinize thematic vari- news), the robust accuracy metric would ideally account
ations and breadth. This metric is instrumental in detecting for the model’s resilience against such adversarial con-
aberrant topic distributions, a hallmark of disinformation. By ditions. The following provides the formulation of Boa:
leveraging language-specific models, it significantly refines
the process of discerning authentic news from deceptive N 
narratives in Bengali content. The Topic Diversity (TD) can 1 X
wadv · I ŷiadv = yi · ri
 
Boa =
be quantified using the entropy formula: N i=1
n  (24)
+ wcln · I ŷicln = yi · (1 − ri )
X  
TD = − p(i) log p(i) (20)
i=1
− λ · MCR
Here, n is the total number of unique topics in the text.
p(i) is the probability of the occurrence of topic i. log is the where, N is the total number of samples. ŷiadv and
logarithm function, typically the natural logarithm. ŷicln are the predicted labels under adversarial and clean
VOLUME 4, 2016 13

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3457860

M. Ahammad et al.: RoBERTa-GCN: A Novel Approach for Combating Fake News in Bangla Using Advanced Language Processing

TABLE 6. Performance comparison of traditional Machine Learning algorithms based on various performance metrics such as Accuracy (ACC), Precision
(PRE), Recall (REC), Matthews Correlation Coefficient (MCC), Hinge Loss (HL), Sensitivity(SEN), Specificity (SPE), Positive Predictive Value (PPV),
Negative Predictive Value (NPV) without employing any classification thresholds.

Models ACC PRE REC MCC HL SEN SPE PPV NPV

Linear Regression (LR) 0.879 0.920 0.971 0.475 0.037 0.921 0.323 0.910 0.871
Random Forest (RF) 0.917 0.837 0.866 0.580 0.041 0.805 0.438 0.879 0.926
Support Vector Machine (SVC) 0.853 0.941 0.905 0.695 0.029 0.934 0.249 0.961 0.845
XGBClassifier (XGBC) 0.890 0.921 0.807 0.515 0.058 0.780 0.964 0.809 0.868
Decision Tree (DT) 0.917 0.958 0.982 0.545 0.071 0.393 0.883 0.953 0.792
Gradient Boosting (GBC) 0.884 0.889 0.941 0.747 0.021 0.378 0.923 0.940 0.880
K-Nearest Neighbors (KNN) 0.954 0.923 0.972 0.632 0.304 0.892 0.332 0.871 0.814
AdaBoost 0.836 0.859 0.879 0.485 0.287 0.987 0.931 0.846 0.915
Stochastic Gradient Decent (SGD) 0.769 0.897 0.729 0.428 0.039 0.181 0.845 0.778 0.894
Linear SVC 0.937 0.889 0.918 0.325 0.031 0.952 0.935 0.912 0.884

conditions, respectively, yi is the true label, wadv and at 0.982 and precision at 0.958, highlighting its capability to
wcln are weights for adversarial and clean condition identify true positives, while Random Forest (RF) led to a
performances.ri is the resilience factor for each sample, negative predictive value of 0.926.
I is the indicator function. λ is a regularization parame- Our extensive evaluations of the same models for different
ter, MCR is the Misclassification Rate. classification thresholds are presented in Table 7. At the 0.05
• Probability of Attack Success: In the Probability of threshold, Random Forest showed superior results with the
Attack Success (Succ) in Bangla Fake News Detection, highest accuracy of 0.992, MCC of 0.908, and specificity of
where the system categorizes news into two classes 0.965. XGBClassifier led the recall at 0.998 and exhibited
(’Fake News’ and ’Authentic News’), we can consider the lowest hinge loss of 0.018. At the 0.1 threshold, K-
a scenario where the system is under adversarial attack. Nearest Neighbors stood out with the highest accuracy of
The aim would be to measure how often such attacks 0.966, recall of 0.996, and sensitivity of 0.996, making it
successfully deceive the system. This metric is particu- the most reliable for detecting true positives. The Decision
larly relevant in the context of robust machine learning, Tree maintained its strength in MCC at 0.915 and specificity
where models must resist adversarial examples designed at 0.942, indicating robust classification capabilities. Support
to cause misclassification. The complex equation for Vector Machine models also performed well, particularly in
Succ is defined as follows: precision and positive predictive value. Overall, K-Nearest
Neighbors, Random Forest, and Decision Tree models con-
1 X
N  sistently ranked among the top performers, with specific
si · I ŷiadv ̸= yi · patk
 
