Domain alignment of large language models for aspect based sentiment analysis of architectural heritage perception in Huishan Ancient Town

Asian historic districts, exemplified by Huishan Ancient Town, face significant pressures from intensive tourism and commercialization, often leading to structural aging and a loss of authenticity. Conventional sentiment analysis tools frequently struggle with specialized architectural terminology, resulting in a superficial understanding of public perception. This study introduces a domain-specific alignment approach using a Qwen3-14B Large Language Model fine-tuned via Quantized Low-Rank Adaptation (QLoRA). By integrating heritage knowledge into the model’s training objective, we enhance its ability to interpret professional terms – such as “leaking windows” and “borrowed scenery” – within informal visitor narratives. Results demonstrate that the aligned LLM achieves a 78.72% F1-score in heritage value classification, effectively bridging the gap between vernacular language and professional conservation frameworks. This research provides a high-precision digital tool for monitoring sentiment and cultural integrity, offering a scientific basis for urban planners to balance commercial development with the preservation of tangible and intangible values in Asian historic environments.