With the advent of the 21st century, advancements in the IT industry, alongside the proliferation of smartphones and the internet, have brought significant changes to people's daily lives and consumption of culture. Particularly, platforms like YouTub...
With the advent of the 21st century, advancements in the IT industry, alongside the proliferation of smartphones and the internet, have brought significant changes to people's daily lives and consumption of culture. Particularly, platforms like YouTube have disrupted the traditional paradigm of broadcast content production, opening new realms for individual creators and small-scale content production. These shifts have led to a wide consumption of diverse contents alongside the innovation of streaming services and the emergence of OTT (Over-The-Top) platforms, enabling both corporations and individual creators to engage in active marketing and content creation.
However, this modern social trend has not only benefitted OTT platforms. Notably, terms like 'poverty in abundance' and 'Netflix Syndrome' have emerged around Netflix, highlighting a phenomenon where the expansion of choices in movies and dramas induces users' decision-making time and mental stress. Additionally, the surge in small-scale content production on social network services like YouTube has increased the tendency of users to refer to relatively shorter review videos than the somewhat lengthy movies or dramas offered on OTT platforms.
The popularity of small-scale content production and short review videos reflects users' emotions and preferences. Through emotion analysis, it's possible to understand these trends and establish more effective content production and marketing strategies.
This study utilizes the BERT – BASE version of the BERT – Multilingual model to transcribe voices from videos containing YouTube creators' poetic interpretations into subtitles and experiments with emotion analysis through dual fine-tuning without pre-training, using comments from viewers. The emotion analysis execution mechanism is divided into two experiments: the first using comments and the second using subtitles, involving data exploration, preprocessing, model training, and performance evaluation based on accuracy.
The study examines the correlation of labels using binary methods based on the emotion analysis of sampled subtitles and comments. For instance, subtitles of a 'cohabitation drama review' video predominantly showed 'jealousy' (Emotion Label, 31), while the comments reflected 'satisfaction' (Emotion Label, 54) and 'excitement' (Emotion Label, 55). This difference is attributed to factors like content and user response disparity, the diversity and subjectivity of emotions, and the dramatization and direction style. This natural variance between the emotions in drama subtitles and viewer comments illustrates how each individual's unique experiences, interpretations, and responses generate diverse emotional reactions.
Particularly, performance evaluation results can be qualitatively compared with related studies. Despite the same learning environment, an increase in accuracy by 0.43% was proven, and BERT demonstrated somewhat higher performance through dual fine-tuning without special pre-training.