Convolutional Neural Network Approach for Aspect-Based Sentiment Analysis of Tourism Reviews

The tourism industry is a key economic sector in Indonesia, with East Java ranking highest in tourist visits. This study aims to enhance tourism development by applying aspect-based sentiment analysis (ABSA) using convolutional neural networks (CNN) to analyze online reviews. CNN was selected for this study due to its proven efficiency in capturing local n-gram features and its relatively lower computational cost compared to other deep learning model. Reviews from TripAdvisor and Google Maps were collected focusing on four aspects: attraction, amenities, access, and price. Five different models were developed in this research: one multilabel aspect classifier designed to identify multiple aspects mentioned within each review, and four sentiment classifiers focused on evaluating the sentiment polarity for each specific aspect. These models were trained and evaluated using various combinations of word embeddings, including static embeddings like Word2Vec, and contextualized embeddings such as BERT and IndoBERT. Additionally, the impact of preprocessing through stemming was investigated to understand how simplifying word forms affects model performance. Results indicate that IndoBERT-CNN attains the best overall sentiment classification, reaching F1-scores up to 0.71 for attraction and 0.93 for amenities, while Word2Vec-CNN with stemming leads multilabel classification. Meanwhile stemming improves performance for static embeddings like Word2Vec by simplifying word forms, it reduces effectiveness in transformer-based models like BERT and IndoBERT that rely on natural language context. These findings highlight the benefit of choosing appropriate embeddings and preprocessing for different tasks, thus providing practical insights for improving tourism services through better tourist reviews analysis.