AI Guardian - Multilingual Hate Speech and Cyberbullying Detection System

Authors: Mrs. V. Pallavi, CH Varshith, R Akhil, D Sathwik

DOI Link: https://doi.org/10.22214/ijraset.2026.79178

Abstract

The exponential rise of social media and dig-ital communication tools has resulted in a sharp rise of hate speech, offensive language, and cyberbullying, which are posing a major threat to the safety and mental health of internet users. Recent research studies have shown that millions of harmful messages are being generated every day in a multilingual and code-mixed environment. The existing methods of handling these are found to be inefficient and ineffective. The traditional methods are mostly monolingual and cannot handle implicit and contextual abuse patterns. In this paper, a novel and intelligent content moderation framework named AI Guardian is proposed for the detection of hate speech and cyberbullying in a multilingual environment. The proposed framework utilizes the latest transformer-based models such as Multilingual BERT (mBERT) and XLM-RoBERTa (XLM-R) and combines them with sequential models such as LSTM and CNN for the detection of hate speech and cyberbullying. The proposed framework utilizes a hybrid pipe-line for the detection of hate speech and cyberbullying. This paper proposes a framework for scalable and intelligent multilingual content moderation known as AI Guardian, specifically for the detection of hate speech and cyberbullying in real-time. The proposed system utilizes the power of advanced state-of-the-art models such as transformers, specifically Multilingual BERT (mBERT) and XLM-RoBERTa (XLM-R), in combination with other sequential models such as LSTM and CNN. The system utilizes a hybrid pipeline consisting of pre-processing, feature extraction, context-based classification, and severity scoring.

Introduction

The rise of digital communication and social media in India, with over 950 million internet users by 2024, has led to a surge in online abuse, hate speech, and cyberbullying, amplified by multilingual and code-mixed content like Hinglish and Tanglish. Traditional moderation methods are insufficient due to linguistic diversity, informal language, and complex context. AI Guardian addresses these challenges by using deep learning and transformer-based models (mBERT, XLM-R) for real-time, multilingual detection, handling both monolingual and code-mixed text. The system incorporates Explainable AI, confidence scoring, severity levels, and a user-friendly dashboard, providing scalable, interpretable, and effective online content moderation across platforms.

Conclusion

This paper has presented a novel, intelligent, and scalable framework, namely AI Guardian, which is capable of effectively detecting multilingual hate speech and cyberbullying in real-time digital communication environments. The proposed system effectively overcomes the major shortcomings of existing content moderation techniques, which are incapable of dealing with multilingual content and lack contextual understanding, and are not transparent in their decision-making process. The proposed system effectively leverages advanced Natural Language Processing techniques along with transformer-based models, namely Multilingual BERT and XLM-RoBERTa, to effectively comprehend semantic and contextual relationships in text-based user-generated content. The severity scoring module further boosts the capabilities of the proposed system, which categorizes hate speech and cyberbullying based on their severity level, thus prioritizing their moderation process. The proposed system, namely AI Guardian, possesses a significant advantage in terms of its Explainable AI module, which provides clear and interpretable outputs in terms of offensive words and their reasoning, thus promoting higher levels of user trust and ease of usage. The proposed system is designed with a modular and efficient architecture, which enables its real-time usage with minimal latency levels. The experimental results have proved the high accuracy levels of the proposed system, which are in the range of 92-95%, thus validating its effectiveness in dealing with multilingual and code-mixed hate speech and cyberbullying. The proposed system, namely AI Guardian, is a robust, interpretable, and scalable solution in dealing with harmful online content, which contributes significantly to promoting a safe digital communication environment.

References

[1] OpenAI / Google Research, “Large Language Models for Toxicity Detection and Content Moderation,” 2025. [2] A. Bhardwaj et al., “Hinglish Offensive Language Detection using Deep Learning Techniques,” 2024. [3] R. Kumar et al., “Cyberbullying Detection using Hierarchical Transformer Networks,” 2024. [4] M. Ribeiro et al., “Explainable Content Moderation using Attention-Based Models,” 2024. [5] T. Ranasinghe et al., “Cross-lingual Hate Speech Detection using XLM-RoBERTa,” 2023. [6] T. Mandl et al., “Overview of the HASOC 2023 Shared Task: Hate Speech and Offensive Content Identification in Indo-European Languages,” 2023. [7] A. Saha et al., “Contextual Hate Speech Detection in Social Media Conversations,” 2023. [8] A. Mozafari et al., “Hate Speech Detection with BERT and CNN: A Study on Explainability,” 2023. [9] F. Ousidhoum et al., “Multilingual and Multi-Aspect Hate Speech Analysis Dataset,” 2022. [10] I. Glavaš et al., “Zero-Shot Cross-Lingual Hate Speech Detection,” 2022. [11] Z. Zhang et al., “Multimodal Hate Speech Detection in Memes using Vision-Language Models,” 2022. [12] P. Fortuna and S. Nunes, “A Survey on Automatic Detection of Hate Speech in Text,” ACM Computing Surveys, 2018. [13] M. Zampieri et al., “SemEval-2019 Task 6: Identifying and Categorizing Offensive Language in Social Media (OffensEval),” 2019. [14] Z. Waseem and D. Hovy, “Hateful Symbols or Hateful People? Predictive Features for Hate Speech Detection on Twitter,” NAACL, 2016. [15] A. Das et al., “Hate Speech Detection in Code-Mixed Hindi-English Text,” 2021. [16] J. Devlin et al., “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding,” NAACL, 2019. [17] A. Conneau et al., “Unsupervised Cross-lingual Representation Learning at Scale (XLM-R),” 2020. [18] T. Wolf et al., “HuggingFace Transformers: State-of-the-Art Natural Language Processing,” 2020. [19] S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,” Neural Computation, 1997. [20] Y. Kim, “Convolutional Neural Networks for Sentence Classification,” EMNLP, 2014.

Copyright

Copyright © 2026 Mrs. V. Pallavi, CH Varshith, R Akhil, D Sathwik. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Download Paper

Paper Id : IJRASET79178

Publish Date : 2026-03-31

ISSN : 2321-9653

Publisher Name : IJRASET

DOI Link : Click Here