AI‑Powered Fake Review Detection for College Admission Using

BERT and DeBERT

Savitha P., Christopher R., Dinesh M. and Jeevabharathi A.

Department of Computer Science and Engineering, Nandha Engineering College, Erode, Tamil Nadu, India

Keywords: Fake Review Detection, AI System, College Admissions, NLP, Machine Learning, BERT, DeBERTa, XLNet,

Sentiment Analysis, Review Classification, Web Plugin, Real‑Time Analysis, Supervised Learning,

Reputation Management, Decision‑Making Support.

Abstract: Online audits plays a crucial role in influencing college admission decisions. However, the presence of fake

reviews, whether excessively positive or misleading negative can distort perceptions and misleading

prospective students. This model proposes an AI-powered fake audit discovery framework outlined

particularly for college affirmation audits. The framework leverages Common Dialect Preparing (NLP)

procedures and directed learning models to distinguish beguiling substances. A web plugin is created to

analyze surveys in real-time, with the dataset collected from different college audit stages and preprocessed

utilizing methods like stop word expulsion, lemmatization, and estimation investigation. Progressed models

counting BERT, DeBERTa, and XLNet are prepared to classify surveys as either honest to goodness or fake,

with DeBERT a conveying the most noteworthy exactness. This framework progresses straight forwardness

within the college choice preparation, enabling students and guardians to form educated choices based on

bonafide surveys, whereas too helping educate in keeping up their notoriety by sifting out deceiving or false

surveys.

1 INTRODUCTION

The expanding dependence on online surveys plays a

significant part in forming students' choices with

respect to college affirmations. In any case, fake

reviews whether too positive or misleadingly

negative can make mutilated discernments and delude

candidates. Conventional strategies for recognizing

fake audits are regularly moderate, error-prone, and

not adaptable (Rayana, S., & Akoglu, L. 2015). With

the developing request for dependable surveys,

computerized arrangements for fake audit locations

are getting to be basic. Machine learning models,

especially those utilizing Normal Dialect Preparing

(NLP), are a capable apparatus for distinguishing fake

surveys by analyzing phonetic designs and opinion

(Asaad, W. H et al., 2023). In this venture, we utilize

NLP methods, such as BERT, DeBERTa, and XLNet,

to classify surveys from different college survey

stages (Liu, M., & Poesio, M. 2023). The extension is

executed as a full- stack web application utilizing

Streamlit, permitting real-time survey examination.

The proposed framework consolidates Reasonable AI

(XAI) to supply clients with straightforward and

interpretable approaches, making a difference

between understudies and guardians making educated

choices. By centering on exactness, interpretability,

and client involvement, this extends points to upgrade

the college determination handle, guaranteeing that

clients can distinguish true surveys and maintain a

strategic distance from deluding data (Wu et al.,

2020).

2 RELATED WORKS

Dummy audit, to this end, has been a crucial question

of scope, with a focus on normal ml and NLP

(Custom Dialect Preparing) strategies (Muhawesh, R

et al., 2021). Traditional models (for example,

Gullible Bayes, Back Vector Machines (SVM) and

Arbitrary Timberlands), have been related to fake

survey areas, with some achievements in

distinguishing misleading substances (Asaad, W. H et

al., 2023). Wang et al. (2020) researched about how

assumption investigation utilizing SVM and ML

models can help in classifying fake audits.

Nonetheless, such techniques often struggle with real

244

P., S., R., C., M., D. and A., J.

AI-Powered Fake Review Detection for College Admission Using BERT and DeBERT.

DOI: 10.5220/0013880800004919

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

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 2, pages

244-248

ISBN: 978-989-758-777-1

time analysis and processing of large datasets

efficiently (Ren et al., 2016). These last few years

have witnessed a surge of transformer-based models

like BERT, DeBERTa, XLNet which have

significantly improved fake audit detection (Liu, M.,

& Poesio, M. 2023). These models make them

sensible for fetch up comprehension tasks, despite the

fact that they exceed desires in comprehending at the

scene and associations in the substance. Such ideas as

those of Kundu et al. (2021) it shows that these

models outshine traditional methods such as SVM

and logistic regression in identifying fake reviews

with greater accuracy (Wang, B., & Kuan, K. K. Y.

2022).

