Does Entropy in the Sentiment of TikTok Videos Point to Polarisation?

The next speaker in this ECREA 2024 session is Petro Tolochko, whose focus is on affective polarisation in climate activism visuals. Such content can be highly affective in climate activist communication, spark audience reactions, and spread online to promote the emergence of like-minded or opposing groups. The analysis here might include aspects of structural polarisation (using network analysis) and reactionary polarisation (using communication analysis).

An initial question might thus be which types of images lead to increased polarisation online; more recently, however, with the shift from Xitter to TikTok the role of videos in such activist communication has grown. Polarisation might then be studied using a measure of the emotion entropy within a given comment thread: how strongly do emotions in different comments swing between extremes from angry to happy (indicating more polarisation), or how consistently are they angry, neutral, or happy (indicating less polarisation)?

The project explored this with some 10,000 TikTok videos on climate change, attracting some 1 million comments from Austria and Germany, with texts in both German and English. Emotion was classified using a local Llama 3 model, using anger, disgust, fear, happiness, neutral, sad, and surprised emotions. Anger, surprise, and happiness were most prevalent here.

The videos themselves were also examined, using keyframe extraction and feature analysis; based on this analysis the videos were then clustered based on their similarities. This resulted in nature, protest, politics, and talking heads videos.

Emotional entropy was broadly similar across these different video types; there were slightly greater differences based on hashtags, but they are still rather similar in their entropy measures. Further and more sophisticated data processing may be required here, for both videos and comments.