A Multilingual Explainable NLP and Deep Learning‑Based

Framework for Intelligent Plagiarism Detection and Academic

Content Validation

Dondeti Rammohanreddy

, P. Pradeep

, Oviyasri G. K.

, M. K. Kirubakaran

Allam Balaram

and Angel Jency V.

Department of CSE, Newton's Institute of Engineering, Andhra Pradesh, India

Department of Mechanical Engineering, Kumaraguru College of Technology, Coimbatore 641049, Tamil Nadu, India

Department of Management Studies, Nandha Engineering College, Vaikkalmedu, Erode - 638052, Tamil Nadu, India

Department of Artificial Intelligence and Data Science, St. Josephs Institute of Technology, Chennai‑600119, Tamil Nadu,

India

Department of Computer Science and Engineering, MLR Institute of Technology, Hyderabad‑500043, Telangana, India

Department of ECE, New Prince Shri Bhavani College of Engineering and Technology, Chennai, Tamil Nadu, India

Keywords: Plagiarism Detection, Deep Learning, Natural Language Processing, Explainable AI, Multilingual Analysis.

Abstract: For Academic research writing, plagiarism checking has moved from simple text matching to context based

matching through sophisticated natural language processing (NLP) and deep learning. This research presents

a multilingual, explainable, and scalable approach to intelligent plagiarism detection and content validation

effort for academic integrity. By combining BERT and XLM-R model with semantic similarity measurement,

the system can effectively detect paraphrased, cross-lingual and AI-generated plagiarism. The model, in

contrast to available systems, include citation context awareness, real-time response and domain-based

thresholds, which accounts for fairness and transparency in an evaluation. Explainable AI components such

as attention visualization and token-level attribution provide interpretability for students, teachers, and

reviewers. It also has the ability to detect code and figure plagiarism and it is appropriate for science,

technology, engineering and mathematics disciplines. Experimental results on benchmark and real world

academic datasets show higher accuracy, fewer false positives, and better cross-language and cross-content

type performance. This work is a first step towards the ethical, smart, and inclusive validation of academic

content.

1 INTRODUCTION

In the era of electronic communication, academic

writing and circulation of research has accelerated to

an unprecedented rate, allowing scholars access to a

wealth of knowledge. This rapid growth, however,

has also raised issues of intellectual property

infringement, especially plagiarism. Traditional

detection methods built on shallow text matching or

database search have difficulty dealing with advanced

plagiarism techniques, including paraphrasing,

cross-language translation, AI-based generation of

content, etc.

With the rise of sophisticated Natural Language

Processing (NLP) methods and deep learning

architecture, we can move beyond shallow similarity

comparison to deep semantic analysis. The

development of language models such as BERT,

RoBERTa, and XLM-R has made it possible for

systems to comprehend the context, intent, and subtle

linguistic differences, A â€ providing a fertile child

for intelligent plagiarism detection. And by including

elements of explainable AI, transparency is also

brought to a new high, letting educators and

institutions of all stripes make the right, less-biased

calls.

Despite advances, most current systems are

hindered by language coverage, black-box nature,

failure to generalize to obfuscated or synthesized text,

and poor interaction with practice tools. In this paper,

we present new multilingual and explainable

Rammohanreddy, D., Pradeep, P., K., O. G., Kirubakaran, M. K., Balaram, A. and V., A. J.

A Multilingual Explainable NLP and Deep Learning-Based Framework for Intelligent Plagiarism Detection and Academic Content Validation.

DOI: 10.5220/0013865700004919

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 1, pages

357-362

ISBN: 978-989-758-777-1

357

framework to overcome these issues. The use of

transformer-based deep learning with citation-

awareness, real-time semantic comparison, and visual

interpretability is beneficial not only in achieving

state-of-the-art detection accuracy, but also to gain

trust and ensure the usability in academic settings.

1.1 Problem Statement

With growing dependence of academic institutions

over digital content generation and usage the menace

of advanced plagiarism is more pervasive and

complicated. The traditional tools for detecting

plagiarism, which include those relying on keyword

searching, n-gram overlap, and rule-based

comparison, are not effective in detecting

semantically paraphrased, cross-lingual- or AI-

generated content. Such systems do not necessarily

take into account contextual similarity, the citation's

intent or are not designed to support domain-specific

and multilingual content. And they are of little to no

interpretability, so teachers and students do not know

how or why they are being flagged.

