The perception of tourists holds significant importance in understanding people’s needs and enhancing the quality of urban development. Training an NLP customized model using park network text as data is more suitable for the demands of the landscape architecture field, enabling intelligent and efficient governance and design of parks. Using a deep learning platform, three models were trained for multi-label text classification, sentiment analysis, and comment viewpoint extraction to process park network text. A multi-level analysis of “time-evaluated object-perceived elements under the evaluated object” was conducted to identify key perception features and explore the significant elements. The research revealed the following findings: (1) Overall, visitors’ perception of Eling Park was predominantly positive, with the highest attention given to external landscapes among the six evaluated objects. Natural landscapes had the most positive impact on visitors’ emotions, while negative emotions were associated with facilities and the need for parking space. (2) Among the 60 high-frequency perception elements, 7 elements are significantly proportional to the probability of tourists’ positive emotions, of which 5 are positively and 2 are negatively correlated. (3) The analysis sequence of “text classification - high-frequency word extraction - sentiment analysis” allowed the identification of low-frequency perceived elements that had essential impacts on corresponding landscapes. (4) The attribute-level sentiment analysis provided by the NLP customized model reduced sentiment analysis errors, leading to more accurate research outcomes. This study examined the emotional perception of visitors and the key perceived elements of Eling Park, providing optimization suggestions for its development and contributing to the application of natural language processing in landscape architecture.