Shield AI: Capsule Network Integrated with Multilingual Transformer for Abusive Text Detection in Online Social Networks | IJCT Volume 13 – Issue 3 | IJCT-V13I3P102
Shield AI: Capsule Network Integrated with Multilingual Transformer for Abusive Text Detection in Online Social Networks | IJCT Volume 13 – Issue 3 | IJCT-V13I3P102
Online social networks generate a massive amount of user-generated text, including abusive, hateful, and offensive content. Manual moderation of such content is fundamentally inefficient, particularly when diverse grammatical structures and regional linguistics are involved simultaneously. This paper presents Shield AI, an intelligent, autonomously scaling abusive text detection system that integrates a Deep Multilingual Transformer with a localized Capsule Network matrix. Utilizing a distilled, multi-layer BERT framework (distilbert-base-multilingual-cased), the system precisely maps the semantic nuances of textual payloads across English, Hindi, Telugu, Tamil, Malayalam, and Kannada within a unified, invariant dimensional vector space. Crucially, instead of flattening these embeddings into a standard dense classification head, we incorporate a Capsule Network hierarchy. This structural configuration significantly amplifies classification accuracy by preserving deep contextual relationships and hierarchical spatial features naturally encoded within the transformer’s multi-head attention blocks. The theoretical implementation of Dynamic Routing by Agreement effectively minimizes the semantic information loss characteristic of traditional Max-Pooling operations. Evaluated against the robust Multilingual Abusive Comment Dataset (MACD) consisting of 5,000 highly heterogeneous cross-lingual samples, the proposed hybrid pipeline establishes a test accuracy of 73.85% and an F1-score of 72.95% running natively on cost-efficient, CPU-optimized hardware arrays (Intel Core i5, 8GB RAM threshold target). The comprehensive experimental validation incorporates rigorous ablation studies comparing recurrent baselines against the proposed hybrid structure, definitively demonstrating its efficacy as a scalable, high-throughput analytical instrument for real-time cyberbullying prevention, digital platform safeguarding, and cognitive content moderation.
Keywords
Capsule Network, DistilBERT, Multilingual Transformer, Abusive Text Detection, Deep Learning, Cybersecurity, Content Moderation, Nano Banana Analysis, Dynamic Routing.
Conclusion
The proliferation of multilingual social media fundamentally invalidates conventional deterministic text moderation philosophies. We proposed Shield AI: an elegant, highly scalable algorithmic apparatus designed to execute complex abusive text classification over severely fragmented linguistic boundaries. By intricately amalgamating the overwhelming multi-head vocabulary matrices embedded within the DistilBERT framework with the revolutionary spatial memory mechanics of a deep Capsule Network, Shield AI achieves profound text validation.
Our theoretical implementation overcomes the deeply established data-compression losses typical of legacy classifiers by actively modeling syntactical token hierarchies via mathematical squashing operations. With a confirmed validation F1 baseline of 72.95% on notoriously complicated datasets, the hybridized format securely outperforms linear benchmarks without necessitating catastrophic inference latencies. Our strategic architectural choices guarantee that cost-restricted Web and independent Mobile applications can autonomously sanitize immense data streams reliably, accelerating the broader goal of securing global digital equity.
References
[1]A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. Gomez, ‘Attention Is All You Need’, NIPS 2017.
[2]J. Devlin, M. Chang, ‘BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding’, NAACL 2019.
[3]S. Sabour, N. Frosst, and G. E. Hinton, ‘Dynamic Routing Between Capsules’, NIPS 2017.
[4]V. Sanh, L. Debut, ‘DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter’, arXiv 2019.
[5]T. Wolf, L. Debut, ‘HuggingFace’s Transformers: State-of-the-art Natural Language Processing’, EMNLP 2020.
[6]L. Dong, N. Yang, ‘Unified Language Model Pre-training for Natural Language Understanding’, NIPS 2019.
[7]P. Bojanowski, E. Grave, ‘Enriching Word Vectors with Subword Information’, TACL 2017.
[8]I. Sutskever, O. Vinyals, ‘Sequence to Sequence Learning with Neural Networks’, NIPS 2014.
[9]A. Joulin, E. Grave, ‘Bag of Tricks for Efficient Text Classification’, EACL 2017.
[10]D. Bahdanau, K. Cho, ‘Neural Machine Translation by Jointly Learning to Align and Translate’, ICLR 2015.
[11]Y. Kim, ‘Convolutional Neural Networks for Sentence Classification’, EMNLP 2014.
[12]C. Sun, X. Qiu, ‘How to Fine-Tune BERT for Text Classification?’, CCF International Conference, 2019.
[13]W. Yin, K. Kann, ‘Comparative Study of CNN and RNN for Natural Language Processing’, arXiv 2017.
[14]M. E. Peters, M. Neumann, ‘Deep Contextualized Word Representations’, NAACL 2018.
[15]K. Clark, U. Khandelwal, ‘What Does BERT Look At? An Analysis of BERT’s Attention’, ACL 2019.
[16]G. Hinton, S. Sabour, ‘Matrix Capsules with EM Routing’, ICLR 2018.
[17]A. Conneau, K. Khandelwal, ‘Unsupervised Cross-lingual Representation Learning at Scale (XLM-R)’, ACL 2020.
[18]J. Zhao, T. Wang, ‘Gender Bias in Coreference Resolution’, NAACL 2018.
[19]E. Almazrouei, ‘Arabic Hate Speech Detection using Deep Learning’,
Data in Brief 2020. IEEE RESEARCH PAPER – Shield AI – Page 5
[20]S. Hochreiter, J. Schmidhuber, ‘Long Short-Term Memory’, Neural Computation 1997.
How to Cite This Paper
vamsi krishna Alam, Prem Siddartha reddy Ankinapalli, G.Mahalakshmi (2026). Shield AI: Capsule Network Integrated with Multilingual Transformer for Abusive Text Detection in Online Social Networks. International Journal of Computer Techniques, 13(3). ISSN: 2394-2231.
