International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
Multimodal Cyberbullying Detection and Neutralization System in Social Media using Generative AI
Santhi1 , Preethiga2 , Prema3 , Sudarshan4
1Professor, Dept. of Information Technology, Puducherry Technological University, Puducherry, India
234Under Graduate Students, Dept. of Information Technology, Puducherry Technological University, Puducherry, India
Abstract - The exponential growth of social media platformshasprecipitatedariseincyberbullying,presenting severe psychological and emotional risks that traditional reactive, text-based detection mechanisms fail to mitigate. Modern harassment is increasingly characterized by contextualcomplexity,sarcasm,andmultimodalintegration, where abusive content is often embedded within images, emojis,ormemestobypassstandardfilters.Toaddressthese challenges, this paper proposes a Multimodal Cyberbullying Detection and Generative Semantic Neutralization System (MCDNS-SMGAI),aframeworkdesignedtofacilitatereal-time online safety. The system employs a sophisticated ingestion pipeline utilizing Pytesseract for OCR-based text extraction and the Demoji library for graphical icon normalization. A fine-tuned RoBERTa transformer model serves as the core analytical engine, leveraging bidirectional self-attention mechanisms to categorize content across six granular demographic classes with a peak detection accuracy of 96.44%.Beyondmere detection,theframeworkintroduces a Generative AI-driven neutralization module powered by the Gemini-1.5-Flash model. Through targeted prompt engineering, the module identifies toxic linguistic spans and reformulates them into polite, constructive alternatives, successfullyreducingtoxicityscoresfrom0.94toanegligible 0.03 while meticulously preserving the user's original communicative intent. Integrated via a Flask-React architecture with a total end-to-end latency of 450 milliseconds, the proposed system offers a scalable, transparent, and proactive solution for fostering inclusive digitalenvironments
Key Words: Cyberbullying Detection, Generative AI, Multimodal Fusion, RoBERTa, Natural Language Processing, Semantic Neutralization, Social Media Moderation, Context-Aware AI, Real-Time Monitoring.
1.INTRODUCTION
The rapid proliferation of social media platforms has revolutionized global communication, enabling users to interactthroughdiversemodalitiesincludingtext,captions, emojis, and high-resolution imagery. While this digital connectivityfosterssocialengagement,ithassimultaneously precipitated a significant rise in cyberbullying and online harassment, frequently targeting women and vulnerable demographics.Moderncyberbullyingisnolongerconfined
to direct offensive text; it manifests through complex channelssuchassarcasm,symbolicexpressions,andabusive textembeddedwithinmemesorscreenshots.
Conventional moderation systems and rule-based detectionapproachesprimarilyrelyonkeywordfilteringand static text analysis, which are increasingly insufficient to addressthecontextualnuancesofmoderndigitaldiscourse. Thesetraditionalmethodsoftenlacktheintegrationrequired tosynthesizetextandvisualdata,resultingintheincomplete detection of harmful interactions. Furthermore, many existingsystemsarepurelyreactive,eitherblockingcontent entirelyorutilizingpassivemaskingtechniquesthatdisrupt theconversationalflowandlacktransparencyfortheenduser.
Recent advancements in Natural Language Processing (NLP) and transformer-based architectures have enabled moresophisticatedsentimentandtoxicityanalysis.Models such as Bidirectional Encoder Representations from Transformers (BERT) and Robustly Optimized BERT approach (RoBERTa) have demonstrated strong results in capturingbidirectionalsemanticrelationships.Additionally, GenerativeAImodelsprovidethecapabilitytomovebeyond mere detection, offering the potential to transform hostile language into constructive communication through automatedsemanticneutralization.However,manycurrent implementationsfunctionasseparatetasks,lackingaunified real-time pipeline that combines multimodal extraction, context-awareclassification,andactivegenerativecorrection.
