MUMBAI, India, July 7 -- Intellectual Property India has published a patent application (202641076888 A) filed by Sr University on June 22, 2026, for A System Of Efficient Toxic-Comment-Classification Models Using Bert.
Inventors include Sukhaverji Ghate; Prof Deepak Garg; and Dr. Dilbag Singh.
The application for the patent was published on July 03, 2026, under issue no. 27/2026.
Abstract: A SYSTEM OF EFFICIENT TOXIC-COMMENT-CLASSIFICATION MODELS USING BERT The present invention discloses an efficient toxic-comment-classification system leveraging the contextual learning capabilities of BERT combined with a CNN-based ensemble. The system preprocesses and tokenizes online comments using BERT’s WordPiece tokenizer, generating deep bidirectional contextual embeddings that capture semantic nuances, sarcasm, and implicit toxicity. A customized multi-layer neural network classifier performs binary classification of harmful versus non-harmful comments. To enhance robustness, a convolutional neural network is trained independently to capture local textual patterns, and its outputs are integrated with those of the BERT classifier through a weighted ensemble mechanism. This hybrid approach combines global contextual semantics with local feature detection, achieving superior accuracy and F1-scores compared to standalone transformer or CNN models. Tested on a balanced dataset of 50,000 comments, the system achieved an F1-score of 0.94 and accuracy of 95.2%. The invention provides a scalable, real-time moderation solution for online platforms, enabling reliable detection of subtle and context-dependent toxic language.
Disclaimer: Curated by HT Syndication.