We advance execution through BERT

(Bidirectional Encoder Representations from

Transformers) and DeBERTa (Decoding-enhanced

BERT) by recognizing the subtleties of dialect and

sentence structures (Liu, M., & Poesio, M. 2023).

XLNet overcomes them by better managing context

in text. Transformer models have been particularly

fascinating to process for tasks that require deep

content comprehension, such as determining whether

audits are genuine or fraud (He, S., Hollenbeck, B.,

Proserpio, D., & Thies,A. 2022). This work expands

on these advances using BERT, DeBERTa, and

XLNet to classify college confirmations surveys, with

an application utilizing Streamlit to provide students

with real-time predictions to help them make

educated decisions (He, S., Hollenbeck, B.,

Proserpio, D., & Tosyali, A. 2022).

3 METHODOLOGY

3.1 Data Collection and Pre-Planning

For this consideration, we utilized datasets sourced

from trusted online review stages, tallying college

overview websites, understudy social occasions, and

social media dialogs (Li, H et al., 2015). The dataset

comprises printed overviews in conjunction with

metadata such as timestamps, commentator

information, and assessments. Each review segment

talks to a supposition around a college, containing

critical properties that offer help recognizing fake or

misdirecting reviews (Muhawesh et al., 2021).

Additionally, the dataset includes sentiment scores

and linguistic patterns that help in detecting biased or

spam reviews.

The metadata also captures reviewer engagement

levels, such as the number of reviews posted and

interaction history, which aids in credibility

assessment. Furthermore, natural language

processing

(NLP)

techniques

are

employed

analyze contextual cues and detect inconsistencies

within the reviews.

3.2 Data Preprocessing

To guarantee information quality and move forward

show execution, the taking after preprocessing steps

were connected:

• Handling Missing Data: Lost values within the

dataset were recognized and ascribed utilizing

procedures like supplanting lost values with the

mode or cruel (for numerical metadata) (Ren et

al., 2016).

• Text Cleaning: The content surveys

experienced preprocessing steps, counting:

• Changing over content to lowercase for

consistency (He, S., Hollenbeck, B., Proserpio,

D., & Thies,A. 2022).

• Evacuating extraordinary characters, stop words,

and intemperate accentuation (Wu et al., 2020).

• Tokenization and stemming/lemmatization to

normalize words (Liu, M., & Poesio, M. 2023).

• Feature Scaling: Numerical metadata

highlights (such as audit length, word check,

and time-based qualities) were normalized to

make strides show joining (Rayana, S., &

Akoglu, L. 2015).

3.3 Data Split

The data file was separated within coordinating,

invigorate, and test groups within an 80-10-10 degree

to guarantee an adjusted representation (11. He,

S., Hollenbeck et al., 2022).

Figure 1 shows the

Dataset Split for Fake Review Detection.

Figure 1: Dataset Split for Fake Review Detection.

AI-Powered Fake Review Detection for College Admission Using BERT and DeBERT

245

3.4 Feature Extraction and Selection

Highlight extraction and assurance play an imperative

portion in recognizing fake and veritable reviews.

The key highlights were chosen based on space data

and quantifiable examination, emphasizing qualities

that earnestly relate with overview validity (Wang,

B., & Kuan, K. K. Y. 2022). Additionally, machine

learning algorithms are employed to classify reviews

by analyzing linguistic patterns and reviewer

behavior (Rayana, S., & Akoglu, L. 2015).

3.5 Feature Engineering

Linguistic Features: Conclusion score (limit and

subjectivity), (Li, H et al., 2015) Lucidness and

substance complexity (e.g., Flesch Scrutinizing Ease

score) Overview length and word repeat scattering.

Metadata-Based Features: Commentator validity

(account age, survey recurrence, etc.) (He, S.,

Hollenbeck et al., 2022) Review posting designs (e.g.,

sudden spikes in comparative reviews) Rating makes

irregularity (e.g., excessively positive/negative audits

with negating content).

3.6 Machine Learning Algorithms

A rule-based approach was to begin with to set up and

arrange criteria for recognizing fake reviews in

college declarations (Muhawesh et al., 2021). This

included recognizing common phonetic plans,

estimation peculiarities, and emphasized expressions

that are as regularly as conceivable found in boggling

reviews (Asaad et al., 2023) Following this, we

orchestrated progressed transformer-based models to

create a solid classification system (Li, H et al., 2015).