Succ = i accuracy, precision, and recall strengths, making them suit-
N i=1
 (25) able for different aspects of fake news classification tasks
 cln  atk depending on the specific metric requirements.
+ (1 − si ) · I ŷi ̸= yi · (1 − pi )

where, N is the total number of samples. si is the B. RESULTS OF DEEP LEARNING MODELS
sensitivity factor for each sample. ŷiadv and ŷicln are the A comparative analysis of the performance metrics for var-
predicted labels under adversarial and clean conditions, ious deep learning models, including BERT + BiLSTM,
respectively. yi is the true label. patki represents the BERT + BiGRU, XLM-RoBERTa, LSTM, CNN, CNN-
probability of an attack on each sample. I[·] is the LSTM, and the proposed RoBERTa-GCN model is presented
indicator function. in Table 8. Among the models, the proposed RoBERTa-
GCN achieves the highest accuracy, with a score of 0.986,
V. EXPERIMENTAL RESULTS demonstrating its superior ability to classify Bangla news
The following sections discuss our experimental results and articles as authentic or fake correctly. It also outperforms the
compare different traditional Machine learning and Deep other models in precision, achieving 0.972, Matthews corre-
learning models with our proposed RoBERTa-GCN. lation coefficient with a score of 0.957, sensitivity at 0.986,
and negative predictive value reaching 0.978. These results
A. RESULTS OF MACHINE LEARNING MODELS indicate its robustness and reliability in detecting fake news.
The performance evaluation of various machine learning Additionally, RoBERTa-GCN exhibits the lowest Hamming
models for fake news classification, as shown in Tables 6 loss at 0.017 and specificity at 0.985, further confirming its
and 7, reveals distinct strengths across different metrics. In effectiveness in minimizing classification errors.
the analysis without classification thresholds as presented in Compared to traditional models like BERT + BiLSTM and
Table 6, K-Nearest Neighbors emerged as the top performer CNN-LSTM, which achieved accuracies of 0.944 and 0.913,
with the highest accuracy of 0.954, MCC of 0.632, and respectively, RoBERTa-GCN’s performance is significantly
specificity of 0.871. The decision Tree excelled in the recall higher. This result highlights the advantage of integrating
14 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3457860

M. Ahammad et al.: RoBERTa-GCN: A Novel Approach for Combating Fake News in Bangla Using Advanced Language Processing

TABLE 7. Comparative of various machine learning models based on different performance metrics for the classification thresholds of 0.05 and 0.1.

Fake ML 0.05 Fake ML 0.1

Models
ACC PRE REC MCC HL SPE PPV NPV ACC PRE REC MCC HL SPE PPV NPV
LR 0.970 0.973 0.997 0.583 0.031 0.785 0.943 0.875 0.900 0.874 0.935 0.785 0.045 0.786 0.914 0.895
RF 0.992 0.992 0.941 0.908 0.085 0.965 0.893 0.936 0.894 0.965 0.892 0.698 0.023 0.887 0.897 0.912
SVC 0.986 0.988 0.998 0.829 0.014 0.895 0.885 0.882 0.908 0.895 0.910 0.723 0.012 0.935 0.793 0.874
XGBC 0.992 0.917 0.998 0.906 0.018 0.912 0.895 0.910 0.914 0.905 0.941 0.866 0.085 0.787 0.698 0.932
DT 0.979 0.994 0.985 0.787 0.208 0.935 0.784 0.897 0.936 0.963 0.932 0.915 0.036 0.894 0.940 0.913
GBC 0.965 0.967 0.996 0.477 0.036 0.893 0.936 0.940 0.943 0.918 0.894 0.943 0.047 0.942 0.895 0.923
KNN 0.969 0.971 0.997 0.564 0.031 0.782 0.890 0.935 0.893 0.936 0.967 0.896 0.065 0.914 0.918 0.914
AdaBoost 0.961 0.970 0.991 0.460 0.039 0.938 0.901 0.941 0.915 0.841 0.944 0.875 0.074 0.874 0.932 0.895
SGD 0.964 0.944 0.895 0.451 0.036 0.923 0.913 0.785 0.966 0.969 0.996 0.506 0.063 0.941 0.892 0.923
LinearSVC 0.971 0.975 0.896 0.602 0.029 0.789 0.941 0.931 0.945 0.892 0.914 0.651 0.071 0.936 0.969 0.935

TABLE 8. Performance metrics comparison for traditional DL algorithms with RoBERTa-GCN

Models ACC PRE REC MCC HL SEN SPE PPV NPV

BERT + BiLSTMa 0.944 0.941 0.961 0.918 0.024 0.917 0.932 0.935 0.918
BERT + BiGRU 0.936 0.952 0.916 0.880 0.081 0.735 0.978 0.914 0.904
XLM-RoBERTa 0.953 0.948 0.924 0.941 0.125 0.983 0.214 0.951 0.958
LSTM 0.890 0.954 0.934 0.935 0.061 0.940 0.964 0.969 0.898
CNN 0.917 0.938 0.932 0.947 0.042 0.823 0.883 0.974 0.897
CNN-LSTM 0.913 0.789 0.941 0.937 0.018 0.878 0.923 0.957 0.934
RoBERTa-GCN 0.986 0.972 0.913 0.957 0.017 0.986 0.985 0.969 0.978

