multiBERT: A Classifier for Sponsored Social Media Content

Kshitij Salil Malvankar

1,2

, Enda Fallon

, Paul Connolly

and Kieran Flanagan

Software Research Institute, Technological University of Shannon, Athlone, Ireland

Circana Inc, Athlone, Ireland

Keywords: Bert, Social Media, Influencer.

Abstract: Social media's rise has given birth to a new class of celebrities called influencers. People who have amassed

a following on social media sites like Twitter, YouTube, and Instagram are known as influencers. These

people have the ability to sway the beliefs and purchase choices of those who follow them. Consequently,

companies have looked to collaborate with influencers in order to market their goods and services. But as

sponsored content has grown in popularity, it has becoming harder to tell if a piece is an independent opinion

of an influencer or was sponsored by a company. This study investigates the use of machine learning models

to categorise influencer tweets as either sponsored or unsponsored. By utilising transformer language models,

like BERT, we are able to discover relationships and patterns between a brand and an influencer. Machine

learning algorithms may assist in determining if a tweet or Instagram post is a sponsored post or not by

examining the context and content of influencer tweets and their Instagram post captions. To evaluate data

from Instagram and Twitter together, this work presents a novel method that compares the models while

accounting for performance criteria including accuracy, precision, recall, and F1 score.

1 INTRODUCTION

The social media sector has experienced significant

growth, not only enabling individuals to communicate

with one another, but also creating career prospects that

were previously unimaginable. Social media has

provided opportunities for content writers and

influencers to gain recognition, popularity, and

financial success. Additionally, it serves as a platform

for online purchasing. One kind of social media

marketing is Influencer Marketing, when an individual

with expertise in a certain sector use their knowledge

to promote the brand and products of others.

Many businesses these days make use of

influencer marketing as part of their overall

marketing strategy. When the influencer promotes the

brand or it’s products, the brand then compensates the

influencer appropriately. Potential methods of

marketing including product placement or

evaluations, sponsored postings, or sponsored events.

The influencer's objective is to enhance brand

recognition among their followers and stimulate the

purchase of the brand's items. The influencer

marketing industry is seeing rapid expansion.

Influencer marketing offers a distinct advantage for

businesses by providing more precise targeting

capabilities in comparison to traditional advertising

tactics. Brands can engage in partnerships with

influencers that specialise in targeting certain

demographics that align well with their business

objectives, therefore streamlining the process of

reaching their desired audience. For example, a Sport

or Health drink brand may form a partnership with a

fitness influencer or an online fitness coach who

possesses a significant following of other fitness

enthusiasts. The followers of the influencer place their

trust in the influencer and their recommendations, thus

leading to an increase in the brand’s sales figures.

With the growth of the industry, it has become

essential to differentiate between genuine sponsored

content and non-sponsored content. This is true from

both the consumer's and the business's point of view.

For the consumer, it is essential to be aware of and

differentiate between whether the influencers they

follow are posting organic content or whether that

content is being paid for. Additionally, for the brand,

this would provide the brand with a more

comprehensive understanding of the shifting

dynamics of influencers and assist them in

quantifying the impact of such influencers.

The objective of this study is to utilise actual

influencer data collected from Twitter and Instagram

706

Malvankar, K., Fallon, E., Connolly, P. and Flanagan, K.

multiBERT: A Classiﬁer for Sponsored Social Media Content.

DOI: 10.5220/0012632400003690

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 26th International Conference on Enterprise Information Systems (ICEIS 2024) - Volume 1, pages 706-713

ISBN: 978-989-758-692-7; ISSN: 2184-4992

to train BERT, along with a modified approach based

on BERT, named multiBERT in order to effectively

determine whether a tweet or Instagram post from an

influencer is sponsored or not by the brand. The

performance of these models will also be compared

and evaluated based on evaluation metrics such as

accuracy, precision, recall and F1 score.

2 RELATED WORK

Regarding work that involves data derived from

Twitter and Instagram, multiple studies have been

done to explore the potential of using this data,

providing valuable insights into its applicability.

