ESFMS: Design of an Ensemble Sentiment Analysis Model for Feedback Evaluation via Multimodal Feature Selection Process

Sentiment analysis for identification of feedback is a multi-domain language processing task that involves pre-processing of text samples, representation of pre-processed samples into feature sets, selection of these features, their classification into different sentiment categories, and post-processing to obtain final feedback classes. Researchers have developed a wide range of sentiment analysis models to carry out this task, and the majority of them are either quite sophisticated or perform less well when tested on multiclass feedback. These models’ need for context-specific inputs further restricts their capacity to scale for real-time use cases, and therefore in order to get high performance, a novel ensemble sentiment analysis model for feedback evaluation via a multimodal feature selection process is proposed. The model initially collects entity-level feedback and converts them into high-density feature sets via a customized Word2Vec process. This process uses context-specific dictionaries to train the Word2Vec model for extraction of application-specific feature sets. These feature sets are further processed via an extra trees classifier (ETC) for inter-sentiment variance maximizations. This assists the ensemble classifier to categorize the selected features into multiple feedback classes with high-accuracy levels. The ensemble classifier uses a combination of Naïve Bayes (NB), k-nearest neighbors (kNN), support vector machine (SVM), logistic regression (LR), and multilayer perceptron (MLP) techniques. Due to which the model is able to identify sentiment classes with 98.5% accuracy across heterogeneous feedback data samples. The model was also able to achieve 3.9% higher precision, 2.5% higher recall, and increase classification speed by 8.3% when compared with existing techniques.