TABLE 9. Comparison of our proposed RoBERTa-GCN with different

model learning techniques and their performances based on accuracy BanglaBERT [49], which achieved 0.90 and 0.834 accuracy,
(ACC) metrics. respectively in 2022, and traditional machine learning tech-
Models Year Methods ACC niques like SVM [3] and RF [26], which reported 0.910 and
SVM [3] 2020 Machine Learning 0.910 0.85 accuracy, respectively in 2020. Transfer learning also
Random Forest [26] 2020 Machine Learning 0.850 proved effective, with XLM-RoBERTa [29] achieving 0.9331
Hybrid CNN-RNN [24] 2021 Deep Learning 0.600 accuracy in 2023. Among deep learning models, BerCon-
BerConvoNet [31] 2021 Deep Learning 0.974 voNet [31] and FNDNet [70] showed strong performance
BERT [27] 2022 Machine Learning 0.900 with accuracies of 0.974 and 0.983, respectively, while the
BanglaBERT [49] 2022 Deep Learning 0.834
CT-BERT+BiGRU [33] model reached 0.984 in 2023. The
XLM-RoBERT [29] 2023 Deep Learning 0.933
FNDNet [70] 2023 Deep Learning 0.983
key finding from this table is that our proposed RoBERTa-
CT-BERT+BiGRU [33] 2023 Deep Learning 0.984 GCN model outperforms all other models with the highest
RoBERTa-GCN (Our study) 2024 Deep Learning 0.986 accuracy of 0.986. This highlights the superior effective-
ness of integrating RoBERTa’s language understanding with
GCN’s relational learning capabilities, marking a significant
RoBERTa’s contextual language understanding with the rela- advancement in the field of Bangla fake news detection.
tional learning capabilities of GCN, making RoBERTa-GCN Overall, the table underscores the steady progression of
the most effective model in this study for Bangla fake news model performance over time, culminating in the SOTA
detection. results achieved by RoBERTa-GCN.
The ROC curve in Figure 9 displays the performance of a
C. COMPARATIVE ANALYSIS binary classifier, which could be the RoBERTa-GCN model,
Figure 8 presents a comprehensive evaluation of various ma- by plotting the True Positive Rate (TPR) against the False
chine learning models against multiple performance metrics. Positive Rate (FPR). The ROC curve shows an excellent
The performance scores are plotted to facilitate a visual com- diagnostic ability of the model, as indicated by the AUC
parison, highlighting the efficacy of each model across the of 0.96, suggesting that the model can distinguish between
metrics. Adding RoBERTa-GCN makes it possible to com- classes with high accuracy. The curve closely approaches
pare it to more advanced models that use both transformer the upper left corner of the graph, which represents an ideal
architectures and GCN. These models are hybrid models. classifier yielding a high TPR and a low FPR. This is indica-
Moreover, Table 9 compares the performance accuracy of tive of a highly effective model in terms of both sensitivity
various learning techniques published between 2020 and (identifying true positives) and specificity (correctly avoiding
2023 for fake news detection. Early approaches, such as the false positives). The model’s performance, as illustrated by
Hybrid CNN-RNN [24] model in 2021, achieved a moder- this ROC curve, is therefore exemplary, with the curve nearly
ate accuracy of 0.600, indicating room for improvement in reaching the perfect classification point at the top left.
deep learning models at that time. As the field advanced,
significant gains were seen with models like BERT [27] and
VOLUME 4, 2016 15

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3457860

M. Ahammad et al.: RoBERTa-GCN: A Novel Approach for Combating Fake News in Bangla Using Advanced Language Processing

FIGURE 8. Comparison of proposed RoBERTa-GCN with other models based on different performance metrics.

D. EXPLAINABLITY OF ROBERTA-GCN as the first one, with different words highlighted. This
Figure 10 depicts the output of a Local Interpretable Model- suggests that those words have a significant impact on
agnostic Explanations (LIME) analysis applied to our pro- the model’s prediction of authenticity.
posed RoBERTa-GCN. It helps to explain the predictions of 3) The third text is predicted with absolute certainty
our RoBERTa-GCN model by highlighting the features, in (probability 1.00) to be fake, and no probability is
this case, words that contribute most to the model’s predic- assigned to it being authentic. Highlighted words in
tion. Based on the Figure 10 we can get some insights as the text are the features that the model found most
follows: indicative of it being fake.
1) The first text has a prediction probability of 0.92 for The right side of the image, where the texts with high-
being authentic and 0.08 for being fake. Certain words lighted words are shown, illustrates how LIME provides local
are highlighted, which likely influenced the model’s interpretability. For each prediction, it points out the specific
decision to classify it as authentic. words that have the highest weight in the model’s decision-
2) The second text has the same prediction probabilities making process. This is crucial for understanding why the
model makes certain decisions and can be used to improve
the model by, for example, adjusting feature weightings or
by providing more training data to reduce biases.
Moreover, the LIME analysis reveals that our proposed
RoBERTa-GCN model does not rely solely on superficial
cues or common phrases but rather on a nuanced under-
standing of the text’s context, enabled by the integration of
RoBERTa’s language understanding and GCN’s relational
reasoning. This interpretability is crucial for validating the
model’s reliability, especially in sensitive applications like
fake news detection, where understanding the rationale be-
hind a classification is as important as the classification itself.