Yadav et al. proposed utilising a machine learning

classifier to do sentiment analysis on Twitter. The

researchers utilised the Kaggle dataset, which

comprised phrases and keywords pertaining to a

certain product. The writers intended to categorise

types of positive and negative attitudes. During the

pre-processing stage, the tweets' case was modified,

appropriate spaces were inserted, unnecessary spaces

were deleted. Making use of both unigrams and

bigrams, the features for the model were retrieved.

The process of lemmatization and removal of stop

words was thereafter carried out. Their recommended

technique was assessed for its efficacy using three

prominent machine learning classifiers: Naive Bayes,

Logistic Regression, and Support Vector Machine.

(Yadav, Kudale, Rao, Gupta, & Shitole, 2021).

Twitter has shown to be beneficial in the realm of

disaster management. Shah et al. (2018) conducted a

comprehensive analysis of Twitter data to examine

the Nepal Earthquake and identify several aspects

related to the catastrophe. Using Twitter data

acquired during the final week of April and the first

week of May 2015, a total of 40,236 unprocessed

tweets that appeared to be relevant to the Nepal

earthquake were obtained. These messages were then

pre-processed for analysis and subsequently

analysed. This study demonstrates the utilisation of

geolocation tag to identify hazardous areas and using

visual analytics to analyse the dataset. The use of

automated keyword identification led to the

development of a disaster management module. This

module is capable of identifying keywords associated

with any particular catastrophe, enabling further

investigation. The findings indicate that the disaster

module used in the research may effectively operate

on various hashtags without requiring manual

parameter definition, as long as the dataset specific to

the given situation is accessible. (Shah, Agarwal,

Dubey, & Correia, 2018).

The results of recent research have shown that

tweets may be used to make predictions about a wide

range of significant events, such as elections and

national revolutions as well as criminal activity. This

is the primary concept, which states that the context,

timing, and content of tweets can give insight into

what will occur in the future. Whether local criminal

activities can be predicted based on the tweets sent

was the question that was addressed in a report that

was published by the University of Virginia in

November of 2014. According to the findings of the

study, the inclusion of information from Twitter

improves the accuracy of prediction for 19 out of 25

different types of criminal activity, and it does so

considerably for a number of different surveillance

scopes. (Gerber, 2014).

When it comes to twitter data, a lot of studies have

been conducted, but most of them pertain to sentiment

analysis regarding a particular topic. There exists a

substantial gap in the research when it comes to

influencers on Twitter.

In a study that was carried out by Briliani et al.,

the researchers looked for instances of hate speech

that were found in the comments area of an Instagram

post. Responses can be either good or negative,

depending on the context. The use of hate speech is

included in the negative comments on Instagram.

Speech that promotes hatred is one of the most

significant issues, and it is extremely hard for

authorities to combat. In light of this, the K-Nearest

Neighbour classification approach was utilised in this

research project to develop a system that was capable

of determining whether or not one was engaging in

hate speech in the Instagram comment area. The

results of this study have produced an accuracy of

98.13 percent, as well as precision, recall, and F1-

score of 98 percent when employing K-Nearest

Neighbour with K equal to three. (Briliani, Irawan, &

Setianingsih, 2019).

Ekosputra et al. in their 2021 study made use of

Supervised Machine Learning algorithms to detect

fake accounts on Instagram. Logistic Regression,

Bernoulli Naive Bayes, Random Forest, Support

Vector Machine, and Artificial Neural Network

(ANN) are the techniques that were utilised in the

process of developing the supervised machine

learning model. In this study, two tests were

conducted. In the first test, the model is in its default

state, which means that it does not have any

parameters and no features are introduced.

Furthermore, in order to enhance the precision of the

experiment, new features and tuning factors were

incorporated into the process in the second test.

Logistic Regression and Random Forest, both of

multiBERT: A Classiﬁer for Sponsored Social Media Content

707

which have an accuracy of 0.93, are the models that

perform better than other models based on the second

experiment with additional variables and parameters

thanks to their superior performance. (Ekosputra,

Susanto, Haryanto, & Suhartono, 2021).

