Calculating tourist sentiment ambivalence through aspect-level sentiment analysis: Infusing tourism domain knowledge into a pre-trained language model

Abstract

Effectively capturing sentiment ambivalence—where individuals simultaneously experience both positive and negative sentiments—enables a more nuanced understanding of tourist sentiments for both academic and industry. Prior studies have measured ambivalence using self-reported overall sentiments data, which may suffer from bias and ambiguity. Building on schema theory, we proposed assessing tourists' sentiment ambivalence based on their aspect-level sentiment toward tourism objects. Tourism domain knowledge was incorporated into the Bidirectional Encoder Representations from Transformers (BERT) model to develop the TK-BERT. Online reviews of a Hong Kong attraction from multiple online platforms were used as a case study. TK-BERT demonstrated higher accuracy compared to the original BERT and other state-of-the-art models. This study advanced the understanding of sentiment ambivalence by operationalizing the concept and identifying the roles of different aspects in its formation. Methodologically, this paper provided a valuable tool for ambivalence calculation and aspect-level sentiment categorization.