VI. CONCLUSION
This paper represents a pivotal stride in combating the es-
FIGURE 9. The obtained ROC curve of our proposed RoBERTa-GCN calating problem of misinformation in the digital age, par-
model. ticularly within the context of the Bengali language. As the
16 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3457860

M. Ahammad et al.: RoBERTa-GCN: A Novel Approach for Combating Fake News in Bangla Using Advanced Language Processing

FIGURE 10. LIME analysis for Bangla fake news detection, highlighting words influencing model predictions. Blue indicates words supporting
authenticity; orange indicates words supporting fake news.

digital ecosystem becomes increasingly saturated with de- includes a new model design and a deep understanding of
ceptive content, the need for sophisticated, language-specific regional linguistic traits, opens the door for more research in
solutions has never been more pressing. Addressing this this area and shows how important customized AI solutions
need, the research introduces the RoBERTa-GCN model, a are for keeping information safe in a world that is becoming
fusion of RoBERTa’s advanced natural language processing more and more connected.
capabilities and GCN’s adeptness at managing relational
data. This model is meticulously tailored to grasp the intrica- A. LIMITATIONS AND FUTURE SCOPES
cies of the Bangla language, encompassing its rich linguistic
Although RoBERTa-GCN is good at identifying Bangla fake
nuances and cultural contexts. Trained on a comprehensive
news, it has certain shortcomings that might make it less
dataset collected from diverse Bangladeshi news sources, the
applicable to other languages and training datasets. A poten-
RoBERTa-GCN model is rigorously optimized to distinguish
tial future direction for improving the model’s usefulness is
authentic news from fabricated narratives effectively. Its
to investigate ways to include other South Asian languages.
performance, benchmarked against several baseline models,
In addition, there is potential for modifying RoBERTa-GCN
demonstrates a marked improvement, affirming its potential
to identify more nuanced types of disinformation, such as
as a critical tool in the arsenal against misinformation.
biased journalism or satire. To further explore the model’s
The significance of the RoBERTa-GCN model transcends
efficacy across other social media platforms and include
its technical achievements; it addresses a crucial gap in the
real-time data analysis, more studies might be conducted to
field of fake news detection for languages that have tradi-
expand its practical applicability.
tionally been underrepresented in global NLP research. By
leveraging state-of-the-art NLP techniques and adapting to
the unique challenges posed by the Bangla language, the ACKNOWLEDGEMENT
model not only sets a new standard for fake news detection We would like to acknowledge Dr. Sami Azam and Dr. Asif
but also underscores the importance of incorporating lin- Karim from the Faculty of Science and Technology, Charles
guistic and cultural idiosyncrasies into AI-driven solutions. Darwin University, Casuarina, NT 0909, Australia, for their
The model’s evaluation, employing metrics like accuracy, assistance in securing funding and their thorough review of
precision, recall, and F1-score, attests to its robustness and this paper.
reliability in real-world settings. Furthermore, the use of
advanced coherence measures like UCI, PMI, and NPMI REFERENCES
ensures that the model’s performance is both consistent and
[1] M. H. Goldani, R. Safabakhsh, and S. Momtazi, “Convolutional neural
dependable, making it a formidable tool against the spread network with margin loss for fake news detection,” Information Processing
of misinformation. The study’s all-around approach, which & Management, vol. 58, no. 1, p. 102418, 2021.

