Discovering a tourism destination with social media data: BERT-based sentiment analysis

Purpose The main purpose of this paper is to analyze a tourist destination using sentiment analysis techniques with data from Twitter and Instagram to find the most representative entities (or places) and perceptions (or aspects) of the users. Design/methodology/approach The authors used 90,725 Instagram posts and 235,755 Twitter tweets to analyze tourism in Granada (Spain) to identify the important places and perceptions mentioned by travelers on both social media sites. The authors used several approaches for sentiment classification for English and Spanish texts, including deep learning models. Findings The best results in a test set were obtained using a bidirectional encoder representations from transformers (BERT) model for Spanish texts and Tweeteval for English texts, and these were subsequently used to analyze the data sets. It was then possible to identify the most important entities and aspects, and this, in turn, provided interesting insights for researchers, practitioners, travelers and tourism managers so that services could be improved and better marketing strategies formulated. Research limitations/implications The authors propose a Spanish-Tourism-BERT model for performing sentiment classification together with a process to find places through hashtags and to reveal the important negative aspects of each place. Practical implications The study enables managers and practitioners to implement the Spanish-BERT model with our Spanish Tourism data set that the authors released for adoption in applications to find both positive and negative perceptions. Originality/value This study presents a novel approach on how to apply sentiment analysis in the tourism domain. First, the way to evaluate the different existing models and tools is presented; second, a model is trained using BERT (deep learning model); third, an approach of how to identify the acceptance of the places of a destination through hashtags is presented and, finally, the evaluation of why the users express positivity (negativity) through the identification of entities and aspects.