As large language models become increasingly

sophisticated and generative AI tools become more

popular, however, plagiarism has become much more

sophisticated than mere direct copying, spawning

different types of academic dishonesty that evade

legacy systems. The lack of explainability in

detection results and the inability to handle a wide

range of academic writing styles and languages add

up this challenge.

Such deep plagiarism detection was proposed in

response to a very important requirement for an

intelligent, scalable, and explainable framework that

identifies and adjudicates textual similarity at a deep

semantic level, while remaining multi-lingual, able to

be made aware of citations, code or figures and that is

able to validate code and figures. The aim is to narrow

the separation between NLP technology development

and academic integrity practical needs, to offer a fair,

rigorous and interpretable system to evidence

scholarly work.

2 LITERATURE SURVEY

Plagiarism identification has evolved rapidly in the

past decade, from heuristic-based string matching

algorithms towards intelligent systems based on NLP

and deep learning. Earlier studies were based on

lexical and syntactic overlap, which was however

found to be less effective for paraphrased or garbled

text. To address this, Wahle et al. (2021) also

proposed a benchmark to assess the performance of

neural language models in detecting paraphrased

plagiarism and found traditional detection methods

do not handle deep semantics very well.

The development of transformer based models

such as BERT and RoBERTa has allowed systems to

understand context, tone, and meaning at a more

granular level. Quidwai et al. (2023) used sentence-

level transformers to detect plagiarism with a better

accuracy than traditional methods. In a related work,

Shouman (2022) conducted deep learning models

identifying academic plagiarism, the models were

reliable but lacked scalability and had expensive

inference time.

In addition, mul- tilingual and cross-lingual

plagiarism detection has been focussed. Bakhteev and

Ivanov (2021) investigated strategies for translated

plagiarism detection and highlighted the necessity of

applying language-independent word embeddings.

Chang et al. (2024) extended such approach, and

proposed transformer-based semantic relation

extraction models to detect contextual plagiarism in

more than one language, but the models have the

limitations of using large amount of resources.

Explainability in plagiarism detection is now an

emerging imperative. Most of the time, traditional

systems would be a “black box” without any

explanation for why something gets identified as

abusive content. To bring a solution to this, Amzuloiu

and coauthor published a paper. (2021) suggested to

enhance Encoplot with deep learning to generate

interpretable predictions, however, their approach has

a high preprocessing overhead. They also noted that

it was difficult to detect AI-created content, and

suggested that a detection algorithm should be trained

on model outputs, not generative techniques in

general.

Many methods lacked citation-awareness and

context awareness. Wahle et al. (2021) stressed the

need to differentiate between correctly cited versus

plagiarized text in general and academic areas of

interest. The absence of this kind of awareness creates

false positives and negatives, which results in a trust

gap in institutional use.

A number of studies also explored non-textual

plagiarism coverage. Madhavan et al. (2023)

addressed the necessity for multimodal systems that

can spot plagiarised code and images, while Ahire et

al. (2021) Proposed the NLP-based mechanism to

detect the attempted plagiarism which considers the

structure and formatting of document.

However, the existing models remain

unsatisfactory in the performance in real-time, the

domain generalization ability and the interpretability.

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

358

Explainable AI was proposed to be incorporated by

Miller and Davis (2023) to enhance trust and utility.

In addition, Ravichandran and Kumar (2022)

emphasized the importance of fairness in detection

and that class imbalance and bias in training datasets

need to be taken into consideration.

These deficiencies motivate for the development

of a unified, multilingual, and explainable plagiarism

detection framework that is capable of accurately

detecting content reuse, while also reporting its

findings in a transparent and didactic way.

3 METHODOLOGY

The proposed multilingual NLP-based transformer-

deep-learning-based and explainable-AI supported

educational content to check for plagiarism model can

ensure a reliable, transparent, and context attended

educational content validation. First, a source corpus is

obtained from a variety of genres, such as academic corpus,

AI-generated texts and cross-lingual paraphrases.

They preprocess these documents including

tokenization, normalization, and citation extraction.

Multilingual support is ensured through language

detection and translation using MarianMT, while

semantic enrichment is performed using tools like

spaCy and NLTK. For representation, sentence and

paragraph embeddings are generated using XLM-

RoBERTa, Sentence-BERT, and SciBERT, capturing

deep semantic relationships across languages and

disciplines. Semantic similarity is calculated using

cosine and STS-based metrics at the sentence,

paragraph, and document levels. Dynamic

thresholding, based on Z-score normalization, adapts

sensitivity based on content length and structure. The

overall pipeline of the prosed system is illustrated in

Figure 1.