VOLUME 4, 2016 17

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3457860

M. Ahammad et al.: RoBERTa-GCN: A Novel Approach for Combating Fake News in Bangla Using Advanced Language Processing

[2] M. Narra, M. Umer, S. Sadiq, H. Karamti, A. Mohamed, I. Ashraf, et al., International Conference on Computing Communication and Networking
“Selective feature sets based fake news detection for covid-19 to manage Technologies (ICCCNT), pp. 1–5, IEEE, 2021.
infodemic,” IEEE Access, vol. 10, pp. 98724–98736, 2022. [26] F. Islam, M. M. Alam, S. S. Hossain, A. Motaleb, S. Yeasmin, M. Hasan,
[3] M. Z. Hossain, M. A. Rahman, M. S. Islam, and S. Kar, “Banfak- and R. M. Rahman, “Bengali fake news detection,” in 2020 IEEE 10th
enews: A dataset for detecting fake news in bangla,” arXiv preprint International Conference on Intelligent Systems (IS), pp. 281–287, IEEE,
arXiv:2004.08789, 2020. 2020.
[4] D. Wang, W. Zhang, W. Wu, and X. Guo, “Soft-label for multi-domain [27] M. A. Al Ghamdi, M. S. Bhatti, A. Saeed, Z. Gillani, and S. H. Almotiri,
fake news detection,” IEEE Access, 2023. “A fusion of bert, machine learning and manual approach for fake news
[5] T. Jiang, J. P. Li, A. U. Haq, A. Saboor, and A. Ali, “A novel stacking detection,” Multimedia Tools and Applications, pp. 1–18, 2023.
approach for accurate detection of fake news,” IEEE Access, vol. 9, [28] E. Raja, B. Soni, and S. K. Borgohain, “Fake news detection in dravidian
pp. 22626–22639, 2021. languages using transfer learning with adaptive finetuning,” Engineering
[6] M. G. Hussain, M. R. Hasan, M. Rahman, J. Protim, and S. Al Hasan, Applications of Artificial Intelligence, vol. 126, p. 106877, 2023.
“Detection of bangla fake news using mnb and svm classifier,” in 2020 [29] P. B. Pranto, S. Z.-U.-H. Navid, P. Dey, G. Uddin, and A. Iqbal, “Are you
International Conference on Computing, Electronics & Communications misinformed? a study of covid-related fake news in bengali on facebook,”
Engineering (iCCECE), pp. 81–85, IEEE, 2020. arXiv preprint arXiv:2203.11669, 2022.
[7] A. Roets et al., “‘fake news’: Incorrect, but hard to correct. the role of [30] S. Rohman, J. Ferdous, S. M. R. Ullah, and M. A. Rahman, “Ibfnd: An
cognitive ability on the impact of false information on social impressions,” improved dataset for bangla fake news detection and comparative analysis
Intelligence, vol. 65, pp. 107–110, 2017. of performance of baseline models,” in 2023 International Conference on
[8] H. Saleh, A. Alharbi, and S. H. Alsamhi, “Opcnn-fake: Optimized con- Next-Generation Computing, IoT and Machine Learning (NCIM), pp. 1–6,
volutional neural network for fake news detection,” IEEE Access, vol. 9, IEEE, 2023.
pp. 129471–129489, 2021. [31] M. Choudhary, S. S. Chouhan, E. S. Pilli, and S. K. Vipparthi, “Bercon-
[9] E. Grave, P. Bojanowski, P. Gupta, A. Joulin, and T. Mikolov, “Learning vonet: A deep learning framework for fake news classification,” Applied
word vectors for 157 languages,” arXiv preprint arXiv:1802.06893, 2018. Soft Computing, vol. 110, p. 107614, 2021.
[10] S. Vosoughi, D. Roy, and S. Aral, “The spread of true and false news [32] M. A. Ali, M. L. Matubber, V. Sharma, and B. Balamurugan, “An improved
online,” science, vol. 359, no. 6380, pp. 1146–1151, 2018. and efficient technique for detecting bengali fake news using machine
[11] S. KEMP, “Digital 2023: Bangladesh — datareportal – global digital in- learning algorithms,” in 2022 6th International Conference On Computing,
sights.” https://datareportal.com/reports/digital-2023-bangladesh, Feb.13, Communication, Control And Automation (ICCUBEA, pp. 1–4, IEEE,
2023. (Accessed on 12/21/2023). 2022.
[33] J. Alghamdi, Y. Lin, and S. Luo, “Towards covid-19 fake news detection
[12] M. Hossain, “Nothing but fb. why bangladeshis never
using transformer-based models,” Knowledge-Based Systems, vol. 274,
took to twitter, threads and the like | the business
p. 110642, 2023.
standard.” https://www.tbsnews.net/features/panorama/
[34] D. K. Dixit, A. Bhagat, and D. Dangi, “Automating fake news detection
nothing-fb-why-bangladeshis-never-took-twitter-threads-and-667186,
using ppca and levy flight-based lstm,” Soft Computing, vol. 26, no. 22,
July.18, 2023. (Accessed on 11/15/2023).
pp. 12545–12557, 2022.
[13] S. Report, “Mobs beat 2 dead for ‘kidnapping’ | daily star.” https://www.
[35] N. Rai, D. Kumar, N. Kaushik, C. Raj, and A. Ali, “Fake news classifica-
thedailystar.net/frontpage/news/mobs-beat-2-dead-kidnapping-1774471,
tion using transformer based enhanced lstm and bert,” International Journal
