The opinions and feelings expressed by tourists in their reviews intuitively represent tourists' evaluation of travel destinations with distinct tones and strong emotions. Both consumers and product/service providers need help understanding and navigating the resulting information spaces, which are vast and dynamic. Traditional sentiment analysis is mostly based on statistics, which can analyze the sentiment of a large number of texts. However, it is difficult to classify the overall sentiment of a text, and the context-independent nature limits their representative power in a rich context, hurting performance in Natural Language Processing (NLP) tasks. This work proposes an aspect-based sentiment analysis model by extracting aspect-category and corresponding sentiment polarity from tourists' reviews, based on the Bidirectional Encoder Representation from Transformers (BERT) model. First, we design a text enhancement strategy which utilizes iterative translation across multiple languages, to generate a dataset of 4,000 reviews by extending a dataset of 2,000 online reviews on 1,000 tourist attractions. Then, the enhanced dataset is reorganized into 10 classifications by the Term Frequency-Inverse Document Frequency (TF-IDF) method. Finally, the aspect-based sentiment analysis is performed on the enhanced dataset, and the obtained sentiment polarity classification and prediction of the tourism review data make the expectations and appeals in tourists' language available. The experimental study generates generic and personalized recommendations for users based on the emotions in the language and helps merchants achieve more effective service and product upgrades.