A specialized citation-aware module

distinguishes plagiarized text from properly

referenced material by classifying citation intent and

analyzing reference alignment. To enhance

transparency, the system integrates explainable AI

using SHAP values, attention heatmaps, and token-

level visualizations, allowing users to understand

which sections triggered plagiarism alerts. The model

also supports multimodal detection, identifying code

plagiarism using abstract syntax tree (AST)

comparison and figure/text duplication through OCR

and NLP caption matching. The overall system

architecture is shown in Figure 1.

Figure 1: Workflow of the Proposed Plagiarism Detection

Framework.

which outlines the sequential flow from data

preprocessing to semantic evaluation and

explainability."The entire framework is implemented

using Python and deployed via Docker containers,

optimized for real-time operation with GPU support,

and integrated into Learning Management Systems

(LMS) through REST APIs. As shown in Figure 2,

English dominates the dataset, followed by regional

and cross-lingual documents including Hindi,

Spanish, French, and Arabic this methodology

ensures comprehensive, interpretable, and scalable

plagiarism detection suitable for modern academic

environments. Table 1 gives the Dataset

Composition.

A Multilingual Explainable NLP and Deep Learning-Based Framework for Intelligent Plagiarism Detection and Academic Content

Validation

359

Table 1: Dataset Composition.

Dataset Source

Language(s)

Content Type

No. of

Documents

Plagiarized

(%)

PAN Plagiarism

Corpus

English

Academic

Papers

1,200

50%

Custom Student

Submissions

English,

Hindi, Spanish

Assignments

& Projects

800

40%

AI-Generated

Corpus

English

GPT/Bard

Outputs

500

100%

(Synthetic)

Multilingual

Academic

French,

Arabic

Journal

Articles

600

35%

Programming

Corpus

Code Snippets

Python, Java

300

45%

4 RESULT AND DISCUSSION

To evaluate the effectiveness of the proposed

plagiarism detection framework, extensive

experiments were conducted using a curated dataset

comprising academic articles, student assignments,

AI-generated content, and multilingual paraphrased

texts. Figure 4 highlights the F1-score comparison of

the proposed model with existing tools,

demonstrating a notable performance gain using

XLM-R and citation modules." The dataset included

both plagiarized and original documents in multiple

languages such as English, Spanish, Hindi, and

French. Evaluation metrics such as Precision, Recall,

F1-score, Semantic Similarity Score (SSS), and Area

Under the ROC Curve (AUC) were used to

benchmark the model against traditional plagiarism

detection systems and recent transformer-based

baselines. Table 2 gives the model comparison on

plagiarism detection accuracy. Table 3 gives the

detection performance on plagiarism types.

Table 2: Model Comparison on Plagiarism Detection Accuracy.

Model

Precision (%)

Recall (%)

F1-Score (%)

AUC

Score

Turnitin (Baseline)

78.1

73.4

75.6

0.81

CopyLeaks (Baseline)

81.3

77.1

79.1

0.84

Sentence-BERT

89.2

88.5

88.8

0.91

XLM-R + Citation

Module

94.3

92.6

93.4

0.96

Table 3: Detection Performance on Plagiarism Types.

Plagiarism Type

Detection

Accuracy (%)

False Positive

Rate (%)

Verbatim

Copying

97.2

1.1

Paraphrased Text

91.3

4.5

AI-Generated

Text

88.7

6.8

Cross-Lingual

Copying

85.4

5.2

Cited but

Improperly

Quoted

89.1

3.4

The results demonstrated that the proposed model

achieved a precision of 94.3%, recall of 92.6%, and

F1-score of 93.4%, outperforming classical tools like

Turnitin and open-source tools such as Moss and

CopyLeaks, especially in detecting paraphrased,

cross-lingual, and AI-generated plagiarism. The

incorporation of multilingual models such as XLM-

RoBERTa led to a notable improvement in cross-

language plagiarism detection accuracy, with a 12–

18% performance boost compared to English-only

models. As shown in Figure 5, the system performs

exceptionally well in detecting verbatim and

paraphrased plagiarism, with slightly lower accuracy

for cross-lingual and AI-generated text."

Furthermore, the citation-aware module significantly

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

360

reduced false positives by accurately identifying

properly cited content, which conventional tools

frequently misclassified as plagiarized. Table 4

shows the Impact of Citation-Aware Module.

Table 4: Impact of Citation-Aware Module.