To overcome these limitations, this paper proposes a Multimodal Cyberbullying Detection and Generative Semantic Neutralization System (MCDNS-SMGAI). The frameworkingestsmultimodalinputs includingcaptions, replies, and text-in-images utilizing Pytesseract for OCR extractionandtheDemojilibraryfornormalization.AfinetunedRoBERTaclassificationnetworkanalyzestheseunified sequences to distinguish between genuine cyberbullying, implicittoxicity,andharmlesssarcasm.Flaggedcontentis subsequentlyprocessedbyagenerativelanguagelayerthat identifies toxic spans and replaces them with polite alternatives using the Gemini-1.5-Flash model. By integrating a multi-model detection pipeline with a generativeinterventionmodule,theproposedsystemoffers aproactiveandinclusivesolutionforfosteringsaferdigital environments.
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
2. LITERATURE REVIEW
Abdullah et al. [1] utilized several machine learning (ML) models,includingSupportVectorMachines(SVM),Random Forest(RF),andNaïveBayes,todetectcyberbullyingthrough Natural Language Processing (NLP) cleaning and Term Frequency-Inverse Document Frequency (TF-IDF) feature extraction.WhiletheirintegrationofBidirectionalEncoder Representations from Transformers (BERT) improved contextual accuracy, the system remains text-centric and does not incorporate multimodal support or generative neutralization.
Pramaetal.[12]developedanAI-enabledframeworkusing Long Short-Term Memory (LSTM) networks to classify cyberbullying severity into mild and severe levels by combining social media text with user-specific attributes. Despite achieving strong results in severity detection, the reliance on user-specific data raises significant privacy concerns, and the architecture lacks a real-time implementationforimmediatecontentmitigation.
Rahman et al. [13] introduced the CAMFusion framework, whichintegratesvideoandtextualcuestodetectsarcasmand humorthroughcontext-awaremultimodalfusion.Although this approach successfully extracts visual and semantic representations using deep learning, it is not specifically optimized for cyberbullying detection and involves high computationalcomplexitythatlimitsreal-timesocialmedia deployment.
Zhang et al. [24] proposed a cross-platform multimodal transfer learning framework that analyzes both text and imagesacrossdifferentsocialplatformstoimprovedetection generalization. While the framework improves accuracy throughsharedknowledge,itlacksamechanismforsarcasm detectionanddoesnotprovideanautomatedGenerativeAI (GenAI)moduleforcontentdetoxification
Dale et al. [5] explored text detoxification using large pretrainedneuralmodelstorewritetoxicsentencesintopolite forms through text style transfer. Their research demonstratesthatsemanticmeaningcanbepreservedwhile reducingharmfulexpressions;however,themodelsoperate purely on textual data and are not integrated with a prior cyberbullyingdetectionstagetoautomatetheintervention process.
Yuetal.[23]presentedatwo-stagedetoxificationframework utilizingLargeLanguageModel(LLM)fine-tuningtoimprove modelgeneralizationandsemanticpreservation.Whiletheir approachensuresthattransformedoutputsretainintended meanings,itremainsaunimodal,text-onlysolutionthatlacks the multimodal support and real-time social media integrationnecessaryfordynamicdigitalenvironments.
3. INFERENCE FROM LITERATURE SURVEY
The comprehensive review of existing literature reveals severalcriticallimitationsincurrentcyberbullyingdetection and content moderation strategies. Existing approaches primarilyaddressisolatedcomponentsoftheproblem,such ascontextualclassificationusingtransformers,imbalanced dataset handling, or text-based style transfer, rather than providinganintegratedandunifiedmitigationframework [1],[5],[12].Mostcurrentdetectionmechanismsarestrictly reactive and text-centric, leading to a failure in capturing multimodalharassment suchasabusivetextembeddedin imagesoremojis resultinginincompleteprotectionwithin dynamicsocialmediaenvironments[15],[25].
