Enhancing products performance evaluation through hybrid DistilRoBERTa and BiGRU models.

Online reviews have a direct bearing on what prospective customers will choose to buy. Consumers need help to make effective use of online reviews for purchasing decisions. The current sentiment analysis methods often overlook the complexity and usability of products in reviews which are critical for potential buyers. The existing approaches solely rely on sentiment polarity but this approach offers a comprehensive way to analyze product' reviews by integrating appraisal theory with hybrid deep learning methods. This research study aims to explore product complexity in customer reviews by utilizing transformer and recurrent neural network-based models. Amazon product reviews dataset is manually annotated according to perceived complexity using attitudinal categories of appraisal theory, i.e., appreciation and judgment and then cross-validated through shapley additive explanations method of eXplainable AI. Appreciation has to do with how much a product is user-friendly or has a learning curve to operate. Judgment deals with evaluation of products on the basis of effectiveness and appropriateness. The proposed model uses a hybrid approach by combining DistilRoBERTa, a pre-trained transformer model, and a bi-directional gated recurrent unit, a recurrent neural network-based model, to learn both contextual and sequential dependencies in product reviews. The proposed approach makes use of fine-tuned DistilRoBERTa embeddings for extracting features, and this model is improved further using a bi-directional gated recurrent unit layer, which takes into account the context of the past and the future. A 5-fold stratified cross-validation method is used to address imbalance learning, with class weighting applied to further balance the impact of sentiment classes in training. The proposed model has achieved a mean fold accuracy of 96.13%,which is higher than that of existing state-of-the-art approaches such as Random Forest (89.93%) [11], DistilBERT with advanced embeddings (92%) [25], XLNet (89.62%) [28], and GPT-based sentiment models (94.5%) [30]. By utilizing Shapley Additive Explanations (SHAP) for explainability, this model provides transparency in understanding emotional tendencies, functional effectiveness, and overall user perceptions, offering insights that traditional models lack. This framework provides a scalable, automated solution for Amazon products' performance evaluation and provides insights into emotional tendencies, functional effectiveness and overall perceptions from the users. It will help in optimizing product development strategies using advanced natural language processing techniques with explainability by setting a new standard for understanding products'product feedback.