We endeavored with models tallying BERT

(Bidirectional Encoder Representations from

Transformers) and DeBERTa (Decoding-enhanced

BERT with Unraveled Thought) (Wu et al., 2020).

These models were chosen for their ampleness in

analyzing printed data, capturing apropos affiliations,

and understanding essential tongue collections that

restricted legitimate to goodness to goodness and fake

overviews (Rayana, S., & Akoglu, L. 2015).

3.7 Model Evaluation

The prepared models were then evaluated on the test

set to quantify the effectiveness of the application to

wiederlect against fake surveys in college admissions

(He, S., Hollenbeck et al., 2022). Assessment and

other measurements including precision, F1 score,

exactness, review, and Region Beneath the Bend

(AUC) can be used to evaluate the model’s prescient

execution and unwavering quality (Liu, M., &

Poesio, M. 2023). DeBERTa showed the finest

performance among various approaches through

contrast stop and was chosen as the best fine-tuned

model (Wang et al., 2022). We chose the

demonstration with the most commendable

evaluation scores to incorporate into the framework

last, giving dependable identification of deluding

audits at the same time bringing down the wrong

positive and false negative rate (He, S., Hollenbeck et

al., 2022) Figure 2 shows the Data Processing

Workflow.

Figure 2: Data Processing Workflow.

4 EXPERIMENTAL RESULTS OF

AI: POWERED FAKE REVIEW

DETECTION FOR COLLEGE

ADMISSION USING BERT,

DEBERT ALGORITHMS

Evaluation metrics used show that the BERT and

DeBERTa models are able to label and classify fake

audits successfully in the context of college

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

246

admissions (Wu et al., 2020). Some of these models

use NLP (Normal Dialect Preparing) approaches to

analyze audit genuineness which shows high

precision and robustness (He, S., Hollenbeck et al.,

2022). The evaluation of the tests points out that the

models achieved high accuracy and recall scores Liu,

M., & Poesio, M. 2023), ensuring low false positive

and frame classified (Muhawesh et al., 2021).

Figure 3: Training Accuracy Curve.

The exploratory aims to post some light on the

performance comparison between BERT and

DeBERTa models in the detection of fake college

reviews. The models were evaluated using essential

execution estimates, numbering exactness, accuracy,

review, and F1-score. Figure 1 lists the setups on

which DeBERTa outperforms BERT on all metrics

(Rayana, S., & Akoglu, L. 2015), (Ren, Y., Ji, D., &

Ren, Y. 2016).

Figure 3 shows the Training Accuracy

Curve.

Accuracy: DeBERTa accomplished a precision of

0.92, outperforming BERT's 0.85, demonstrating

made strides by and large classification execution.

Precision: DeBERTa got an exactness score of 0.90,

compared to BERT's 0.83, illustrating its capacity to

decrease wrong positives viably.

Recall: DeBERTa beat BERT with an accuracy of

0.91 vs. 0.86, which demonstrated that DeBERTa

was better than BERT at capturing fake audits.

F1-Score: The DeBERTa062 achieves an F1-score

of 0.91 against BERT 0.84, ensuring a balanced

precise and recall trade-off.

This is done according to confirm that DeBERTa

could be a more successful portrayal for phony audit

area, giving improved exactness and better

generalization. Its predominant review and F1-

score make it more dependable for recognizing false

audits whereas keeping up exactness (Rayana, S., &

Akoglu, L. 2015).

Figure 4 shows the Performance

Comparison: DeBERT and BERT Algorithms. Figure

5 shows the Performance Comparison BERT vs

DeBERT Fake College Review Detection.

Figure 4: Performance Comparison: Debert and Bert

Algorithms.

Figure 5: Performance Comparison BERT vs DeBERT

Fake College Review Detection.

5 CONCLUSIONS

The AI-powered Fake Survey Location framework

for college confirmations gives an imaginative

arrangement to the developing concern of false audits

within the instructive segment. By utilizing cutting-

edge advances such as Common Dialect Preparing

(NLP) and Machine Learning (ML), the framework is

competent at precisely analyzing and classifying

surveys in real-time, guaranteeing the keenness of

online audits. Built as a web application utilizing

Streamlit, the framework presents a user- friendly

interface, making it simple for both specialized and

non-technical clients to get to and navigate.The center

usefulness of the framework incorporates

recognizing fake surveys

based on different

etymological designs, irregularities, and suspicious

behaviors found within the content. Because it works

AI-Powered Fake Review Detection for College Admission Using BERT and DeBERT

247

in real-time, it can swiftly hail audits that are likely to

be false, empowering teach to provoke activity. This

arrangement can be easily integrated into college

affirmation stages, progressing the

straightforwardness of the audit preparation and

empowering understudies to form more educated

choices based on true input.