Machine learning and transformer-based architectures, including BERT and its derivatives, have significantly enhancedsemanticunderstandingandintentclassification [7], [14]. However, standalone detection models lack a proactive intervention layer, often defaulting to binary maskingorcontentdeletionwhichdisruptsuserinteraction andcommunicativeflow[11],[16].Furthermore,noexisting approachsuccessfullybridgesthegapbetweenhigh-fidelity multimodal extraction and real-time generative semantic neutralization,wheretoxiccontentisnotjustidentifiedbut constructivelyreformulatedinsitu[18],[23]
Current activity monitoring and automated moderation methods often operate independently, resulting in a fragmented user experience where the context of a conversationisfrequentlylostduringtheintervalbetween detectionandmanualintervention.Additionally,thelackof sarcasm-aware gating and visual semiotic analysis in traditional models reduces detection accuracy in social media dialects, which are heavily reliant on irony and graphicalicons[2],[13],[24].
Based on these identified gaps, this paper proposes the MCDNS-SMGAI Framework a Multimodal Cyberbullying DetectionandGenerativeSemanticNeutralizationSystem. Theproposedsystemaddressesallidentifiedlimitationsby integratinganOCR-basedmultimodalingestionpipeline,a context-awareRoBERTaclassificationengine,andanactive generativeinterventionmodulepoweredbytheGemini-1.5Flash model. By combining real-time detection with semantic detoxification into a single unified platform, the system shifts the moderation paradigm from passive censorshiptoinclusiveandsafedigitalcommunication.
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
4. PROPOSED SYSTEM
The proposed system introduces the MCDNS-SMGAI Framework, a Multimodal Cyberbullying Detection and Generative Semantic Neutralization System designed specifically for modern social media environments. The framework addresses the escalating psychological risks causedbytheproliferationofcomplex,multi-layereddigital harassment. Unlike traditional unimodal and reactive moderation tools, such as the basic machine learning classifiers proposed by Abdullah et al. [1] or the text-only LSTM severity models discussed by Prama et al. [12], the proposed system integrates real-time multimodal data ingestion, context-aware behavioural analysis, and active generativedetoxificationintoaunifiedplatform
Thesystemcontinuouslyprocessesdiverseuser-generated inputs,includingdirecttext,captions,graphicalemojis,and text-embeddedimagessuchasmemesorscreenshots.These multimodalelementsarecapturedandstandardizedthrough an ingestion pipeline utilizing Pytesseract for Optical Character Recognition (OCR) and the Demoji library for visualsemioticdecoding.Theaggregateddataisstructured into unified sequence tensors and fed into a dual-gate Context-AwareDetectionEngine.Thiscoreenginecombines afine-tunedRoBERTa(RobustlyOptimizedBERTapproach) transformer whichgenerates768-dimensionalcontextual embeddings using bidirectional self-attention with a VADER-based sentiment discordance calculator and a heuristicemoji-mappingdictionarytodetectimplicittoxicity andmaskedsarcasm
Theextractedcontextualfeaturesareprocessedthrougha fully connected dense layer and a Softmax classification function,whichcomputesaprobabilitydistributionacross six granular categories: Age, Ethnicity, Gender, Religion, OtherBullying,andNotBullying.Tominimizefalsepositives, the system employs a rigid probability gating and thresholding mechanism. Sequences with a toxicity probabilityexceeding0.8oranegativesentimentcompound scorebelow-0.5areflaggedashigh-risk.Conversely,content fallingbelowthiscriticalthresholdisclassifiedasstandard communication and allowed to pass freely without intervention.
Upon the detection of suspicious or abusive content, the system actively bypasses traditional binary blocking or rudimentary word-masking techniques (e.g., replacing wordswithasterisks),resolvingamajorlimitationfoundin existingNLPmoderationsystems.Instead,itautomatically enforces a Generative Semantic Neutralization protocol poweredbyaLargeLanguageModel(Gemini-1.5-Flash).The generative intervention layer isolates the specific harmful linguistic spans and dynamically substitutes them with polite,constructivegoodwords,meticulouslypreservingthe user's original communicative intent. To ensure transparencyandreal-timemonitoring,theintegratedFlask-
Reactdashboardprovidesimmediatefeedback,displaying the classification label, confidence score, and the safely neutralizedtext.Bycombiningadvancedmultimodalfeature extraction,bidirectionalcontextualclassification,andintentpreservinggenerativereformulation,theproposedMCDNSSMGAI framework significantly reduces the psychological impact of digital harassment and fosters a safe, inclusive socialmediaecosystem
5. SYSTEM ARCHITECTURE
The MCDNS-SMGAI framework is organized into three functional modules that operate sequentially and collaboratively to deliver end-to-end multimodal data ingestion, context-aware threat detection, and generative semanticneutralization.
