Abstract
Social networking platforms generate vast volumes of short, informal user-generated text that encode opinions and attitudes. Organizations seek to analyze this content to derive actionable insights, motivating robust sentiment-analysis methods. This study presents an empirical comparison of deep-learning approaches for sentiment classification of social-media text. We evaluate six standalone models—Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Recursive Neural Network (RecNN), Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and Gated Recurrent Unit (GRU)—and three hybrid variants (CNN–RNN, CNN–LSTM, CNN–BiLSTM). The corpus undergoes standard preprocessing (cleaning, tokenization, lemmatization, stop-word removal) before classification. Across the tested settings, the CNN–BiLSTM consistently achieves the strongest overall performance on key metrics (e.g., recall, F1-score, AUC-ROC), outperforming both standalone architectures and the other hybrids. These findings indicate that combining localized n-gram feature extraction with bidirectional sequence modeling yields a favorable balance of accuracy and robustness for short, noisy posts, and they offer practical guidance for selecting architectures for social-media sentiment analysis.