Test

Group

Without

Citation

Module

With

Citation

Module

False

Positives

(%)

12.5

4.3

Overall

F1-Score

86.8

93.4

User

Satisfacti

71%

91%

A significant insight emerged in handling AI-

generated content. While traditional tools often failed

to flag generated text from models like ChatGPT or

Bard, the proposed framework, trained with a labeled

subset of generative outputs, successfully identified

88.7% of AI-generated plagiarism, thanks to stylistic

feature embeddings and semantic drift analysis. The

inclusion of SHAP explainability and attention

heatmaps allowed both educators and students to

clearly understand the reasons behind detection,

increasing trust in the tool's output. In user studies,

91% of faculty and students found the visual feedback

helpful in learning about proper citation practices and

avoiding unintentional plagiarism. Figure 6 offers an

interpretability view using SHAP, indicating which

tokens most influenced the plagiarism classification

decision

Multimodal support for detecting plagiarized code

and figures further distinguished the framework.

Using abstract syntax trees and embedding-based

code comparison, the model achieved an 89.2%

detection rate for reused or renamed code segments.

Figure duplication detection, based on caption

similarity and image text analysis via OCR, proved

highly effective in STEM disciplines, identifying

81.5% of reused visual data.

From a computational perspective, despite the

deep learning backbone, the model demonstrated

optimized performance using GPU acceleration and

quantization techniques. On a standard academic

server with 16GB RAM and a mid-range GPU, the

system processed a 15-page document in under 12

seconds, making it viable for real-time LMS

integration.

Overall, the study proves that a hybrid,

multilingual, and explainable plagiarism detection

system not only enhances detection accuracy but also

bridges the gap between machine intelligence and

ethical academic evaluation. By tackling limitations

in language, semantics, and interpretability, the

framework paves the way for the next generation of

intelligent academic validation systems that are fair,

inclusive, and pedagogically supportive. Evaluation

of Explainability Tools is tabulated in table 5.

Table 5: Evaluation of Explainability Tools.

Explainab

ility Tool

Interpretat

ion

Clarity

Average

User

Rating

(/5)

Trainin

Overhe

Attention

Heatmaps

High

4.7

Modera

SHAP

Very

High

4.8

High

Token

Attributio

n Layer

Moderate

4.2

Low

5 CONCLUSIONS

The increasing sophistication of academic plagiarism,

characteristic of multilingual authorship, AI-

produced content, and advanced paraphrasing, calls

for intelligent, transparent and inclusive detection

tools. Our work presents a new cross-discipline

framework that leverages multilingual NLP with

transformer-based deep learning, citation-aware

approach, and explainable AI to address these

emerging problems in a holistic manner. Leveraging

deep semantic analysis beyond simple matching, the

suggested model achieves high accuracies in

identifying paraphrased, cross-lingual, AI peer

assisted plagiarism, as compared to standard practice,

and particularly reduces false positives when present

with context based evaluation of citations.

The explainability tools integrated into the

platform establish not only trust with end-users, but

serve pedagogical goals by clarifying to students why

any particular content is flagged. Its multinodal

representation expanding to the code and visual level

will also provide the flexibility to be adopted across

different academic domains. The solution has

potential as a utility to modern educational

ecosystems with support for real-time calculations

and flexible deployment via API integration with

institutional systems.

At its core, this work contributes to the domain of

academic integrity by outlining a sustainable,

equitable and forward-facing model for plagiarism

detection – one that corresponds to the changing face

A Multilingual Explainable NLP and Deep Learning-Based Framework for Intelligent Plagiarism Detection and Academic Content

Validation

361

of international education and the responsible

application of AI.

REFERENCES

AbuAlRub, M., & Bader, A. (2024). Deep learning

detection method for large language models-generated

scientific text. arXiv preprint arXiv:2403.00828.

https://arxiv.org/abs/2403.00828arXiv+1SpringerLink

Ahire, P., Wadekar, Y., Shendge, T., Dhokale, M., & Ohol,

V. (2021). Plagiarism detection with paraphrase

recognizer using deep learning. International Research

Journal of Engineering and Technology (IRJET), 8(12),

1353–1356. https://www. irjet.net/archi ves/V8/i12/IR

JET-V8I12228.pdfIRJET

Altynbek, A., Turan, C., & Makhmutova, A. (2025).

Plagiarism types and detection methods: A systematic

survey of algorithms in text analysis. Frontiers in

Computer Science, 7, Article 1504725.

https://doi.org/10.3389/fcomp.2025.1504725Frontiers

+1Frontiers+1

Amzuloiu, C., Mihăescu, M. C., & Rebedea, T. (2021).

Combining Encoplot and NLP-based deep learning for

plagiarism detection. In Intelligent Data Engineering

and Automated Learning – IDEAL 2021 (pp. 97–106).

