A Hybrid Schema: LSTM-BiLSTM with Attention Mechanism to Predict Emotion in Twitter Data

Abstract

On the World Wide Web, people may now express their views and opinions in a novel way on a wide variety of subjects, trends, and issues, which includes advertising, political polling, knowledge-based surveys, market prediction, feeling and business intelligence. The user-generated content is available on various platforms, such as internet forums, discussion groups, and blogs, which serve as a concrete and significant basis for decision-making. Analysis of Emotion deals with the issue of extracting feelings from internet-based text data and classifying the author’s reactive mental reply as fear, anger, happiness etc. The underline research proposes a sequentially appended Deep Neural Network architecture to bridge the gap between previous approaches such as Maximum Entropy, Gradient Descent, Random Forest, Na¨ıve Bayes, and SVM(Support Vector Machine) used in machine learning. The model uses a balanced dataset to achieve enhanced accuracy and scalability. In the proposed architecture the first layer is the LSTM layer, which is used to process and sustain data in sequence for a long time. In the second layer, Bi-LSTM is appended for processing the flow of information in forward (past-directed to future)and backward (future-directed to past) directions and attached with an attention mechanism for predicting the output. The proposed framework is evaluated by Utilising various matrices, including the confusion matrix, recall, precision, and F-measure. Consequently, and is compared with the balanced dataset after handling the imbalance issue of different classes in the dataset. The model outperformed the actual dataset, which only had an accuracy of 90.87%, and reached a high accuracy of 96.53% in the sampled dataset.