Hyperparameter and model configurations.

The convergence of Metaverse technologies, Internet of Things (IoT), and consumer electronics has given rise to an imperative need for scalable, real-time sentiment analysis that can process heterogeneous, high-velocity media flows. The traditional approaches tend to fail in preserving the contextual, emotional, and temporal dynamism that pervades cross-platform settings. For these shortcomings, this work proposes a deep learning-based framework for sentiment analysis that integrates IoT-enabled consumer devices and Metaverse media interactions seamlessly. The overall BG-Hybrid model, fundamentally, blends BERT-led bidirectional encoding and GPT-based generative modeling to attain subtle emotion detection and context-aware comprehending. The five interconnected modules constituting the architecture include (i) multi-source data collection using RESTful APIs; (ii) weighted preprocessing pipelines using tokenization, lemmatization, and normalization; (iii) Adam algorithm-optimized model training and cross-entropy loss minimization-based training; (iv) adaptive real-time processing using dynamic window segmentation; and (v) an ongoing refinement loop using continuous user inputs, triggered by a feedback mechanism. Predictive thresholding is employed to manage temporal sentiment variations, and anomaly detection ensures data trustworthiness. Experimental analyses on Twitter Sentiment140 and Amazon Reviews datasets validate the effectiveness of the system, obtaining 94.5% accuracy, 91.5% F1-score, an average response latency of 250 ms, and proved scalability exceeding 91.5%.