Estimating Viewer Emotions Using Comments on Online Video-sharing Services

Abstract

The rapid proliferation of internet access and smartphones has significantly increased engagement with video-sharing platforms like YouTube and TikTok. These platforms have emerged as powerful influencers of individual behavior and societal trends, particularly in the domains of marketing and consumer decision-making. Understanding the emotional responses elicited in viewers during video consumption holds substantial value for content creators and marketers. This study presents a novel approach for estimating viewer emotions based on user-generated comments associated with online videos. By leveraging Bidirectional Encoder Representations from Transformers (BERT) and various large language models (LLMs), the emotion estimation capabilities of different models are systematically compared. Emotions are represented as a seven-dimensional vector, where each component reflects the intensity of a specific emotional state. Experimental evaluations reveal that LLMs perform better in estimating nuanced emotional intensities when a larger set of comments (e.g., 100) is available. Conversely, BERT performs better in identifying the predominant emotion when only a limited number of comments (e.g., 10) are analyzed. These findings highlight the complementary strengths of different models and offer practical insights into emotion estimation from social media data.