Purpose: This paper aims to classify the sentiment of online tourism-hospitality reviews at an aspect level. A new aspect-oriented sentiment classification method is proposed based on a neural network model. Design/methodology/approach: This study constructs an aspect-oriented sentiment classification model using an integrated four-layer neural network: the bidirectional encoder representation from transformers (BERT) word vector model, long short-term memory, interactive attention-over-attention (IAOA) mechanism and a linear output layer. The model was trained, tested and validated on an open training data set and 92,905 reviews extrapolated from restaurants in Tokyo. Findings: The model achieves significantly better performance compared with other neural networks. The findings provide empirical evidence to validate the suitability of this new approach in the tourism-hospitality domain. Research limitations/implications: More sentiments should be identified to measure more fine-grained tourism-hospitality experience, and new aspects are recommended that can be automatically added into the aspect set to provide dynamic support for new dining experiences. Originality/value: This study provides an update to the literature with respect to how a neural network could improve the performance of aspect-oriented sentiment classification for tourism-hospitality online reviews.