M. Singh presented an alternative method for

identifying fraudulent accounts on Instagram in

research done in 2023. One form of malicious

behaviour on the Instagram platform is the creation

and use of counterfeit accounts. This study employs a

hybrid technique that takes into consideration both

the content of the post and the photographs to identify

phoney accounts on Instagram. The author assessed

the presence of text spam using machine learning

models such as Random Forest classification and

identified picture spam using CNN models. The

picture dataset was sourced from picture Spam

Hunter, while the model was trained using a Kaggle

dataset to categorise images based on their content.

The suggested hybrid model has also undergone

testing using the dataset obtained through web

scraping from Instagram. The experimental

classification results demonstrate that the suggested

model achieves a classification accuracy of 97.1%.

(Singh, 2023).

When it comes to BERT, there are a lot of studies

that have been published.

M.T. Riaz et al. published a study in 2022 which

introduced TM-BERT or twitter modified BERT for

COVID 19 vaccination sentiment analysis. Within the

scope of this research, a Twitter Modified BERT

(TM-BERT) that is based on Transformer

architecture is shown. Additionally, a new Covid-19

Vaccination Sentiment Analysis Task (CV-SAT) and

a COVID-19 unsupervised pre-training dataset

consisting of 70,000 tweets have been produced by

this group. After being fine-tuned on CV-SAT, BERT

attained an accuracy of 0.70 and 0.76, however TM-

BERT achieved an accuracy of 0.89, which is a 19%

and 13% improvement over BERT respectively.

(Riaz, Shah Jahan, Khawaja, Shaukat, & Zeb, 2022).

The application of BERT for the detection of

cyberbullying in the digital age is discussed by Yadav

et al. in their article that was released in the year 2020.

Using a novel pre-trained BERT model with a single

linear neural network layer on top as a classifier, a

new strategy is suggested to the identification of

cyberbullying in social media platforms. This

approach is an improvement over the results that have

been obtained previously. During the training and

evaluation process, the model is trained on two

different social media datasets, one of which is very

small in size, and the other of which is fairly large in

size. (Yadav, Kumar, & Chauhan, 2020).

Software vulnerabilities pose a significant risk to

the security of computer systems, and there has been

a recent increase in the discovery and disclosure of

these weaknesses. Ni et al. did a study in which they

introduced a novel approach called BERT-CNN. This

approach combines the specialised task layer of Bert

with CNN to effectively collect crucial contextual

information in the text. Initially, a BERT model is

employed to analyse the vulnerability description and

other data, such as Access Gained, Attack Origin, and

Authentication Required, in order to provide the

feature vectors. Subsequently, the feature vectors

representing vulnerabilities together with their

corresponding severity levels are fed into a

Convolutional Neural Network (CNN), from which

the CNN parameters are obtained. Subsequently, the

fine-tuned Bert model and the trained CNN model are

employed to predict the degree of severity associated

with a vulnerability. This method has demonstrated

superior performance compared to the current leading

method, with an F1-score of 91.31%. (Ni, Zheng,

Guo, Jin, & Li, 2022).

Guo et al. did a study in 2022 focusing on

developing methods to detect false news. Current

suggested methods for false news identification in

centralised platforms do not consider the location of

news announcements, but rather prioritise the

analysis of news content. This study presents a

distributed architecture for detecting false news based

on regions. The framework is used inside a mobile

crowdsensing (MCS) setting, where a group of

workers are chosen to collect news depending on their

availability in a particular location. The chosen

workers disseminate the news to the closest edge

node, where the local execution of pre-processing and

detection of counterfeit news takes place. The

detection technique used a pre-trained BERT model,

which attained a 91% accuracy rate. (Guo, Lamaazi,

& Mizouni, 2022).

Text categorization has consistently been a

significant undertaking in the field of natural

language processing. Text categorization has become

extensively utilised in several domains such as

emotion analysis, intention identification, and

intelligent question answering in recent years. In a

2021 publication, Y. Cui et al. introduced a novel

methodology. This study used the Bert model to

produce word vectors. The text characteristics

collected by a Convolutional Neural Network (CNN)

were then combined to get more efficient features,

enabling the completion of Chinese text

classification. Experiments were performed using a

publicly available dataset. Recent studies have

demonstrated that the Bert+CNN model outperforms

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

708

other text classification models in properly

categorising Chinese text, mitigating overfitting, and

exhibiting strong generalisation capabilities. (Cui &

Huang, 2021).