5.1 Module I: Multimodal Data Collection and Preprocessing
This module serves as the foundational entry point of the framework, responsible for ingesting multimodal content, establishing a structured data format, and initializing the normalized textual baseline for downstream contextual analysis. The module processes incoming image-based mediautilizingPytesseractOpticalCharacterRecognitionto extract embedded text and applies the Demoji library to translate graphical emojis into descriptive equivalents. Extracted text and raw strings undergo rigorous heuristic refinement, including regex-based entity stripping, case normalization,andmorphologicallemmatizationtoreduce wordstotheirfundamentalrootforms.Themoduleaccepts rawsocialmediaposts,directtextsubmissions,embedded emojis,andtext-embeddedimagefilesasinputandproduces unified, sanitized, lemmatized textual sequences, and aggregatedfeaturevectorsforbaselinecontextualevaluation asoutput.
5.2Module II: Cyberbullying and Sarcasm Detection
Thismoduleformstheanalyticalcoreoftheframework.It continuouslyanalyzestheunifiedtextsequencesgenerated from the ingestion layer, extracts contextual embeddings, and passes them through a multi-layered classification architecturetocomputeprecisetoxicityprobabilities.The standardizedtextualdata,includingtranslatedemojisand OCR-extracted strings, are segmented and converted into high-dimensional numerical feature vectors using a RoBERTa-specificsub-wordtokenizerandattentionmasks. Thesestructuredtensorsaresimultaneouslyprocessedbya fine-tuned RoBERTa bidirectional transformer, a VADER sentiment discordance engine, and a heuristic emojimappingdictionarytocapturedeepsemanticdependencies andmaskedsarcasm.Theextractedcontextualfeaturesare evaluated by a fully connected dense layer and a Softmax classifier to produce a probability distribution across six
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
distinct demographic classes: Age, Ethnicity, Gender, Religion, Other Bullying, and Not Bullying. The predicted intent is then evaluated against a rigid confidence gate, triggering active moderation if the toxicity probability exceedsa 0.8thresholdorthesentimentcompoundscore falls below -0.5, immediately initiating the generative neutralizationsequenceandreal-timedashboardalerts
5.3 Module III: Generative Semantic Neutralization
Thismoduleenforcesautomatedcontentprotectioncontrols basedontheclassificationprobabilityreceivedfromModule II and provides users with constructive, polite text alternativesforeverydetectedabusiveevent.Uponreceiving a high-risk toxicity score exceeding the 0.8 threshold, the generativeinterventionengineautomaticallyneutralizesthe relevantharmfullinguisticspansinrealtime,substituting offensivewordswithsociallyacceptablelanguageusingthe Gemini-1.5-Flash model. For all flagged events, the integrated Flask-React architecture delivers an instantaneous push update to the user's social media frontend. The generative AI evaluates the underlying contextual meaning of the flagged sentence, producing a refinedsemanticstructurethateffectivelyreplacestoxicity whilestrictlypreservingtheuser'soriginalcommunicative intent.Insteadofpresentingonlyagenericblockedcontent label,thedashboardseamlesslydisplaysthetoxicityclass, theprobabilityconfidencescore,andthesafelyneutralized plain-languagealternative.Administratorsanduserscanuse thisinformationtoreviewmoderationactions,understand contextual flagging, and maintain uninterrupted, inclusive digitalcommunicationacrosstheplatform.
