【Purpose/significance】As the number of travel websites increases, the number of online reviews of tourists is increasing. Aiming at the problems of traditional methods such as feature over-dimensioning and data sparseness in the topic classification of short travel text reviews, this paper proposes a topic discovery method for travel reviews based on convolutional neural networks and SOM.【Method/process】Firstly, word vectors are used to represent the texts, which reduces the problem of over-dimensionality. Secondly, high-order abstract features are extracted from the commentary texts by Convolutional Neural Networks. Finally, the topics are clustered based on the abstract features extracted from the SOM model.【Result/conclusion】The experimental results show that compared with the traditional text clustering algorithm, CNN-SOM algorithm has significantly improved the accuracy rate, recall rate and F-value, which can make thematic discovery of travel reviews better.