Bidirectional LSTM for Spam Detection and Sentimental Analysis

Short Message Service (SMS) and email communication have become primary vectors for spam, placing heavy burdens on users and mobile network operators. This paper proposes a Bidirectional Long Short-Term Memory (BiLSTM) deep learning model for spam detection and sentiment analysis, evaluated on three benchmark datasets: SpamAssassin, SMS, and Email. The model is compared against a Hybrid K-Nearest Neighbors and Support Vector Machine (Hybrid KNN-SVM) classifier from the prior literature. Preprocessing involves stemming, tokenization, and stop-word removal, followed by Word2Vec-based feature extraction. The BiLSTM network captures both past and future contextual information in text sequences, substantially outperforming the hybrid baseline. On the SpamAssassin dataset, BiLSTM achieves an accuracy of 98.77%, and on the Email dataset it reaches 99.11%. Sentiment polarity is classified using AFINN and SentiWordNet lexicons. Experimental results confirm that the proposed BiLSTM model yields superior accuracy, recall, F1-score, Kappa statistics, MAE, and RMSE across all three datasets.