HScodeNet: Combining Hierarchical Sequential and Global Spatial Information of Text for Commodity HS Code Classification

Commodity Harmonization System (HS) code classification is an important customs procedure in cross-border trade. HS code classification is to identify the category (i.e., HS code) of a commodity according to its description information. In fact, HS code classification is essentially a text classification task. However, compared with general text classification, the challenge of this task is that commodity description texts are organized in special hierarchical structures and contain multiple independent semantic segments. What's more, the label space (i.e., the HS code system) has hierarchical correlation. In this paper, we propose a HS code classification neural network (HScodeNet) by incorporating the hierarchical sequential and global spatial information of texts, in which a hierarchical sequence learning module is designed to capture the sequential information and a text graph learning module is designed to capture the spatial information of commodity description texts. In addition, a label correlation loss function is presented to train the model. Extensive experiments on several real-world customs commodity datasets show the superiority of our HScodeNet model.