July .21, 2019. (Accessed on 12/07/2023).
of Cognitive Computing in Engineering, vol. 3, pp. 98–105, 2022.
[14] I. Ahmed and J. A. Manik, “A hazy picture appears | the daily star.” https:
[36] M. Sudhakar and K. Kaliyamurthie, “Effective prediction of fake news
//www.thedailystar.net/news-detail-252212, Oct .3, 2012. (Accessed on
using two machine learning algorithms,” Measurement: Sensors, vol. 24,
12/02/2023).
p. 100495, 2022.
[15] R. Rafe, “Misinformation mars bangladesh vaccina-
[37] A. Choudhary and A. Arora, “Linguistic feature based learning model for
tion drive – dw – 01/27/2021.” https://www.dw.com/en/
fake news detection and classification,” Expert Systems with Applications,
covid-bangladesh-vaccination-drive-marred-by-misinformation/
vol. 169, p. 114171, 2021.
a-56360529, Jan .27, 2021. (Accessed on 01/05/2024).
[38] V. Khullar and H. P. Singh, “f-fnc: Privacy concerned efficient federated
[16] T. Report, “Busting the top 3 fake news of the week | approach for fake news classification,” Information Sciences, vol. 639,
the business standard.” https://www.tbsnews.net/thoughts/ p. 119017, 2023.
busting-top-3-fake-news-week-173236, Dec .18, 2020. (Accessed [39] H. Xia, Y. Wang, J. Z. Zhang, L. J. Zheng, M. M. Kamal, and V. Arya,
on 01/10/2024). “Covid-19 fake news detection: A hybrid cnn-bilstm-am model,” Techno-
[17] “Politifact.” https://www.politifact.com/. (Accessed on 01/25/2024). logical Forecasting and Social Change, vol. 195, p. 122746, 2023.
[18] “Factcheck.org - a project of the annenberg public policy center.” https: [40] M. Hosseini, A. J. Sabet, S. He, and D. Aguiar, “Interpretable fake news
//www.factcheck.org/. (Accessed on 01/09/2024). detection with topic and deep variational models,” Online Social Networks
[19] “Verification - fact check bangladesh. fact-check bangladesh.” https:// and Media, vol. 36, p. 100249, 2023.
www.jachai.org/. (Accessed on 01/20/2024). [41] B. Palani and S. Elango, “Bbc-fnd: An ensemble of deep learning frame-
[20] Y. Long, Q. Lu, R. Xiang, M. Li, and C.-R. Huang, “Fake news detection work for textual fake news detection,” Computers and Electrical Engineer-
through multi-perspective speaker profiles,” in Proceedings of the eighth ing, vol. 110, p. 108866, 2023.
international joint conference on natural language processing (volume 2: [42] O. B. Okunoye and A. E. Ibor, “Hybrid fake news detection technique with
Short papers), pp. 252–256, 2017. genetic search and deep learning,” Computers and Electrical Engineering,
[21] G. Karadzhov, P. Nakov, L. Màrquez, A. Barrón-Cedeño, and I. Koychev, vol. 103, p. 108344, 2022.
“Fully automated fact checking using external sources,” arXiv preprint [43] P. Malhotra and S. K. Malik, “Fake news detection using ensemble
arXiv:1710.00341, 2017. techniques,” Multimedia Tools and Applications, 2023.
[22] M. A. Haque Palash, A. Khan, K. Islam, M. A. Al Nasim, and R. M. [44] R. Mohawesh, S. Maqsood, and Q. Althebyan, “Multilingual deep learning
Bin Shahjahan, “Incongruity detection between bangla news headline and framework for fake news detection using capsule neural network,” Journal
body content through graph neural network,” in The Fourth Industrial of Intelligent Information Systems, pp. 1–17, 2023.
Revolution and Beyond: Select Proceedings of IC4IR+, pp. 375–387, [45] V. Jain, R. K. Kaliyar, A. Goswami, P. Narang, and Y. Sharma, “Aenet:
Springer, 2023. an attention-enabled neural architecture for fake news detection using
[23] R. Kawser, “Bangla ranked at 7th among 100 most spoken contextual features,” Neural Computing and Applications, vol. 34, no. 1,
languages worldwide.” https://www.dhakatribune.com/world/201648/ pp. 771–782, 2022.
bangla-ranked-at-7th-among-100-most-spoken, Feb .17, 2020. [46] B. Palani, S. Elango, and V. Viswanathan K, “Cb-fake: A multimodal deep
(Accessed on 01/01/2024). learning framework for automatic fake news detection using capsule neural
[24] J. A. Nasir, O. S. Khan, and I. Varlamis, “Fake news detection: A hybrid network and bert,” Multimedia Tools and Applications, vol. 81, no. 4,
cnn-rnn based deep learning approach,” International Journal of Informa- pp. 5587–5620, 2022.
tion Management Data Insights, vol. 1, no. 1, p. 100007, 2021. [47] R. Katarya, D. Dahiya, S. Checker, et al., “Fake news detection system us-
[25] A. Anjum, M. Keya, A. K. M. Masum, and S. R. H. Noori, “Fake ing featured-based optimized msvm classification,” IEEE Access, vol. 10,
and authentic news detection using social data strivings,” in 2021 12th pp. 113184–113199, 2022.

