Sentiment Data Analysis for Detecting Social Sense after COVID-19 using Hybrid Optimization Method

Abstract

Sentiment analysis is the most effective way to understand opinions presented in digital media. Social sense is the concept of understanding every user's emotions connected through an online social media platform. This emotion helps to understand the mental health of people. Every stage of extracting social impact contains several problems with several methods associated with it, like lexicon-based techniques, which were proposed, developed, and evaluated but sometimes had poor accuracy. Machine Learning (ML) is useful to recognize patterns, in various aspects of sentiment analysis and it will best suited for this traditional algorithm where the large dataset is chunked into the smaller dataset to train the model. NLP methods do not account for different domains when modeling sentiment information. In a classification dilemma, reducing the feature set size diminishes the algorithm's time demand while improving the method's accuracy to obtain the optimal features. Scalability refers to handling large amounts of data and performing several computations like time and cost efficiently. This paper's perspective is to overcome the shortcomings of each system, and this article proposes a hybrid approach to sentiment analysis. Data has been collected from Twitter via the Twitter API. This research investigates the effects of negations, URLs, usernames, punctuation, repeated character normalization, and hashtags on sentiment classification output using an n-gram representation model, which is a kind of probabilistic language model and also integrates a cross-domain sentiment-aware learning model that can collect both sentiment awareness and domain validity of a word simultaneously to solve the problem of words from various domains. To fix the scalability problem, the article offers the novel Tunicate Swarm Algorithm (TSA), which increases scalability and reduces computation time. Meanwhile, this paper custom a hybrid Harris Hawks Optimization (HHO) algorithm based on simulated annealing (SA) and bitwise operations to solve the local optima problem. This proposed work results in the sentiment score that people with positive sentiments are more than those with negative. The proposed model gives better feature size reduction measured using the correlation coefficient metric and comparing other state-of-the-art algorithms like the wrapper approach, particle swarm optimization, and greedy feature selection with a reduction in feature size of up to 64%. The precision–recall curve ranges toward one, which means the classifier does classification accurately. In the sentiment analysis of Twitter feedback, the suggested solution has greater exactness and scalability with 96 s. The proposed model is well-trained, and the best validation is achieved at the 4th epoch with 0.6672. The area under the ROC curve evaluates the separability near one that shows the proposed model works excellently. The claim of a proposed model has been validated through comparison from different classifiers like Naïve Bayes, kNN, Random Forest, CNN-RNN, and AC-BiLSTM with admissible accuracy of 96.37%.