A lot of studies have been conducted on the

application of BERT and also on the data collected

from social media sites such as Twitter and

Instagram. However, there exists a significant gap in

research when it comes to influencer behaviour on

such sites. While there have been studies regarding

the authenticity of accounts on such sites, detecting

the authenticity of the content posted on such sites is

also of equal importance.

3 METHODOLOGY

The general research methodology is outlined in

figure 1. The data pre-processing step will be

completed first. The data has been cleansed at this

point. Once the data pre-processing step is completed,

the next step is to determine the classification model

that will be applied. To determine which model

performs best, a classification evaluation will be done

once each model has been trained. The methodology

is visualized in Figure 1.

Figure 1: Methodology flow diagram.

3.1 Dataset

For the purpose of this study, real world user data was

collected from Instagram and Twitter. The data was

collected using the respective APIs provided by

Twitter and Instagram. Twitter API allows us to

extract either Tweets pertaining to a hashtag ("#”)

passed as keywords into the API or user information

as data. All the tweets pertained to “#beauty” and

“#gaming”. The Instagram API allows us to extract

post captions along with URL links to the photo itself.

However, for the purpose of this study, only captions

related to the hashtags mentioned above were

collected. The data was divided into 4 datasets.

Dataset 1 (D1) comprised of 1000 samples of Twitter

data, Dataset 2 (D2) comprised of 80000 samples of

Twitter Data, Dataset 3 (D3) comprised of 1000

samples of Instagram Data and Dataset 4 (D4)

comprised of 80000 samples of Instagram data. 2 test

sets were also created, Test set 1 (T1) comprised of

200 samples of Instagram and Twitter data and Test

set 2 (T2) which comprised of 20000 samples of

Instagram and Twitter data combined. The

tweets/Instagram captions were manually annotated

based on the “sponsored” and “Paid Partnership

With” tags available in Twitter and Instagram

respectively. The details of the datasets are detailed in

Table I.

Table 1: Datasets.

Dataset Sample size Source

D1 1000 Twitter

D2 80000 Twitter

D3 1000 Instagram

D4 80000 Instagram

T1 200 Twitter + Instagram

T2 20000 Twitter + Instagram

As it can be observed in Table 1, the sample size for

training was increased from 1000 samples to 80000

samples. This was done to measure the performance

of the model when limited data is available and also

when there is no limit on the data.

3.2 Data Preprocessing

The significance of the data pre-processing step lies

in the fact that it has an impact on the efficiency of

the future phases. Altering the syntax of tweets,

removing information that is not essential from the

text, and identifying any additional elements that are

helpful are all included in this process. Python regular

expressions were utilised in order to exclude special

characters, emoji symbols, hashtags, and links from

the text in order to accomplish the objectives of this

study. It was decided not to remove the user handles

and handles that were referenced in the tweets since

doing so would be necessary in order to establish the

relationships between influencers in the later phases

of this experiment.

Preprocessing steps followed:

• Removal of hashtags

• Removal of links

• Removal of emoji

• Removal of special characters

• Removing extra spaces

3.3 Models

BERT (Bidirectional Encoder Representations from

Transformers) is a publicly available machine

multiBERT: A Classiﬁer for Sponsored Social Media Content

709

learning framework designed for the purpose of

natural language processing (NLP). BERT is

specifically engineered to enhance computers'

comprehension of the semantic nuances in ambiguous

textual language by using the surrounding text to

construct a comprehensive context. The BERT

framework underwent pre-training using textual data

sourced from Wikipedia and may thereafter be

refined through the use of question-and-answer

datasets. reference

BERT is a deep learning model that utilises

Transformers. In this model, each output element is

linked to every input element, and the weightings

between them are dynamically computed depending

on their relationship. This technique is referred to as

attention in the field of natural language processing

(NLP).

In the past, language models were limited to

reading text input in a sequential manner, either from

left to right or from right to left but were unable to do

both simultaneously. BERT stands out due to its

unique ability to do bidirectional reading

simultaneously. The capacity to process information

in both forward and backward directions, made

possible by the use of Transformers, is referred to as

bidirectionality.

