Audience perception and decoding of Chinese cultural content in global cross-media context based on LLM-enhanced temporal topic modeling

Abstract

Methodological limitations in cross-cultural social media research—temporal blindness, semantic misalignment, and data sparsity—constrain understanding of how global audiences decode Chinese cultural content. This study integrates Large Language Models with temporally-weighted topic modeling to examine non-Chinese audience reception of Chinese cultural videos on YouTube within an algorithmically-mediated environment. Using an LLM-enhanced Seeded-LDA model (k = 11), we analyzed 24,847 non-Chinese comments from high-engagement videos (2014–2025). DeepSeek-v4-flash facilitated multilingual preprocessing, comment filtering, and sentiment analysis. Temporal evolution was tracked via Dynamic Topic Modeling, and reply-thread networks were analyzed to detect politicization spillover. Eleven topics emerged, organized into four decoding positions: algorithmic empathy (travel and food content, >60% positive), value friction, civilizational comparison, and ideological contestation (narrative framing, >50% negative). Network analysis revealed platform-mediated politicization spillover in one-third of high-engagement threads. Three temporal regimes were identified: aesthetic curiosity (2014–2018), polarization (2018–2022), and complex bifurcation (2023–2025). Platform affordances and source attribution cues appear systematically linked to divergent decoding patterns, with self-media triggering empathetic responses and official media activating ideological contestation. The proposed computational framework advances large-scale analysis of platform-mediated cultural communication, though observational constraints limit causal claims.