인공지능을 이용한 인간의 감정 인식 연구

초록

This study developed an emotion recognition model using voice, facial expression, and text information to enable artificial intelligence (AI) to more accurately recognize human emotions. Predicting human emotions such as joy, neutrality, sadness, and anger is highly important in both industry and academia, as it helps build empathetic, adaptive, and intelligent systems that enhance human well-being, communication, and decision-making. In this study, a video dataset from AI Hub was utilized to train the model on four emotions—joy, neutrality, sadness, and anger—by applying the HuBERT, ArcFace, and Llama3-8B models to each modality. The data used in this research referred to the emotion-related video dataset provided by AI Hub, which includes resources for emotion recognition, entity recognition, gender/age identification, relationship analysis, multimodal video question answering, single-utterance intent analysis, multi-utterance intent analysis, and speech recognition. The results showed that the text-based model achieved overall higher emotion prediction accuracy compared to the voice and facial expression models. The final emotion prediction accuracies of each model were 39% for voice, 41% for face, and 45% for text. These figures represent the accuracy of emotion classification on the final test set, which was prepared separately from the training data. This study is significant in that it demonstrates the importance of emotion recognition models and empirically verifies both the potential and limitations of incorporating context as an additional modality.

키워드