OSA-UER: An Optimized RNN-Based Sentiment Classification Framework Using GloVe Word Embeddings

Abstract

The ability to turn extensive online reviews from customers into comprehensible insight on sentiment works for high-quality decision making in service-focused platforms. Traditionally deep learning models like CNNs, vanilla RNNs and LSTMs usually fail to capture long sequential dependencies and sentiment flow for long textual reviews. In order to overcome these pitfalls, this paper implements an Optimized-Sentiment Analysis (OSA-UER) framework with Enhanced Recurrent Neural Network (ERNN) architecture for multi-class opinion mining of customer feedback analysis. This framework combines word embeddings with a GloVe trained embedding layer, preprocessing, attention and a sequence feature extraction layer with a stacked powerful sequential architecture with hierarchical Bidirectional GRUs 128–256–512 units, Dropout regularisation, attention for contextual modelling and generalisation. The Zomato customer review dataset is collected from the internet and is divided into test and train sets with a simple supervised learning protocol to train and test the mentioned model. The Adam optimizer with categorical cross-entropy loss is used for training, while accuracy, precision, recall, and F1-score are utilized for evaluation. While experimental results show that the highest accuracy achieved by Enhanced RNN is 98.91% — a greater performance than the baseline CNN (92.17%), standard RNN (94.83%), LSTM (96.17%), and ResNet-50 (89.61%) cannot be compared reliably under the same experimental conditions — We can see that the training and validation curves are very close to one another which indicates that the learning behaviour is stable and the generalization is strong. These results validate the use of bidirectional sequential modeling and attention-based token weighting provide more optimum integration for sentiment discrimination than traditional architectures. Our framework can be applied to customer feedback mining with high scalability and interpretability for practical applications in sentiment caption.