BERT is trained bidirectionally, meaning it is trained

on two distinct yet interconnected NLP tasks: Masked

Language Modelling and Next Sentence Prediction.

The primary goal of Masked Language Model

(MLM) training is to obfuscate a word within a phrase

and thereafter enable the program to forecast the

concealed word (masked) by leveraging the

contextual cues of the hidden word. The goal of Next

Sentence Prediction training is to enable the program

to accurately determine whether two provided

phrases exhibit a coherent, sequential link or whether

their relationship is just arbitrary.

There are two primary variants of pre-trained BERT

models, which differ based on the magnitude of their

architectural design.

• The BERT-Base model consists of 12 layers,

each with 768 hidden nodes and 12 attention

heads. In all, it includes 110 million training

parameters.

• The BERT-Large model consists of 24 layers,

1024 hidden nodes, 16 attention heads, and a

total of 340 million training parameters.

The BERT-Base model was utilised for this

project as it is relatively easier on the system as

compared to BERT large model while keeping the

architectural intricacies.

3.3.1 Proposed Model

To incorporate the use of multiple datasets in one

model, an approach based on the BERT architecture

is proposed. The architecture is outlined in Figure 2.

Figure 2: multiBERT.

Two BERT models, one to be trained on twitter data

(BERT1) and one to be trained on Instagram data

(BERT2), were paired together, with a fully

connected layer and a 2-class linear layer on top.

To set the appropriate hyperparameters for tuning,

a simple grid search strategy was utilized. The search

space utilized was the search space recommended by

the authors of BERT (Devlin, Chang, Lee, &

Toutanova, 2019).

Parameter search space:

• Num_epochs: [2,3,4,5]

• Learning_rate: [2e-5, 3e-5, 5e-5]

• Batch_size: [16, 32]

• Weight_decay: [0, 0.1, 0.3]

A total of 72 trials were run, with each combination

of the parameters for both BERT1 and BERT2. The

best performing combination of parameters was

selected for the model.

The hyper parameters set for BERT1 were:

• Epochs: 5

• Batch size: 16

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

710

• Learning rate: 3e-5

• Weight decay: 0.1

The hyper parameters set for BERT2 were:

• Epochs: 5

• Batch size: 16

• Learning rate: 2e-5

• Weight decay: 0.1

3.4 Experiments

A total of 6 experiments were conducted as part of

this study. The BERT model was trained and tested

separately on both Twitter and Instagram data to

evaluate the results. The combined model was also

trained and tested along with the BERT models. The

first rounds of experiments were conducted with 1000

samples in the training set followed by the second

round of experiments, where the training sets were

populated with 80000 samples. The details of the

experiments are given below:

• Experiment 1: BERT

trained on Twitter Data

(1000 Samples)

• Experiment 2: BERT

trained on Twitter Data

(1000 Samples)

• Experiment 3: BERT

trained on Instagram

data (1000 Samples)

• Experiment 4: BERT

trained on Instagram

Data (80000 Samples)

• Experiment 5: multiBERT (1000 Samples)

• Experiment 6: multiBERT (80000 Samples)

4 RESULTS

The objective of this study was to evaluate whether

BERT-base and the proposed model multiBERT can

be effectively used to classify user tweets and

Instagram caption as sponsored or not sponsored. The

results are detailed in Table 2.

Table 2 details the results of Experiment 1. While

conducting training with 1000 samples, the model

BERT

achieved an accuracy of 73.5%. The model

had a precision score of 74.6% and a recall score of

73.5%, with the F1 score being 73.02.

Results observed for Experiment 2 are detailed in

Table 2. While training with 80000 records on the

twitter dataset, the model BERT

was able to achieve

an accuracy of 79.5%, showing a 6% increase in

performance due to additional training. The precision

score of the model went up by 4.6% to achieve 79.2%

precision and the recall score increased by 6%, to

achieve a recall score of 79.5%. The model has a F1

score of 79.3%.

Table 2: Experimental Results.