6. METHODOLOGY
The MCDNS-SMGAI framework follows a structured methodologythatbeginswiththecontinuouscollectionand standardization of multimodal user-generated content across social media platforms. Diverse behavioural inputs includingdirecttextcomments,graphicalemojis,andtextembedded media such as memes and screenshots are captured per user interaction. These inputs are preprocessed utilizing Pytesseract for OCR-based text extraction and the Demoji library for graphical icon normalization. The aggregated strings undergo rigorous morphological refinement including regex-based entity stripping and lemmatization and are subsequently convertedintohigh-dimensionalnumericalfeaturevectors using transformer-specific sub-word tokenization and dynamicattentionmasks.
Theseunifiedfeaturetensorsaresimultaneouslyprocessed byamulti-layeredanalyticalenginetodetectnuancedabuse. The core component, a fine-tuned RoBERTa transformer, leveragesbidirectionalself-attentionmechanismstomodel complex semantic dependencies and contextual relationshipswithinthetext.Concurrently,aVADER-based
sentiment discordance calculator evaluates the emotional polarity of the sequence to detect implicit hostility. A heuristic emoji-mapping module further analyzes the translated graphical icons to pinpoint disguised malicious intentandmaskedsarcasmthatstandardglobalNLPmodels typicallyoverlook.
The extracted contextual features from these components areaggregatedandpassedthroughafullyconnecteddense layerpairedwithaSoftmaxactivationfunctiontoproducea normalizedprobabilitydistribution.Thecontentisclassified across six specific categories: Age, Ethnicity, Gender, Religion, Other Bullying, and Not Bullying. The computed riskisevaluatedagainstarigidconfidencegatingmechanism to determine the intervention tier. A toxicity probability between 0.0 and 0.79, coupled with a neutral or positive sentimentscore,isclassifiedassafeandrequiresnoaction. Conversely, a toxicity probability of 0.8 or higher, or a negativesentimentcompoundscorebelow-0.5,isclassified asahigh-riskviolation,triggeringfullincidentloggingand immediateautomatedneutralizationenforcement.
Upon high-risk classification, the generative semantic mitigationlayerisautomaticallyactivated.Theframework utilizes the Gemini-1.5-Flash Large Language Model to calculate the contextual weight of the harmful linguistic spans contributing to the high toxicity score. Instead of executing a binary deletion, the generative model actively reformulatesthesentence,replacingexplicitandoffensive terminology with polite, constructive alternatives while strictlypreservingtheuser'soriginalcommunicativeintent. These final outputs are presented instantaneously on the userdashboardviaaFlask-Reactarchitecture,displayingthe specifictoxicityclass,theprobabilityconfidencescore,and thesafelyneutralizedtext,ensuringtransparent,inclusive, anduninterrupteddigitalcommunication.
7. CONCLUSIONS
The proposed Multimodal Cyberbullying Detection and GenerativeSemanticNeutralizationSystem(MCDNS-SMGAI) addresses the escalating psychological risks posed by complex,multi-layeredharassmentindynamicsocialmedia environments.Byintegratingarobustmultimodalingestion pipeline with a fine-tuned RoBERTa transformer, VADER sentiment discordance, and an automated generative intervention layer, the system achieves highly accurate detection of cyberbullying across diverse forms of digital communication,includingdirecttext,emojis,andembedded images. The rigid confidence gating and classification mechanism ensures precise, context-aware responses, minimizingfalsepositivesandreducingrelianceonmanual moderation.Real-timesemanticneutralization,poweredby theGemini-1.5-Flashmodel,seamlesslyreformulatestoxic content into constructive language immediately upon detection,mitigatingemotionalharmbeforetheend-useris exposed. The integration of this generative capability
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
transformstraditional,passivecensorshipintoaninclusive, transparent moderation process, improving the digital experience while maintaining an uninterrupted communicativeflow.
The framework is highly applicable to diverse digital ecosystems,includingmainstreamsocialmediaplatforms, multiplayer online gaming communities, educational discussionforums,andenterprisecommunicationnetworks requiring continuous content moderation. It effectively supports the protection of vulnerable demographics engaging in online interactions by providing a proactive, automatedsafetynet.Thesystemishighlysuitableforsocial technologyconglomerates,educationalinstitutions,digital mentalhealthplatforms,andanyorganizationprioritizing safe,inclusive,andconstructivedigitalenvironments.
