Bullet Screen is a novel innovation allowing viewers to watch internet films and engage in real-time comment reading and sharing opinions at the same time. This paper aims to adopt machine learning algorithms to develop a model based on Bag of Words to conduct sentiment analysis on the reviewer’s comments on the bullet screen. This paper adopted an experimental design with a dataset of 6,300 valid comment datasets retrieved from bilibili.tv. This research used accuracy and loss as performance metrics and compared the performance of the BOW model with the performance of the CNN and RNN models. The result suggested that the accuracy of the BOW model is the highest among all the three compared models, and it has low losses in its performance. This paper contributes new findings to Natural Language Processing (NLP) in social media, enriches the analytical approaches for sentiment analysis on audience expression online, integrates AI in social media, and drives further business innovation in the online video sector. Finally, a methodological limitation of this research is that the measuring metrics used did not incorporate precision and recall. Furthermore, the utilisation of BOW disregards both the arrangement of words and the surrounding context. Furthermore, our scope constraint fails to take into account the utilisation of emoji expressions. In the coming years, the authors intend to train and evaluate the precision of the model using various segments of social media sites. Additionally, authors endeavour to develop a model utilising the CNN architecture to accurately detect emotional expression through emotion classification, based on the findings of current research.
Wang et al. (Tue,) studied this question.