18 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3457860

M. Ahammad et al.: RoBERTa-GCN: A Novel Approach for Combating Fake News in Bangla Using Advanced Language Processing

[48] K. M. Hasib, N. A. Towhid, K. O. Faruk, J. Al Mahmud, and M. Mridha, MEJBAH AHAMMAD received his BSc in com-
“Strategies for enhancing the performance of news article classification in puter science and engineering with the major of
bangla: Handling imbalance and interpretation,” Engineering Applications computer vision and his MSc in computer science
of Artificial Intelligence, vol. 125, p. 106688, 2023. with the major of artificial intelligence from Amer-
[49] A. Bhattacharjee, T. Hasan, W. U. Ahmad, K. Samin, M. S. Islam, A. Iqbal, ican International University, Bangladesh. He is
M. S. Rahman, and R. Shahriyar, “Banglabert: Language model pretrain- currently the CEO of Bytes of Intelligence, New
ing and benchmarks for low-resource language understanding evaluation York, USA, and works as a deep learning and
in bangla,” arXiv preprint arXiv:2101.00204, 2021.
AI-specialized consultant and instructor at aiQuest
[50] W. A. Qader, M. M. Ameen, and B. I. Ahmed, “An overview of bag of
Intelligence, Dhaka, Bangladesh. He has authored
words; importance, implementation, applications, and challenges,” in 2019
international engineering conference (IEC), pp. 200–204, IEEE, 2019. over 11 research papers. He has the Best Research
[51] J. Ramos et al., “Using tf-idf to determine word relevance in document Award from the International Conference ICETET-SIP-22.
queries,” in Proceedings of the first instructional conference on machine
learning, vol. 242, pp. 29–48, Citeseer, 2003.
[52] Z. Yin and Y. Shen, “On the dimensionality of word embedding,” Ad-
vances in neural information processing systems, vol. 31, 2018.
[53] G. Sidorov, F. Velasquez, E. Stamatatos, A. Gelbukh, and L. Chanona-
Hernández, “Syntactic n-grams as machine learning features for natural
language processing,” Expert Systems with Applications, vol. 41, no. 3, AL SANI received his B.Sc. degree in Indus-
pp. 853–860, 2014. trial and Production Engineering from Jashore
[54] Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, University of Science and Technology (JUST),
L. Zettlemoyer, and V. Stoyanov, “Roberta: A robustly optimized bert Jashore–7408, Bangladesh. His current research
pretraining approach,” arXiv preprint arXiv:1907.11692, 2019. interests include artificial intelligence (AI), ma-
[55] T. N. Kipf and M. Welling, “Semi-supervised classification with graph chine learning (ML), and deep learning (DL),
convolutional networks,” arXiv preprint arXiv:1609.02907, 2016. specifically image processing, computer vision,
[56] L. E. Peterson, “K-nearest neighbor,” Scholarpedia, vol. 4, no. 2, p. 1883. and natural language processing. Additionally, he
[57] L. Breiman, “Random forests,” Machine learning, vol. 45, pp. 5–32, 2001. is passionate about exploring the integration of
[58] Y. Freund and R. E. Schapire, “A decision-theoretic generalization of on- industrial engineering principles with AI technolo-
line learning and an application to boosting,” Journal of computer and
gies to optimize manufacturing processes, enhance production efficiency,
system sciences, vol. 55, no. 1, pp. 119–139, 1997.
and innovate smart industrial systems. His academic and research career
[59] C. Cortes and V. Vapnik, “Support-vector networks,” Machine learning,
vol. 20, pp. 273–297, 1995. demonstrates a deep fascination with the intersection of engineering and
[60] L. Bottou, “Stochastic gradient descent tricks,” in Neural Networks: Tricks cutting-edge technologies, emphasizing an interdisciplinary approach. AL
of the Trade: Second Edition, pp. 421–436, Springer, 2012. SANI is eager to contribute to the advancement of intelligent industrial
[61] T. Chen and C. Guestrin, “Xgboost: A scalable tree boosting system,” in systems through groundbreaking research and innovative solutions.
Proceedings of the 22nd acm sigkdd international conference on knowl-
edge discovery and data mining, pp. 785–794, 2016.
[62] W.-Y. Loh, “Classification and regression trees,” Wiley interdisciplinary
reviews: data mining and knowledge discovery, vol. 1, no. 1, pp. 14–23,
2011.
[63] L. Breiman, “Bagging predictors,” Machine learning, vol. 24, pp. 123–140,
1996.
[64] J. H. Friedman, “Greedy function approximation: a gradient boosting KHALILUR RAHMAN received a B.Sc. degree
machine,” Annals of statistics, pp. 1189–1232, 2001. in Computer Science and Engineering from Sylhet
[65] A. Conneau, K. Khandelwal, N. Goyal, V. Chaudhary, G. Wenzek, Engineering College, Sylhet-3100, Bangladesh.
F. Guzmán, E. Grave, M. Ott, L. Zettlemoyer, and V. Stoyanov, “Un- His research interest in Deep learning, Ma-
supervised cross-lingual representation learning at scale,” arXiv preprint chine learning, and Artificial intelligence. He has
arXiv:1911.02116, 2019. worked on various projects in the field of appli-
[66] M. Sundermeyer, R. Schlüter, and H. Ney, “Lstm neural networks for cation of machine learning. His focus lies specifi-
language modeling.,” in Interspeech, vol. 2012, pp. 194–197, 2012.
cally in Computational Bioinformatics, Computer
[67] K. W. Church, “Word2vec,” Natural Language Engineering, vol. 23, no. 1,
Vision, and Image Processing.
pp. 155–162, 2017.
[68] J. Pennington, R. Socher, and C. D. Manning, “Glove: Global vectors for
word representation,” in Proceedings of the 2014 conference on empirical
methods in natural language processing (EMNLP), pp. 1532–1543, 2014.
[69] A. Joulin, E. Grave, P. Bojanowski, M. Douze, H. Jégou, and T. Mikolov,
“Fasttext. zip: Compressing text classification models,” arXiv preprint
arXiv:1612.03651, 2016.
[70] R. K. Kaliyar, A. Goswami, P. Narang, and S. Sinha, “Fndnet–a deep
convolutional neural network for fake news detection,” Cognitive Systems MD TANVIR ISLAM is a graduate student pursu-
Research, vol. 61, pp. 32–44, 2020. ing an M.Sc. in Computer Science and Engineer-
ing at Sungkyunkwan University (SKKU), South
Korea. Specializing in AI-driven image process-
ing, his research has led to publications in top-
tier journals and conferences such as "Alexan-
dria Engineering Journal" and "ACM Multime-
dia". He has experience as a Research Assistant
at "VIS2KNOW Lab", focusing on computer vi-
sion, deep learning, and machine learning. He is
passionate about applying advanced computational techniques to solve real-
world challenges and has been actively involved in academic and research
communities.