6 FUTURE WORK

Within the future, the framework will be extended to

cover an assortment of segments, counting e-

commerce, neighborliness, and healthcare, permitting

it to distinguish fake surveys past college

confirmations. This will help in progress and

straightforwardness over different stages where client

input plays a basic role. To enhance the system's

precision, we'll center on refining the machine

learning calculations and investigating progressed

NLP procedures. This will empower the framework

to better recognize unobtrusive false designs and

adjust to advancing strategies utilized by fake

reviewers. Additionally, the framework will be

upgraded to analyze numerous dialects for fake

audits, broadening its worldwide appropriateness. We

too arrange to coordinate a confirmed audit database

and present real-time location, permitting the

framework to hail fake audits instantly upon

accommodation, guaranteeing the judgment of survey

stages.

REFERENCES

Alshehri, A. H. (2023). "An Online Fake Review Detection

Approach Using Famous Machine Learning

Algorithms." Computers, Materials & Continua, 78(2),

2767-2786.

Asaad, W. H., Allami, R., & Ali, Y. H. (2023)."Fake

Review Detection Using Machine Learning. "Revued'

Intelligence Artificielle, 37(5), 1159-1166.

He, S., Hollenbeck, B., Proserpio, D., & Thies,A. (2022).

"Detecting Fake-Review Buyers Using Network

Structure." Proceedings of the National AcademyofSci

ences,119(47), e2211932119. (CVPR), 2017, pp. 7263–

7271.

He, S., Hollenbeck, B., Proserpio, D., & Tosyali, A. (2022).

"Detecting Fake-Review Buyers Using Network

Structure." Proceedings of the National Academy of

Sciences, 119(47), e2211932119. DOI:

10.1073/pnas.2211932119.

Li, H., Chen, Z., Mukherjee, A., Liu, B., & Shao, J. (2015).

"Analyzing and Detecting Opinion Spam on a Large-

Scale Dataset via Temporal and Spatial Patterns."

Proceedings of the Ninth International AAAI

Conference on Web and Social Media, pp. 634–637.

Liu, M., & Poesio, M. (2023). "Data Augmentation for Fake

Reviews Detection." Proceedings of Recent Advances

in Natural Language Processing, pp. 673–680.

DOI:10.26615/978-954-452-092-2_073.

Muhawesh, R., Xu, S., Tran, S. N., & Maqsood, S. (2021).

"Fake Reviews Detection: ASurvey." arXiv preprint

arXiv:2104. 08521.Engineering Access, 8(2), 192-197.

Rayana, S., & Akoglu, L. (2015). "Collective Opinion

Spam Detection: Bridging Review Networks and

Metadata." Proceedings of the 21th ACM SIGKDD

International Conference on Knowledge Discovery and

Data Mining, pp. 985– 994. DOI: 10.1145/2783258.27

83370

Ren, Y., Ji, D., & Ren, Y. (2016). "Detecting Online

Review Spam: A Linguistic Feature Engineering

Approach." Applied Intelligence, 45(3), 868–882. DOI:

10.1007/s10489-016-0794-5

Wang, B., & Kuan, K. K. Y. (2022)."Understanding the

Message and Formulation of Fake Online Reviews: A

Language-production Model Perspective. "Frontiers in

Artificial Intelligence.:2207.02696, 2022.

Wani, M. A., ElAffendi, M., & Shakil, K. A. (2024). "AI-

Generated Spam Review Detection Framework with

Deep Learning Algorithms and Natural Language

Processing. "Computers, 13(10), 264.vol.7, p.1419,

Oct.2016.DOI:10.3389/fpls.2016.01419.

Wu, Y., Ngai, E. W., Wu, P., & Wu, C. (2020)."Fake Online

Reviews: Literature Review, Synthesis, and Directions

for Future Research." Knowledge and Information

Systems, 62(6), 2133–2170.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

248