Understanding users’ experiences and evaluations in urban public spaces is essential for human-centric design and management. While traditional survey methods are costly, large-scale user-generated content (UGC) offers new possibilities for capturing public insights. However, existing approaches mainly assess evaluations and experiences through oversimplified, single-dimensional assessments, while deriving users’ multidimensional experiences beyond basic sentiments and capturing their underlying mechanisms remain critical research gaps. This study proposes a novel integrative analytical model for assessing urban public space experiences and evaluations. Drawing on large-scale online review data, it employs a Low-Rank Adaptation (LoRA)-fine-tuned RoBERTa language model for multidimensional user evaluation identification, combines a lexicon-based method with an Aspect-based Sentiment Analysis (ABSA) pipeline to examine specific experiential qualities and satisfaction, and integrates textual and visual modalities through Bootstrapping Language-Image Pre-training (BLIP) vision-language model. Focusing on Amsterdam as a case study, validation demonstrates that the proposed model outperforms baselines, achieving a mean accuracy of 85% and exceeding widely used XGBoost and LSTM methods by 12% and 23%, respectively. The application effectively quantifies nuanced and detailed user experience and evaluation patterns, revealing differentiated associations between experiential qualities and overall evaluation positivity, thereby providing a deeper understanding of human-environment interactions in urban public spaces. The proposed approach offers an integrated and scalable method for investigating user insights and can inform future responsive, human-centric public space design and management.