Model Accuracy

Precision Recall

Score

BERT

(1000 samples)

73.5% 74.6% 73.5% 73.02%

BERT

(80000 samples)

79.5% 79.2% 79.5% 79.3%

BERT

(1000 samples)

78.3% 78.9% 78.5% 78.7%

BERT

(80000 samples)

84% 85.05% 84% 84.5

multiBERT

(1000 samples)

82% 83.1% 82% 82.5%

multiBERT

(80000 samples)

89% 90.1% 89.3% 89.7%

Results observed for Experiment 3 are detailed in

Table 2. While conducting training with 1000

samples, it is observed that BERT

achieves an

accuracy of 78.3%. The model achieved a precision

score of 78.9% and a recall score of 78.5%, with an

F1 score of 78.7%. As compared to the BERT model

trained on twitter data, the model trained on

Instagram data is able to achieve a 6.5% increase in

performance at the same task and same number of

training samples.

Results observed from Experiment 4 are detailed

in Table 2. While conducting training with 80000

samples, it is observed that in this experiment, the

BERT

achieves an accuracy of 84%, showing an 4%

increase in performance with additional training. The

model achieved a precision score of 85.05% and a

recall score of 84%, getting a F1 score of 84.5%. As

seen with the previous experiment, even with

additional training, the model trained on Instagram

data with 80000 records achieves an accuracy

improvement of 4.5% over the model trained on

twitter data.

Results observed in Experiment 5 are detailed in

Table 2. While training on 1000 samples, multiBERT

was able to achieve an accuracy of 82% which is an

8.5% increase over the BERT

model trained on the

same number of samples in Experiment 1 and a 2.1%

increase over the BERT

model trained on the same

number of samples in Experiment III. The model was

also able to achieve a precision score of 83.1% and a

recall score of 82%, with a F1 score of 82.5%

Results observed in Experiment 6 are detailed in

Table 2. While training on 80000 samples,

multiBERT was able to achieve an accuracy of 89%

which is an 9.5% increase over the BERT

model

trained on the same number of samples in Experiment

2 and a 5% increase over the BERT

model trained on

the same number of samples in Experiment 4. The

model was also able to achieve a precision score of

90.1% and a recall score of 89.3%, with a F1 score of

89.7%

multiBERT: A Classiﬁer for Sponsored Social Media Content

711

Figure 3: Training samples used & accuracy score of the

models.

Figure 3 shows the results of all the experiments put

together in terms of the accuracy of the models. It can

be observed that the proposed model multiBERT

achieves better performance while being trained on

the same amount of data.

5 CONCLUSIONS AND FUTURE

WORK

In this study, both BERT-Base as well as the

proposed model, multiBERT were correctly able to

classify Influencer posts as sponsored or not. With

regards to performance, it was found that the new

proposed model was the best in terms of accuracy,

both with limited data as well as a large sample size.

The multiBERT model achieved an accuracy of 82%

when trained on 1000 samples, which is a 3.7%

increase over the BERT-Base model trained on

Instagram data (BERT

) and an 8.5% increase over

the BERT-Base model trained on Twitter (BERT

The performance increase is slightly more significant

when the multiBERT model is trained on 80000

samples. The multiBERT model trained on 80000

samples achieved an accuracy of 89%, which is a 5%

increase over the BERT-Base model trained on

Instagram data and 9.5% increase over the model

trained on Twitter data. An interesting observation

here would be that the models trained with Instagram

data outperformed the models trained with Twitter

data consistently. When trained on 1000 samples, the

BERT

achieved

an accuracy of 78.3% which is an

increase of 4.8% and when trained on 80000 samples,

BERT

achieved an accuracy of 84% which was a

4.5% increase over the BERT

model. This inferior

performance of the models trained on Twitter data

may be attributed to the use of more slang words or

the increased use of abbreviations as compared to

Instagram.

Transformer models, like BERT, have previously

been proven to be extremely effective and applicable

in a wide range of machine learning applications. The

findings of this study demonstrate that the proposed

model, multiBERT, is capable of effectively

categorizing tweets or Instagram posts as either

sponsored or non-sponsored. Future work will focus

on correlating influencer data with real world sales

data using knowledge graphs. These graphs would

effectively illustrate the evolving dynamics between

influencers and ordinary users, in order to determine

the effect that these influencers have on social media

users and their behaviour and spread of online trends.

