Abstract :

Automatic Identification Of Harmful Content On Social Platforms Requires Systems That Maintain Classification Precision Without Sacrificing Processing Speed. In Real-world Moderation Pipelines, This Balance Is Critical Because Largescale User-generated Content Must Be Filtered In Near Real Time While Preserving Contextual Understanding. Deep Learning Architectures Such As BERT Deliver Superior Accuracy But Demand Extensive Computing Resources, Whereas Conventional Algorithms Like SVMs Execute Rapidly Yet Fail To Capture Semantic Context Adequately. We Introduce A Dual-stage Classification Framework Employing Confidencebased Decision Routing To Optimize The Trade-off Between Computational Efficiency And Detection Performance. Our Methodology Begins With An SVM Utilizing TF-IDF Representations For Initial Assessment, Then Forwards Uncertain Predictions To A Fine-tuned BERT Module For Comprehensive Semantic Evaluation. By Restricting Intensive Processing To Ambiguous Cases Only, This Strategy Minimizes Unnecessary Computational Expense. Evaluation Was Conducted On The Davidson Et Al. Benchmark Dataset Containing Three Categories: Hate Speech, Offensive But Nonhateful Language, And Neutral Messages. Our Experimental Findings Demonstrate That The Proposed Framework Achieves 93.2% Classification Accuracy With 30 Ms Mean Inference Time, Approximating BERT S Performance While Operating Approximately Five Times Faster Than Standalone Transformer Implementations. Analysis Of Prediction Confidence Levels, Processing Time Distributions, And Misclassification Patterns Validates The Architectural Design. Results Indicate That Hybrid Systems With Adaptive Routing Offer A Practical Solution For Deploying Sophisticated Hate Speech Detectors Under Real-world Operational Constraints.

Published:

12-5-2026

Issue:

Vol. 26 No. 5 (2026)

Page Nos:

1093-1098

Section:

Articles

License:

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

How to Cite