Aspect Extraction for Tourist Spot Review in Indonesian Language using BERT

Abstract

Review is an essential aspect to support decision making in various fields and industries, with tourism being one of them. With the emergence of Internet, especially the Web 2.0, feedbacks and reviews can now be found easily from forums, blogs, and social websites. However, the sheer amount of data making it a tedious task for user to analyze and extract useful information from all those reviews. Aspect-based sentiment analysis (ABSA) is one of solutions to solve the aforementioned problem with aspect extraction phase as one of the essential component of it. In this paper, we adapt Bidirectional Encoder Representations from Transformers (BERT) to process aspect extraction by adding the Indonesian language specific preprocessing phase and further pretrain the multilingual base model with 4220 review sentences in Indonesian language. For the experiment, we collect 501 review sentences from TripAdvisor and do the performance evaluation using the 10-fold cross validation method. The collected reviews are about tourism spots, specifically amusement park, in Indonesian language. We found our adaptation succeed in improving the performance of BERT for aspect extraction with the new further pretrained base model produced the best accuracy and F1-measure value with 0.799 and 0.738 respectively, while adding preprocessing phase with the new further pretrained base model produced the best accuracy with 0.824. The impact of preprocessing phase can also be seen with the best improvement for each labels ROC-AUC value happens on scenarios that contain preprocessing phase.