Issue Detection and Future Proofing Dutch Government Apps Using Language Technologies

As public services increasingly shift to digital platforms due to e-Government initiatives, understanding and incorporating user feedback has become critical for improving the quality and usability of government applications. The field of Natural Language Processing (NLP) techniques have emerged as a crucial response to the need to process and analyzeand diverse user feedback. Among these techniques, Large Language Models (LLMs) have become key scalable and versatile tools. This research explores the application of LLMs to extract, classify, and forecast issues reported in user reviews of four Dutch government applications, namely KopieID, Reisapp, MijnOverheid, and DigiD. This research is structured around four core tasks: (1) issue extraction, (2) multi-label review classification, (3) assessment of how different issues impact star ratings, including a temporal analysis, and (4) forecasting of future issues and actionable recommendations. A comparative analysis between LLMs and Latent Dirichlet Allocation (LDA) is performed to evaluate coherence and classification confidence (via Shannon Entropy). The results show that LLMs outperform LDA in coherence, flexibility, and interpretability, though challenges such as hallucination and classification ambiguity were observed. The star rating assessment highlights that technical reliability remains a key driver of user dissatisfaction, while usability-related concerns exhibit more variable effects across applications. Forecasting analysis reveals that LLMs can partially identify emerging issues and generate precise, app-specific recommendations, though the prediction of issue frequency remains limited. This research offers a replicable, unsupervised pipeline for multilingual user feedback analysis and provides practical insights for enhancing citizen-centric digital services in the public sector. Government institutions could use and build on this pipeline to identify critical pain points in their applications, create an evidence-based prioritization framework based on the evolution of discovered issues, and employ focused recommendation strategies.