Application of a Neural Network-based Visual Question Answering System in Preschool Language Education

Abstract

The continuous progress of modern science and technology has led to comprehensive innovations in education, and the use of information technology for teaching has become the mainstream in the current education field. For children’s preschool language education, the application of a visual question answering (VQA) system has gradually become a new development power. This research uses a Recurrent Neural Network and a VGGNet-16 network to extract features from text and images, respectively, and applies a Hierarchical Joint Attention (HJA) model to the whole VQA system. Experiment results demonstrate that the HJA model reaches the target accuracy after 125 iterations, and convergence performance is good. When using the VQAv1 dataset, accuracy can stabilize at 88% after 18 iterations, and when using the VQAv2 dataset, the highest and lowest overall accuracy rates are 77% and 72%, respectively. The three question types (Num, Y/N, and Other) are answered with high accuracy when using the chosen preschool language education database for children, providing accuracy rates of 90%, 94%, and 91%, respectively. This new reference technique offers a new method for maximization of a VQA system, and significantly raises the preschool language education level of the children.