A comparative evaluation of machine learning and deep learning algorithms for question categorization of VQA datasets

Question classification primarily involves categorizing questions based on the type of answer, with less emphasis on the words or phrases used to form the query. Question classification is crucial in the Visual Question Answering (VQA) system, and the dataset's quality plays an essential role in the system's development. The available question categorization in the VQA and TDIUC datasets shows imbalance, and the VQA model trained on imbalanced datasets performs poorly in handling language-prior problems, failing to categorize questions, and predicting incorrect outcomes. Therefore, developing a better classification method for classifying questions into appropriate categories based on phrases is necessary. This paper examines the effectiveness of the synthetic minority oversampling technique (SMOTE) in addressing the class imbalance problem within the question classification task using the LSTM, selected machine learning models and BERT-based transformer model. The preprocessing and analysis module efficiently categorizes input question sets by identifying valuable phrases and obtaining an evenly distributed dataset based on question categories from both datasets. The performance evaluation of Naive Bayes, SVM, Random Forests, and XGBoost models shows that the XGBoost model outperforms other selected classifiers, and the LSTM model achieves higher accuracy but requires more computation time. The empirical assessment indicates that the BERT-based transformer model exceeds the traditional models employed for comparison. The ablation study also reveals that utilizing SMOTE techniques for question classification tasks achieves slightly improved accuracy at the expense of higher computation time and resources. It is concluded that the BERT-based transformer model efficiently and precisely performs question classification tasks.