Adaptive bi-directional long short-term memory-based sarcasm detection on social media platforms

Social networking is becoming a fantastic resource for gathering feedback from users. As technology advances and the internet becomes more sophisticated, an enormous amount of data is generated by various sources, including websites, social media platforms, and blogs. The proliferation of an excessive number of blogs in the cloud has enabled the production of a vast amount of information in various formats, including opinions, reviews, and personal perspectives. People on social networking sites commonly use sarcasm to communicate their thoughts, which is difficult to analyze, not just for an automated system but also for a person. However, the sarcastic reviews are confusing to predict the class of the review. To avoid the problem, in this paper sarcastic detection-based sentimental analysis is proposed. The suggested method is divided into three stages: processing, extraction, and prediction. To begin, reviews are gathered and pre-processed with tokenization, stemming, stopword removal, and word removal. After pre-processing, feature extraction is performed. For feature extraction, ten sets of features are extracted, namely, the weighting of two unsupervised and eight supervised term methods are utilized. Subsequently, these extracted features are input into the adaptive bi-directional long short-term memory (ABi-LSTM) model to determine whether a product review is sarcastic or not. To improve the classifier, the hyperparameters have been optimized using the adaptive red fox algorithm. The proposed approach's performance is evaluated using several measures and compared to state-of-the-art methodologies. Python is used to implement the proposed method. The proposed ABi-LSTM + Term Weighted model outperforms other methods across all performance metrics. It achieves the highest accuracy (97.487%), sensitivity (97.435%), and specificity (97.537%), along with top precision, recall, and F1-score values. Additionally, it records the lowest false positive rate (2.463%), false negative rate (2.565%), and error rate (2.513%), demonstrating superior reliability and classification strength in sarcasm detection compared to LSTM, RNN, and CNN-based alternatives.