A Topic Modeling Perspective on Fintech User Feedback: Evaluating Transformer-Based Models for Semantic Clustering

The rapid proliferation of fintech applications has generated a massive volume of user-generated feedback, often in the form of unstructured textual reviews. Extracting actionable insights from such data is essential for improving user experience, product design, and service reliability. This study presents a comprehensive comparative evaluation of three topic modeling approaches-TF-IDF with KMeans clustering, Latent Dirichlet Allocation (LDA), and BERT-based clustering using SBERT embeddings-applied to 67,000 real-world reviews from a widely-used fintech mobile application. The analysis employs both intrinsic clustering metrics (Silhouette Score, Davies-Bouldin Index, Calinski-Harabasz Index) and interpretability-oriented heuristics (UMass Coherence, Jaccard Overlap) to assess model performance. Findings demonstrate that BERT + KMeans significantly outperforms traditional methods in capturing semantically rich and distinct topics, with practical implications for fintech stakeholders in product development, customer support, and strategic planning. The study also proposes a hybrid evaluation framework and emphasizes the importance of noise resilience, semantic precision, and scalability in modern NLP pipelines for financial feedback analysis. Limitations and future directions are discussed, including the integration of sentiment, intent recognition, and multilingual adaptations.