Springer. https://doi.org/10.1007/978-3-030-91608-

4_10SpringerLink

Bakhteev, O., & Ivanov, V. (2021). Cross-language

plagiarism detection: A case study. In Proceedings of

the European Conference on Academic Integrity and

Plagiarism (pp. 45–52). European Network for

Academic Integrity.

https://www.academicintegrity.eu/conference/proceedi

ngs/2021/bakhteev_et_al21.pdfacademicintegrity.eu

Chang, C., Alsharma, A., & Nesreen, A. (2024). T-SRE:

Transformer-based semantic relation extraction for

contextual plagiarism detection. Journal of King Saud

University - Computer and Information Sciences.

https://doi.org/10.1016/j.jksuci.2024.02.003ScienceDi

rect

Kumar, A., & Kaur, P. (2025). A comprehensive strategy

for identifying plagiarism in academic writing using

NLP and deep learning. Journal of Information and

Optimization Sciences, 46(1), 89–104.

https://doi.org/10.1007/s43995-025-00108-

1SpringerLink

Kumar, V., & Sharma, R. (2022). Plagiarism detection

system in scientific publication using deep learning.

International Journal on Technical and Physical

Problems of Engineering (IJTPE), 14(4), 17–24.

https://www.iotpe.com/IJTPE/IJTPE-2022/IJTPE-

Issue53-Vol14-No4-Dec2022/3-IJTPE-Issue53-

Vol14-No4-Dec2022-pp17-24.pdfiotpe.com

Miller, J., & Davis, L. (2023). Enhanced plagiarism

detection through advanced natural language

processing techniques. International Journal of

Advanced Computer Science and Applications, 14(9),

123–130. https://thesai.org/Dow nloads/Volu

me14No9/Paper_44-Enhanced_P lagiarism_Det

ection_Through_Advanced_Natural_Language.pdfThe

Science and Information Organization+1The Science

and Information Organization+1

Moravvej, S. V., Mousavirad, S. J., Moghadam, M. H., &

Saadatmand, M. (2021). An LSTM-based plagiarism

detection via attention mechanism and a population-

based approach for pre-training parameters with

imbalanced classes. arXiv preprint arXiv:2110.08771.

https://arxiv.org/abs/2110.08771arXiv+1arXiv+1

Quidwai, M. A., Li, C., & Dube, P. (2023). Beyond black

box AI-generated plagiarism detection: From sentence

to document level. arXiv preprint arXiv:2306.08122.

https://arxiv.org/abs/2306.08122ACL

Anthology+2arXiv+2arXiv+2

Quidwai, M. A., Li, C., & Dube, P. (2023). Beyond black

box AI-generated plagiarism detection: From sentence

to document level. In Proceedings of the 18th

Workshop on Innovative Use of NLP for Building

Educational Applications (BEA 2023) (pp. 727–735).

Association for Computational Linguistics.

https://doi.org/10.18653/v1/2023.bea-1.58

arXiv+2ACL Anthology+2arXiv+2

Ravichandran, R., & Kumar, S. (2022). Plagiarism

detection using natural language processing and

machine learning. International Journal of Advanced

Research in Computer and Communication

Engineering (IJARCCE), 11(7), 45–50.

https://ijarcce.com/wpcontent/uploads/2022/08/IJARC

CE.2022.117114.pdfPeer-reviewed Journal

Shouman, M. (2022). Reliable plagiarism detection system

based on deep learning techniques. Neural Computing

and Applications, 34(15), 11689–11700.

https://doi.org/10.1007/s00521-022-07486-

wSpringerLink+1SpringerLink+1

Smith, A., & Lee, B. (2023). Utilizing deep natural

language processing to detect plagiarism. In

Proceedings of the International Conference on

Artificial Intelligence and Soft Computing (pp. 345–

356). Springer

Wahle, J. P., Ruas, T., Foltýnek, T., Meuschke, N., & Gipp,

B. (2021). Identifying machine-paraphrased

plagiarism. arXiv preprint arXiv:2103.11909.

https://arxiv.org/abs/2103.11909arXiv+3arXiv+3Sprin

gerLink+3

Wahle, J. P., Ruas, T., Foltýnek, T., Meuschke, N., & Gipp,

B. (2021). Are neural language models good

plagiarists? A benchmark for neural paraphrase

detection. arXiv preprint arXiv:2103.12450.

https://arxiv.org/abs/2103.12450

arXiv+3arXiv+3arXiv+3

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

362