Abstract: | With the recent development of deep learning technology comes the wide use of artificial intelligence (AI) models in various domains. AI shows good performance for definite-purpose tasks, such as image recognition and text classification. The recognition performance for every single task has become more accurate than feature engineering, enabling more work that could not be done before. In addition, with the development of generation technology (e.g., GPT-3), AI models are showing stable performances in each recognition and generation task. However, not many studies have focused on how to integrate these models efficiently to achieve comprehensive human interaction. Each model grows in size with improved performance, thereby consequently requiring more computing power and more complicated designs to train than before. This requirement increases the complexity of each model and requires more paired data, making model integration difficult. This study provides a survey on visual language integration with a hierarchical approach for reviewing the recent trends that have already been performed on AI models among research communities as the interaction component. We also compare herein the strengths of existing AI models and integration approaches and the limitations they face. Furthermore, we discuss the current related issues and which research is needed for visual language integration. More specifically, we identify four aspects of visual language integration models: multimodal learning, multi-task learning, end-to-end learning, and embodiment for embodied visual language interaction. Finally, we discuss some current open issues and challenges and conclude our survey by giving possible future directions. |