Explainable AI with fine-tuned large language models for sustainable cultural heritage management

Abstract

The redevelopment of cultural heritage areas, especially in historical urban environments, requires a nuanced understanding of public perceptions to balance preservation with modernization. This study introduces an advanced AI-driven framework for Aspect Sentiment Quadruple Prediction (ASQP) to assess public perceptions of Lijiang Ancient Town, a UNESCO World Heritage site in China. We fine-tuned the large language model Qwen-14B using LoRA-based methods to augment sentiment data, effectively uncovering implicit emotional cues in social media content. The model integrates BERT, multi-layer BiLSTM, self-attention, CNN, and CRF for enhanced entity recognition and sentiment classification. Experimental results show that the enhanced model (Qwen-14B + ASQP) improved F1-score by 0.97% (from 75.42% to 76.39%) and Precision by 4.48% (from 76.14% to 80.62%) compared to the baseline. Analyzing data from platforms such as Weibo, Dazhong Dianping, and Xiaohongshu (2018–2024), the research uncovers factors influencing public perception, offering insights for heritage site management, urban planning, and the sustainable preservation of cultural heritage.