Topic modeling plays a central role in social science research by uncovering thematic structures within textual data across diverse domains. It is particularly valuable in contexts where expert-annotated datasets are scarce and existing benchmarks rely on overly broad categories. Recent advances in dense embedding representations based on Sentence Transformer architectures offer new possibilities for scalable classification and interpretable analysis through embeddings comparison and visualization. This study compares the performance of recent embedding models derived from transformer-based language models with Word2Vec embeddings trained on domain large-scale corpora. Focusing on opinion mining in the hospitality sector, our case study shows that in scenarios involving large-scale unlabeled datasets, domain-specific Word2Vec embeddings remain an effective solution, and can even outperform more recent contextual embeddings. In addition, we provide an overview of the epistemic debate surrounding topic modeling research and highlight that the notion of a universally “best” topic model is meaningless.