CNN-Based Hybrid Performance Evaluation Towards Online News Sentiment Classification Task

CNN is a deep learning model that is effective in extracting features in text data. However, CNN has shortcomings in understanding long-term context and lacks sensitivity to word order. In this study, a CNN model was hybridized with a machine learning model to improve the model's performance. The LSTM model and BiLSTM model were chosen because they have long-term memory features and sensitivity to word order. Apart from that, other machine learning models such as TF-IDF, Word2Vec, and BERT are used to add features that are useful in extracting text data. In model testing, the accuracy, precision, recall, F1-score, and AUC-ROC matrices are used. The evaluation results of all models are compared to determine the influence of vector dimensions and obtain the best model. From the evaluation results, it was found that the RoBERTa-CNN-BiLSTM model had the best performance with a matrix accuracy of 98.18%, precision 98.19%, recall 98.18%, F1-score 98.18%, and AUC-ROC 99.86%. Apart from that, it is known that the vector 38 dimension provides the most superior performance. These results indicate that hybridized deep learning models, especially CNNs, are effective in improving model performance.