Scalable Content Analysis of Social Media Videos Using Multimodal Large Language Models: A Video-to-Text Pipeline for Large-Scale Analysis

The analysis of large-scale video content remains a significant challenge in social science research due to the high cost and complexity of manual annotation. This thesis investigates the use of Multimodal Large Language Models (MLLMs) as a scalable solution for the analysis of video data, with a specific focus on the study of sexualization in social media content. A dataset of TikTok videos from Italy, the United States, and South Korea was constructed and analyzed using a structured codebook derived from prior literature on sexual objectification. A multimodal model was employed to generate both content coding annotations and textual descriptions of video content, enabling a unified video-to-text analytical pipeline. Results indicate that MLLMs can support large-scale analysis of video content, capturing consistent patterns aligned with theoretical expectations. In particular, systematic differences in sexualization were observed across gender, while cross-national variations were also identified. Complementary analyses of the generated textual descriptions provide additional evidence that the extracted signal reflects meaningful characteristics of the underlying content.