Feature classification by free description analysis using machine learning

Abstract

Various methods have been explored for social media text analysis, including sentiment classification, spam detection, and identifying inflammatory videos. However, the format of free text makes it difficult to obtain consistent analysis results. To address this, sentiment-based analysis and classification are conducted. The accuracy of analyzing videos using multi-class sentiment estimation is considered lower compared to using negative and positive sentiment models. To improve accuracy, the WRIME dataset is used, and an emotion intensity estimation model is generated by retraining a pre-trained BERT linguistic model. The study aims to extract objective features from subjective free text data using machine learning and provide classifications and discussions. The experiment focuses on comments from Vocaloid producers' YouTube videos, and the results indicate the effectiveness of 8 emotion estimation for video analysis, suggesting potential applications in emotion-aware recommendation and analytics systems.