Circana provides clients with data, industry insight,

and advanced analytics to enhance their

understanding of the retail sector. With access to

propriety real world sales and POS data provided by

Circana, future work will focus on implementing a

system that can detect and quantify the effects of such

influencers and the impact they can have on the

purchase decisions of their followers by correlating

influencer behaviour with real world sales data. Work

will also focus on developing a system in tandem

which would identify anomalous behaviour of the

influencers on social media.

REFERENCES

Briliani, A., Irawan, B., & Setianingsih, C. (2019). Hate

Speech Detection in Indonesian Language on Instagram

Comment Section Using K-Nearest Neighbor

Classification Method,. IEEE International Conference

on Internet of Things and Intelligence System (IoTaIS),

(pp. 98-104). Bali.

Cui, Y., & Huang, C. (2021). A Chinese Text Classification

Method Based on BERT and Convolutional Neural

Network. 7th International Conference on Systems and

Informatics (ICSAI) (pp. 1-6). Chongqing: IEEE.

Devlin, J., Chang, M., Lee, K., & Toutanova, K. (2019).

BERT: Pre-training of Deep Bidirectional

Transformers for Language Understanding. . North

American Chapter of the Association for

Computational Linguistics.

Ekosputra, M. J., Susanto, A., Haryanto, F., & Suhartono,

D. (2021). Supervised Machine Learning Algorithms to

Detect Instagram Fake Accounts. 4th International

Seminar on Research of Information Technology and

Intelligent Systems (ISRITI) (pp. 396-400). Yogyakarta:

IEEE.

Gerber, M. S. (2014). Predicting crime using Twitter and

kernel dencity estimation. Decision Support Systems,

115-125.

Guo, Y., Lamaazi, H., & Mizouni, R. (2022). Smart Edge-

based Fake News Detection using Pre-trained BERT

Model. 18th International Conference on Wireless and

ICEIS 2024 - 26th International Conference on Enterprise Information Systems

712

Mobile Computing, Networking and Communications

(WiMob) (pp. 437-442). Thessaloniki: IEEE.

Machuca, C., Gallardo, C., & Toasa, R. (2021). Twitter

Sentiment analysis on coronavirus: Machiene Learning

Approach. Journal of Physics: Conference Series.

Ni, X., Zheng, J., Guo, Y., Jin, X., & Li, L. (2022).

Predicting severity of software vulnerability based on

BERT-CNN. International Conference on Computer

Engineering and Artificial Intelligence (ICCEAI) (pp.

711-715). Shijiazhuang: IEEE.

Riaz, M. T., Shah Jahan, M., Khawaja, S. G., Shaukat, A.,

& Zeb, J. (2022). TM-BERT: A Twitter Modified

BERT for Sentiment Analysis on Covid-19 Vaccination

Tweets. 2nd International Conference on Digital

Futures and Transformative Technologies (ICoDT2),

(pp. 1-6). Rawalpindi.

Shah, B., Agarwal, V., Dubey, U., & Correia, S. (2018).

Twitter Analysis for Disaster Management. Fourth

International Conference on Computing

Communication Control and Automation (ICCUBEA)

(pp. 1-4). Pune: IEEE.

Singh, M. (2023). Advanced Machine Learning Model to

Detect Spam on Instagram. 2023 IEEE International

Conference on Blockchain and Distributed Systems

Security (ICBDS) (pp. 1-6). New Raipur: IEEE.

Yadav, J., Kumar, D., & Chauhan, D. (2020).

Cyberbullying Detection using Pre-Trained BERT

Model. International Conference on Electronics and

Sustainable Communication Systems (ICESC), (pp.

1096-1100). Coimbatore.

Yadav, N., Kudale, O., Rao, A., Gupta, S., & Shitole, A.

(2021). Twitter sentiment analysis using supervised

Machine Learning. In Intelligent Data Communication

Technologies and Internet of Things (pp. 631-642).

Zhu, J., Xia, Y., Wu, L., He, D., Qin, T., Zhou, W., . . . Liu,

T.-Y. (2020). Incorporating BERT into Neural Machine

Translation. International Conference on Learning

Representations.

multiBERT: A Classiﬁer for Sponsored Social Media Content

713