Enhancing Sentiment Analysis through Hybrid Transformer Architectures and Contextual Embeddings

Abstract

Understanding emotions in text allows computers to grasp how people truly mean their words. Hidden sentiments within sentences reveal themselves through careful analysis, offering clarity in areas such as social media monitoring or feedback interpretation. A fresh technique appears here, rooted deeply in neural computation, combining stacked architectures with adaptive contextual reasoning to boost performance. Older methods take a back seat while varied strategies emerge - TF-IDF transforms appear next to Word2Vec mappings, GloVe embeddings sit beside dynamic representations drawn from BERT and RoBERTa - to explore which inputs strengthen prediction depth. This approach leads to a system named the Extended Transformer Belief Network (ETBN), shaped by layers that learn both structure and meaning across diverse linguistic patterns. This setup initially picks up structures using unlabeled data, relying on selective focus methods. Then, precision improves when real-world examples steer the adjustment phase. Meaning deepens step by step, mirroring actual human phrasing habits. Testing kicks off across IMDB, Sentiment 140, alongside Amazon feedback - performance reaches 97.5% correct, with mistakes only at 2.5%. Leading models based on transformers beat prior techniques, showing noticeable gains. Still further ahead, it positions new benchmarks alongside today’s hybrid models when identifying emotional states. Evidence appears clear: linking deep learning architectures with refined feature extraction pushes accuracy beyond conventional methods.