Abstract: Customs commodity HS code classification is an important international procedure for cross-border trade of enterprises and individuals.HS code classification can be regarded as a text classification problem,that is,given a paragraph of description for a commodity,to determine the category of the commodity represented by HS code.However,this task is more challenging than general text classification task.First,commodity description texts are organized with special hierarchical structures.Then commodity description texts present sequential features at two levels.In addition,the key information in the commodity description text is scattered and the description forms are diverse.Most of the existing classification methods cannot comprehensivelyconsiderthe above factors to capture key information in the commodity description text.In this paper,we proposes a Text Sequence and Graph Information combination Neural Network(TSGINN) to solve the problem of customs commodity HS code classification.The TSGINN defines the HS code classification problem as a subgraph classification problem based on word co-occurrence network,models association between non-contiguous words through graph attention network,and captures multi-level sequential information through hierarchical long short-term memory network.Experiments on the real-world customs datasets show that the classification effect of TSGINN model is better than that of other methods.

Key words: Customs commodity, Graph attention network, HS code, Multi-level sequential information, Text classification