An AFFA-Integrated Hybrid Deep Learning Framework for Explainable Sentiment Analysis

In recent years, sentiment analysis has become a prominent research area in natural language processing, with advances in deep learning significantly improving multiclass sentiment classification through enhanced modeling of contextual and semantic relationships. In this study, a deep learning–based sentiment analysis framework is evaluated on multisource English user reviews. Sequential representations derived from FastText word embeddings are used to assess baseline recurrent models (RNN, LSTM, Bi-LSTM, and GRU), attention enhanced hybrid architectures, and the proposed Adaptive Fine-Grained Feature Aggregation (AFFA) model. To improve data quality and label consistency, potentially mislabeled instances are identified and removed using a pretrained CardiffNLP Twitter-RoBERTa–based sentiment classifier. The AFFA architecture adaptively fuses representations from multiple encoders by learning encoder-specific contribution weights, which are normalized through a Softmax function to dynamically emphasize complementary feature representations. The experimental results indicate competitive class-wise discrimination performance, with modest improvements observed in selected metrics for the neutral sentiment class. The AFFA–LSTM (3 Layer)–Attention model achieved the best overall performance, obtaining 94.60% Accuracy, 94.52% Macro-F1, and a mean Macro ROC–AUC score of 99.19% ± 0.05 across five independent runs. These results suggest that adaptive feature aggregation provides an effective mechanism for improving both robustness and reliability in sentiment analysis tasks. Explainability is illustrated through representative qualitative analyses based on Integrated Gradients–based token attribution and AFFA fusion-weight analysis, providing illustrative insights into token-level and encoder-level contributions.