BERTopic_Teen: a multi-module optimization approach for short text topic modeling in adolescent health

Adolescent health has become a critical dimension in the digital era, as social media platforms emerge as vital sources of real-time behavioral data for informing sustainable and equitable public health strategies. However, conventional topic modeling methods often struggle with the semantic sparsity and noise inherent in short-form texts. The study proposes BERTopic_Teen, an enhanced topic modeling framework optimized for adolescent health-related tweets. The model incorporates three key innovations: a Popularity Deviation Regularizer (PDR) to suppress high-frequency generic terms and amplify domain-specific vocabulary; a Dynamic Document Embedding Optimizer (DDEO) that adaptively selects optimal UMAP dimensions based on silhouette scores; and a Probabilistic Reassignment Matrix (PRM) to reassign outlier documents to relevant topic clusters. Using a dataset of 64,441 tweets (61,039 successfully classified), experimental results show that BERTopic_Teen outperforms LDA, NMF, Top2Vec, and the original BERTopic in all key evaluation metrics. It achieves a 16.1% improvement in topic coherence (NPMI = 0.2184), higher topic diversity (TD = 0.9935), and lower perplexity (1.7214), indicating superior semantic clarity, topic distinctiveness, and modeling stability. These findings suggest that BERTopic_Teen offers a robust solution for extracting meaningful topics from social media data and advancing public health surveillance.