Machine Learning-Based Theme Classification for Video Content Analysis: A Bilingual Approach on the StoryBox

This research introduces an advanced hybrid machine learning framework for the automatic thematic classification of video content in a bilingual (Turkish–English) setting, with a particular focus on the YouTube StoryBox dataset (172 videos). The proposed pipeline integrates sentence-level embeddings from Sentence-BERT, multilingual zero-shot classification with XLM-RoBERTa, classic clustering algorithms (HDBSCAN, k-means, and spectral clustering), dimensionality reduction via UMAP, and large language model (LLM) based theme labeling with Flan-T5. The StoryBox collection is thematically rich and highly heterogeneous, covering entrepreneurship, education, technology, and industry. As a result, the video embeddings occupy a continuous semantic manifold rather than form compact, well-separated clusters. This leads to weak hard-clustering scores in the original embedding space (negative silhouette values and 100% noise assignments for HDBSCAN), which we interpret as an evidence of intrinsically fuzzy, overlapping themes rather than a failure of the algorithms. Nevertheless, the bilingual use of multilingual transformer models yields a 23% improvement in F1-score over a monolingual baseline for theme consistency. LLM-assisted inspection of UMAP-projected clusters further reveals that “education and development” emerges as the dominant macro-theme, accounting for 54.1% of the corpus. We position the framework as a proof-of-concept for real-world video platforms and discuss its implications for scalable content organization, recommendation systems, and decision support in global, bilingual media environments.