Optimal Feature Selection and Hierarchical Attention Networks for Text Classification

Abstract

In the current era, with the advancement in technologies, an enormous quantity of data is generated daily. Due to the permanently growing amount of textual data, automatic methods for organizing text document into a group of classes depends on the contextual information in the document. Text classification has been employed in large number of applications, like content moderation, user intent classification, news categorization, sentiment analysis, spam detection, Question Answering (QA), and so on. In this research, multiple text classification techniques are proposed. The principal contribution of this research is the conception of an algorithm named Tunicate Swarm Algorithm based Hierarchical Attention Network (TSA-HAN) for the classification of textual data. In this algorithm, the Tunicate Swarm Optimization Algorithm (TSA) is utilized to choose the highly appropriate features applied to the Hierarchical Attention Network (HAN). The next contribution is the development of another text classification approach utilizing the Improved Invasive Weed Optimization-based HAN (IIWO-HAN). Like the previous method, the optimization algorithm named IIWO was employed to find the prominent features in the input text data and HAN trained by the IIWO is applied for classifying the text. The third contribution is the Invasive Weed Tunicate Swarm Optimization-based HAN (IWTSO-based HAN) for text classification, wherein a novel optimization technique named IWTSO is formulated by adjusting the IIWO based on TSA. Here, the IWTSO is employed for feature selection and the weight optimization of the HAN, which classifies the document depending on the selected features. The last contribution is prepared in the form of the Henry Fuzzy Competitive Multi-Verse Optimizer (HFCVO)-based Deep Maxout Network (DMN) for incremental text categorization. In this method, the IWTSO selects the ideal features that are fed to the DMN for classification.