Recent advances in natural language processing (NLP) and computer vision (CV) have accelerated research on multimodal intent recognition, aiming to integrate text, audio, and video modalities to better understand human intentions. However, existing mu...
Recent advances in natural language processing (NLP) and computer vision (CV) have accelerated research on multimodal intent recognition, aiming to integrate text, audio, and video modalities to better understand human intentions. However, existing multimodal intent recognition datasets are predominantly English-based, limited in scale, and often fail to capture the complexity of real-world conversational flows. Particularly, there is a significant lack of multimodal datasets applicable to Korean, hindering the practical expansion of Korean intent recognition technologies. In this study, we constructed K-MIR, a Korean multimodal intent recognition dataset, by benchmarking the English-based MIntRec dataset and utilizing real-world broadcast content. K-MIR comprises 2,270 utterances across 20 intent classes and includes text, audio, and video modalities. For data construction and feature extraction, tools such as Whisper, PySceneDetect, and YOLOv8 were employed, and labeling was conducted using a majority voting scheme to ensure reliability. To evaluate the quality of K-MIR, we conducted experiments using the MAG-BERT model, which not only confirmed that multimodal fusion yields significant performance improvements over text-only inputs, but also demonstrated that the same model architecture performs stably and achieves comparable benchmark-level results on K-MIR, thereby validating its practical utility and reliability. In parallel, to address the structural complexity and computational inefficiency of existing Transformer-based models such as MAG-BERT and MULT, this study proposes three lightweight multimodal classification models: FusionBERT, GatedFusionNet, and AlignContrastiveNet. These models were integrated using a 3-Way Soft Voting ensemble strategy as a practical approach to enhancing performance. To enable fair comparison with previous studies and assess the generalizability of the proposed architectures, training and evaluation were conducted using the publicly available MIntRec benchmark dataset, and the results consistently outperformed baseline models. In particular, the proposed hybrid ensemble model achieved an accuracy of 74% and a macro-F1 score of 0.69, surpassing the performance of MAG-BERT. In addition, the contrastive learning-based hybrid structure was designed to align semantically similar cross-modal representations within a shared embedding space, while the Soft Voting ensemble effectively leveraged the complementary strengths of individual models to further enhance overall performance. This study made three key contributions: providing a practical data resource (K-MIR) for Korean-language multimodal intent recognition, developing lightweight model architectures with enhanced performance, and optimizing modality fusion strategies. These contributions establish a solid foundation for advancing the practicality and future scalability of Korean multimodal AI systems.