Comparative relations behind online reviews contain rich information concerning customers assessments of different products or services, thereby supporting upcoming consumers purchase decisions, as well as helping to identify enterprises market competitiveness. Instead of using the pattern recognition method, this paper proposes a hierarchical multi-attention network (HMAN) model to extract the comparative relations, in order to greatly reduce the requirements of artificial features and the manual annotation in the relation mining process. Such model outperforms both traditional classification models and text classification models in terms of accuracy, with its F1-score up to 81%. Besides, the proposed model has a good performance on extracting comparative relations from long texts where comparison information is relatively scattered. In this study, we visualise results of different experiments in order to demonstrate the interpretability of this model, and furthermore explore the mechanism of multi-attention method in comparative relations mining. This study applies the deep learning method instead of pattern recognition to automatically capture deep features of comparative relations, and therefore it redefines the identification process of comparative relations.