Future enhancements may include the integration of advanced cross-lingual models to support real-time cyberbullying detection and neutralization in resourceconstrained and regional languages, adaptive generative promptingdynamicallytailoredtothespecificagegroupor psychological profile of the user, cross-platform API deployment for seamless integration into diverse social media architectures, and the extension of the multimodal pipelinetoanalyzeliveaudioandvideostreams.Overall,the framework demonstrates the effectiveness of combining sophisticated multimodal feature extraction, bidirectional transformeranalysis,andgenerativesemanticenforcement to achieve a proactive, transparent, and highly scalable solutionfordigitalharassmentmitigationinmodernsocial networks.
ACKNOWLEDGEMENT
TheauthorswouldliketoexpresssinceregratitudetoDr.G. Santhi, Professor, Department of Information Technology, Puducherry Technological University, for her valuable guidanceandcontinuoussupportthroughoutthisresearch work.TheauthorsalsothanktheDepartmentofInformation Technology, Puducherry Technological University, for providingthenecessaryresourcesandfacilitiestocarryout thiswork
REFERENCES
[1] M.Abdullah,I.L.Latif,N.Hafeez,F.Ullah,G.Sidorov,E.F. Riverón, and A. Gelbukh, “Cyberbullying Detection on Social Media Using Machine Learning Techniques,” ComputaciónySistemas,vol.29,no.3,2025.
[2] M. Alharbi, F. Alqahtani, and S. Alsubaie, “Sarcasm detection in online social networks using deep contextual embeddings,” Applied Soft Computing, vol. 148,p.110891,2024.
[3] F.N.Al-Wesabi,M.Obayya,J.Alsamri,R.Alabdan,N.O. Aljehane,S.Alazwari,F.F.Alruwaili,M.A.Hamza,andA. Swathi,“AutomaticRecognitionofCyberbullyinginthe Web of Things and Social Media Using Deep Learning
Framework,”IEEETransactionsonBigData,vol.11,no. 1,pp.259–270,2025.
[4] T. Davidson, D. Warmsley, M. Macy, and I. Weber, “Automatedhatespeechdetectionandtheproblemof offensive language,” in Proc. Int. AAAI Conf. Web Soc. Media,vol.11,no.1,pp.512–515,2017
[5] D. Dale, L. Li, and L. Mohit, “Text Detoxification using LargePre-trainedNeuralModels,”Transactionsofthe AssociationforComputationalLinguistics,vol.9,2025
[6] A.DerbalaYacoub,A.ElsayedAboutabl,andS.O.Slim, “Multilingual sarcasm detection for enhancing sentimentalanalysisusingdeeplearning,”J.Commun. Softw.Syst.,vol.20,no.4,pp.278–280,2024
[7] J.Devlin,M.-W.Chang,K.Lee,andK.Toutanova,“BERT: Pre-training of deep bidirectional transformers for language understanding,” in Proc. NAACL-HLT, pp. 4171–4186,2019
[8] M.Ibrahim,A.Hassan,andR.Khalifa,“FusionBullyNet:A robust English–Arabic cyberbullying detection framework using dual-encoder transformer architecture,”Mathematics,vol.13,p.172,2025
[9] V. S. Kumar, R. Prakash, and N. Menon, “Real-time cyberbullying detection across multimodal platforms usingXLNet,”inProc.ICAECA,p.412,2025
[10] Y.Lu,Y.Zhou,J.Li,Y.Zhang,W.Wang,X.Li,M.Zhang,F. Liu, J. Yu, and M. Zhang, “Adaptive Detoxification: Safeguarding General Capabilities of LLMs through Toxicity-Aware Knowledge Editing,” Findings of ACL, 2025
[11] R.Mishra,S.Patra,andA.K.Tripathy,“Acomprehensive survey on cyberbullying detection using natural language processing and deep learning techniques,” IEEEAccess,vol.12,pp.11245–11268,2024
[12] T.T.Prama,J.F.Amrin,M. M.Anwar,andI.H.Sarker, “AI-Enabled User-Specific Cyberbullying Severity DetectionwithExplainability,”IEEEAccess,vol.13,pp. 12543–12558,2025.
