EmoNet: Deep Attentional Recurrent CNN for X (Formerly Twitter) Emotion Classification

Emotion classification from social media data is critical for market research, sentiment analysis, and understanding human behavior, yet the unstructured nature of Twitter data poses significant challenges for conventional models. To overcome these challenges, we propose EmoNet, a novel Deep Attentional Recurrent Convolutional Neural Network (DAR-CNN) that integrates an advanced attention mechanism to capture intricate contextual dependencies in textual data. EmoNet utilizes an AttentionEmbedder module to generate query, key, and value vectors, which are then processed through an attention layer to compute weighted representations. These representations are concurrently fed into LSTM, CNN, and GRU layers to extract diverse sequential and spatial features, with global max pooling applied to each pathway before concatenation and final classification via fully connected layers. We employ K-fold cross-validation to rigorously evaluate EmoNet against baseline models, including RoBERTa, BERT, and DistilBERT. Our DAR-CNN model achieves a remarkable accuracy of 94.51% and an F1 score of 94.55%, outperforming existing state-of-the-art methods. These results demonstrate our effective combination of attentional mechanisms with recurrent and convolutional architectures, establishing EmoNet as a robust solution for emotion detection on social media platforms.