Sentiment Analysis of Online Streaming Reviews for Movies Using the Dual LSTM-CapsNet Model

Sentiment analysis (SA) extracts customer opinions from reviews, but traditional word-level and even deep learning (DL)-based phrase- or sentence-level methods struggle to capture complex semantics and long-term dependencies. These challenges hinder the effective extraction of contextual information. This research proposes a novel DualLS TMCapsNet model for SA of movie reviews, addressing the limitations of traditional methods in capturing complex semantics and long-term dependencies. The model begins by converting raw text into numerical vectors through embedding, ensuring efficient processing of short text samples. Dropout is used during training to avoid overfitting. The second layer uses a bidirectional Long Short-Term Memory (BiLS TM) system, which captures the text's past and future contextual dependencies. Finally, a Capsule network refines the output of the LSTM, enhancing the model's ability to recognize hierarchical patterns before making the final sentiment classification. This approach improves the model's handling of complex sentiment contexts and relationships in movie reviews. The proposed SA model was used to assess the cleaned IMDb dataset, which achieved an impressive 95% accuracy, outperforming existing methods. These results underscore the efficacy of DL models to improve SA and demonstrate promise for wider uses in natural language processing (NLP).