Implementasi Algoritma Convolutional Neural Network (cnn) Pada Evaluasi Strategi Prompt Engineering Terhadap Akurasi Visual Question Answering (vqa)
Year: 2026
Type: article
Abstract: Visual Question Answering (VQA) is the task of answering natural language questions about the content of an image. This study aims to identify the prompt engineering strategy with the highest VQA accuracy and to test whether the accuracy differs significantly among strategies. The model is a CNN–Transformer architecture with ResNet-50 as the visual encoder, BERT as the text encoder, and a cross-attention fusion layer. Five strategies were evaluated: zero-shot, few-shot, chain-of-thought, instruction-based, and role-prompting. The data consist of 8,000 questions from the VQA v2.0 validation set selected by stratified random sampling. Each strategy was trained with five random seeds and evaluated using VQA Accuracy and One-Way ANOVA. The results show that the instruction-based strategy achieved the highest mean accuracy of 37.54%, while the other strategies ranged from 36.63% to 37.17%. However, the differences were not statistically significant (p = 0.2897). The instruction-based strategy was then implemented in a web application prototype.
Institution: North Sumatra Islamic University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7235912431
