Deep learning based SentiNet architecture with hyperparameter optimization for sentiment analysis of customer reviews

Abstract

Abstract Due to the ever-increasing volume of opinionated content on social media, there is a pressing need for a highly effective process to perform sentiment analysis (NLP) successfully. But, methods nowadays are usually not good enough in context, long-term dependency, and domain-specific areas modelling, especially they lack in modelling of noisy, short texts. In this paper, we introduce SentiNet, a new hybrid deep learning architecture combining multi-layered BiLSTM encoders with convolutional feature extractors and an attention-based fusion mechanism, to achieve better performance in sentiment classification across various datasets. Our model incorporates embedding vectors, parallel convolutional layers, and channel-wise attention over embedded sequential data, as well as one or two bidirectional LSTM units to capture context over the sequence elements. We conduct extensive experiments on the IMDb, Twitter, and Yelp datasets, and show SentiNet yields the highest accuracy of 94.2%, best F1-score of 92.8%, and a balanced precision-recall curve, outperforming competitive baselines. To validate the contribution of each module, ablative experiments are performed, and cross-domain evaluations prove its robustness. The key contribution of this work is that it balances accuracy and interpretability within an efficient processing pipeline, making it applicable in real-world sentiment scenarios, such as e-commerce, customer experience monitoring, and social media analytics. This work extends scalable sentiment analysis by introducing a high-performing, interpretable, and adaptable framework, providing a strong foundation for future explainable AI advancements in text analytics. To enable reproducibility and follow-up work, the code and model will be released publicly.