PCA-Driven Quantum LSTM with Stacked Ensemble Sentiment Analysis for Cyber-Bullying Detection and Hate Speech Classification

Abstract

The rapid growth of user-generated content on social media platforms has reinforced the urgency of detecting abusive language automatically. This includes identifying cyber-bullying, hate speech, and toxic sentiment. This paper introduces PCA-QLSTM-stacked ensemble (SE), a unified hybrid framework that combines principal component analysis (PCA) based quantum long short-term memory (QLSTM) neural network. The framework is optimized by black-winged kite optimization (BWKO) and augmented by a stacked ensemble of support vector machine (SVM), random forest (RF), and extreme gradient boosting (XGBoost [XGB]) classifiers. Firstly, the Text features are extracted using a hybrid term frequency–inverse document frequency (TF-IDF) and multi-scale hash vectorization (HV) pipeline. The features are then reduced by PCA retaining 95% cumulative variance, before classification by the hybrid model. Experiments across five benchmark datasets (internet movie dataset [IMDB], Yelp Polarity, Cyber-bullying Classification, Kaggle Hate Speech, and a combined corpus) demonstrate the accuracy of 96.41%, 93.87%, 98.12%, and 96.80% respectively. Ablation studies confirm the independent contribution of each component, and Friedman-Nemenyi statistical tests validate significance at alpha = 0.05.