VOLUME 4, 2016 19

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3457860

M. Ahammad et al.: RoBERTa-GCN: A Novel Approach for Combating Fake News in Bangla Using Advanced Language Processing

MD MOSTAFIZUR RAHMAN MASUD is now

pursuing a Bachelor of Science in Software Engi-
neering at Universiti Teknologi Malaysia. He has
shown a strong dedication to acquiring informa-
tion in the field of computers, with a specific focus
on the overlap between machine learning, deep
learning, NLP, and computer vision.

MD. MEHEDI HASSAN (MEMBER, IEEE)

is a dedicated and accomplished researcher, is
completed the Master of Science (M.Sc.) degree
in computer science and engineering at Khulna
University, Khulna, Bangladesh in 2024. He is
a member of the prestigious Institute of Electri-
cal and Elec- tronics Engineers (IEEE). Mehedi
completed his BSc degree in Computer Science
and Engineering from North Western University,
Khulna in 2022, where he excelled in his studies
and demonstrated a strong aptitude for research. As the founder and CEO
of The Virtual BD IT Firm and VRD Research Laboratory, Bangladesh,
Mehedi has established himself as a highly respected leader in the fields
of biomedical engineering, data science, and expert systems. Mehedi’s
research interests are broad and include important human diseases, such
as oncology, cancer, and hepatitis, as well as human behavior analysis and
mental health. He is highly skilled in association rule mining, predictive
analysis, machine learning, and data analysis, with a particular focus on
the biomedical sciences. As a young researcher, Mehedi has published
55 articles in various international top journals and conferences, which
is a remarkable achievement. His work has been well-received by the
research community and has significantly contributed to the advancement
of knowledge in his field. Overall, Mehedi is a highly motivated and skilled
researcher with a strong commitment to improving human health and well-
being through cutting-edge scientific research. His accomplishments to date
are impressive, and his potential for future contributions to his field is very
promising. Additionally, he serves as an academic editor of PLOS One,
PLOS Digital Health, Discover Applied Sciences (Springer) and a reviewer
for 56 prestigious journals. Md. Mehedi Hassan has filed more than 3 patents
out of which 3 are granted to his name.

20 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/