Deep Convolutional Networks for Affective Content Extraction from Textual Sources

Abstract This paper proposes a Convolutional Neural Network (CNN) based system for classifying emotions in text data. The system tackles the challenge of automatically recognizing emotions expressed within textual content. The paper presents a comprehensive approach encompassing data preparation, model building, training, evaluation, saving, and prediction. The data preparation stage involves cleaning, balancing for imbalanced datasets, and pre-processing techniques such as tokenization and padding. The CNN model employs two branches with similar CNN layers for feature extraction. These features are then concatenated to create a richer representation. Dropout techniques are utilized during training to prevent over fitting. The system's performance is evaluated using various metrics like accuracy, precision, and recall. We visualize the learning process through training and validation curves. Additionally, confusion matrix and classification reports are employed to analyze the model's performance. Finally, the trained model and tokenizer are saved for future prediction tasks. This work offers a comprehensive framework for text emotion classification using CNNs, demonstrating its effectiveness in identifying emotions from textual data.