ReviewIQ: A Unified NLP, Deep Learning and Advanced Machine Learning Framework for Review Sentiment Classification, Aspect Intelligence and Grounded Question Answering

Abstract ReviewIQ is an integrated framework that combines natural language processing (NLP), deep learning (DL), and advanced machine learning (AML) to analyse Amazon Electronics product reviews. The system supports three-class sentiment classification, aspect-based sentiment analysis, topic modelling, review helpfulness prediction, suspicious-review anomaly detection, and product-scoped question answering grounded in retrieved review evidence. The framework evaluates VADER, TF-IDF-based Logistic Regression, a linear Support Vector Machine (SVM), a bidirectional Long Short-Term Memory (BiLSTM) network with attention, and fine-tuned DistilBERT using a shared test split of 10,000 reviews from a 50,000-review working sample. The BiLSTM achieves a macro-F1 score of 0.718 in the main comparison, while DistilBERT achieves the highest accuracy of 89.9%. Further experiments examine class weighting, attention pooling, computational cost, aspect-level sentiment, helpfulness prediction, and retrieval-based question answering with a similarity threshold to reduce unsupported responses. The aspect analysis identifies battery, connectivity, and customer service as prominent complaint areas. An XGBoost classifier predicts review helpfulness, while Isolation Forest identifies potentially suspicious patterns without claiming that flagged reviews are definitively fraudulent. The study also discusses limitations arising from weak sentiment labels, sampling bias, unequal model-training budgets, lexicon-based aspect extraction, and the need for further evaluation of generated answers. Keywords: Sentiment analysis, product review mining, BiLSTM, attention mechanism, DistilBERT, aspect-based sentiment analysis, XGBoost, Isolation Forest, semantic retrieval, grounded question answering.