Abstract: The current research on fine-grained sentiment analysis of tourists is mostly focused on approximate analysis based on topic clustering and frequency statistics and so on, lacking the meticulous quantitative portrayal of all aspect sentiment features extracted from each review more precisely. In this article, we constructed a large-scale ABSA Chinese annotated dataset of tourism reviews based on the big data of tourists' real reviews of hundreds of national A-level scenic spots in Guizhou Province on four well-known domestic tourism e-commerce platforms. using BERT-based LCF-AEPC aspects sentiment triplet extraction joint model, the prediction output of more than 170,000 aspect sentiment triplets were realized end-to-end on a dataset of nearly 80,000 subjective review texts. Based on the output results, a visual and quantitative analysis was conducted to explore the differences in tourists' sentiment characteristic regarding different aspects of attractions, as well as across different levels and types of attractions. Then the reasons of the differences were tried to be explained and a framework of four-element tourist sentiment impact factors was constructed. This article utilizes the deep learning techniques in the field of artificial intelligence to achieve ABSA multi-task, thereby quickly and effectively capturing the information of tourists' preferences, needs, and opinions on various aspects of tourist attractions. This will be a beneficial attempt and the latest application of fine-grained sentiment research in the field of tourism.

Key words: tourism big data, aspect sentiment analysis, deep learning, triplet extraction, sentiment features