[13] M.Rahman,P.K.Dhar,M.A.M.Provath,K.Deb,andT. Shimamura, “Context-Aware Multi-Modal Fusion Framework for Detecting Sarcasm and Humor IntegratingVideoandTextual Cues,”IEEEAccess,vol. 13,2025.
[14] S. K. Roy, P. Dutta, and M. Banerjee, “Text-based cyberbullying detection on social media using transformer-based models,” Expert Systems with Applications,vol.235,p.121012,2024.
[15] F. R. Sayed, E. H. Elnashar, and F. A. Omara, “Cyberbullyingdetectioninsocialmediausingnatural language processing,” Scientific African, vol. 28, p. e02713,2025.
[16] A. Sharma, N. Gupta, and S. Verma, “Cyberbullying detection using NLP techniques,” Int. J. Adv. Res. Comput.Sci.,vol.15,no.2,pp.45–50,2023–2024
[17] A.Sharma,N.Jain,andV.Kumar,“Detectionofabusive and hateful content against women on social media platforms,”J.Intell.Inf.Syst.,vol.62,no.1,pp.89–110, 2024
[18] K. Singh, R. Mehta, and P. Bansal, “Automatic neutralization of toxic language using NLP-based text transformation,”Inf.Manag.DataInsights,vol.4,no.1,p. 100204,2025
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
[19] X. Song, Y. Liu, and H. Chen, “Assessing the human likeness of AI-generated counterspeech,” in Proc. COLING,p.239,2025.
[20] I.Tabassum,S.Rahman,andM.Hossain,“Multilingual multimodal cyberbullying detection through adaptive andhierarchicalfusion,”IEEEAccess,p.142,2026
[21] Z. Waseem and D. Hovy, “Hateful symbols or hateful people?Predictivefeaturesforhatespeechdetection,” inProc.NAACLHumanLang.Technol.,pp.88–93,2016
[22] Y.Wen,Q.Zhang,andL.Zhao,“DashFusion:Dual-stream alignment with hierarchical bottleneck fusion for multimodal sentiment analysis,” IEEE Trans. Neural Netw.Learn.Syst.,p.89,2025
[23] J.Yu,Y.Zhao,J.Zhu,W.Shao,B.Pang,Z.Zhang,andX.Li, “Text Detoxification: Data Efficiency, Semantic Preservation and Model Generalization,” in Proc. EMNLP,pp.32172–32186,2025
[24] W. Zhang, C. Dong, A. Yao, A. Nazari, and A. Gaddam, “Cross-Platform Multi-Modal Transfer Learning Framework for Cyberbullying Detection,” Electronics, vol.15,2026
[25] S.S.-Us-Sakib,M.R.Rahman,M.S.A.Forhad,andM.A. Aziz,“CyberbullyingDetectionofResource-Constrained LanguagefromSocialMediaUsingTransformer-Based Approach,”NaturalLanguageProcessingJournal,vol.9, p.100104,2024
BIOGRAPHIES
Dr. G. Santhi holdsB.E.,M.E.,andPh.D. degreesinCSE,withspecializationin CN,IS,WirelessNetworks,andCA
Preethiga P isaB.Tech(IT)student atPTU,currentlyfocusedonthefields ofAI&ML.ShehasproficiencyinPython, C,C++,Java.
Prema J isaB.Tech(IT)student atPTU,currentlyfocusedonthefields ofFullStackWebDevelopmentAI&ML. ShehasproficiencyinPython,Java, MySQL
Sudharshan M isaB.Tech(IT)student at PTU, with a keen interest in datadrivensystemdesign.Heisproficientin programminglanguagessuchasPython