Fig. 3

Our proposed architecture combines dynamically produced opinion embeddings with BERT embeddings. We integrate these opinion embeddings with BERT’s contextual representations, which are obtained via sentiment annotations. Our design uses a hybrid feature extraction technique to capture sequential relationships as well as local patterns. This mechanism consists of layers of CNN and BiGRU. We apply concatenation, dropout, and normalizing layers to the final representations, yielding outputs for status prediction and sentiment categorization.