Abstract
This study investigates the structural and semantic characteristics of political polarization on TikTok during the 2024 elections. By harvesting a dataset of over 150,000 user comments across 5,000 highly engaged political videos, we employ a hybrid methodology combining Natural Language Processing (NLP) and Social Network Analysis (SNA). Using a fine-tuned RoBERTa model for sentiment classification and BERTopic for thematic modeling, we map the semantic landscape of the discourse. Concurrently, we construct user-interaction networks based on comment-reply structures to quantify community segregation. Our structural network analysis reveals a high modularity index (Q = 0.74), indicating the presence of deeply entrenched, isolated partisan communities. Semantic analysis demonstrates that while intra-community interactions are characterized by supportive, homophilous sentiment, inter-community dialogue is highly adversarial, marked by elevated levels of toxicity and negative sentiment. Furthermore, topic modeling reveals that algorithmic curation on TikTok reinforces these boundaries by clustering users around emotionally charged, highly specific wedge issues. These findings provide empirical evidence of how TikTok’s unique interest-based recommendation system fosters robust echo chamber dynamics, posing significant challenges for democratic deliberation and digital inclusion in the modern socio-technical landscape.
