Contextual Features and Optimal Hierarchical Attention Networks for Sentiment Classification Under Data Streaming Environment

The endemic growth of online reviews has attracted the users to post their opinions in the social network. This growth made the classification of sentiment process an interesting domain in both industrial and academic research. Various sentiment classification techniques are developed to perform sentiment analysis, but the acquisition of sentiment grade is not precisely performed. Hence, this paper devises novel optimization driven classifier for classifying the sentiment grades. Here, the reviews are considered wherein the features are mined using reviews. In addition, significant features, like, SentiWordNet-based features, statistical features, context-based features, and Term Frequency-Inverse Document Frequency (TF-IDF)-based features are obtained from review. These features are adapted in Hierarchical Attention Network (HAN) to categorize sentiment grade. The training of HAN is performed using the proposed Competitive Swarm Water Wave Optimization (CSWWO) algorithm. The developed CSWWO algorithm is newly designed by integrating the Competitive Swarm Optimizer (CSO) and Water Wave Optimization (WWO) technique. Thus, the proposed CSWWO-based HAN model categorizes the sentiment into five classes, poor, better, good, very good, and excellent. At last, data stream handling is performed by concept drift detection and prototype-based adaption. Hence, the proposed CSWWO-based HAN offered enhanced performance with maximal accuracy of 93%, minimal True Positive Rate (TPR) of 93.9%, and maximal True Negative Rate (TNR) of 91.7%.