Research On Commodity HS Code Classification With Graph Neural Networks

With the rapid development of cross-border trade,commodity Harmonization System(HS)code classification,as an important customs procedure in international import and export trade for enterprises,has attracted more and more attention from relevant departments for its correctness and efficiency.If HS code classification can be performed automatically,accurately and efficiently,it will help customs to smoothly carry out tariff calculations,trade statistics,etc.,and can also help enterprises to improve customs clearance efficiency and reduce customs clearance costs.Commodity HS code classification task can be regarded as a text classification task,that is,given a paragraph of description for a commodity,the goal is to identify its category,i.e.,HS code.However,the commodity HS code classification is more challenging than general text classification.First,commodity description texts are organized with special hierarchical structures.Then,a commodity description text contains several independent semantic segments rather than one uniform continuous semantic topic.What's more,there are hidden hierarchical correlations between class labels(i.e.,HS codes).This paper focuses on the challenges of commodity HS code classification and tries to explore effective HS code classification methods.Firstly,given the structural characteristics of commodity description text and the incomplete feature capture of existing text classification methods,this paper proposes a HS code classification model(TSSINN)which combines text sequential and spatial information.Aiming at the two-level sequential features of commodity description text,a sequential information modeling method is designed to capture the multi-level sequential information in the text.In addition,a spatial information modeling method based on text graph is designed,which constructs a global corpus-level text graph based on the word co-occurrence information of the entire corpus and extracts the corresponding commodity text graph for each commodity.And graph neural network is utilized to model the spatial correlation among word nodes in the graph,to capture the spatial information in the commodity description text.The final classification is achieved by combining the text sequential features and spatial features.Then,on the basis of the TSSINN model,this paper further mines the global spatial information of the commodity description text and the hierarchical correlation constraints in the label space,and proposes a HS code classification model(HScode Net)which utilizes the label hierarchy.Aiming at the problem that TSSINN does not sufficiently consider the long-range dependencies in the text,this paper designs a global spatial information modeling method to learn the long-range spatial correlation in the text and capture the global spatial information of the text.In addition,aiming at the problem that TSSINN does not consider the hidden hierarchical correlation constraints in the label space,a label correlation loss function is designed.The real label distribution of the sample is obtained by calculating the correlation between any labels and the sample,and the whole model is optimized by fitting the label distribution.Finally,extensive experiments on four real-world customs commodity datasets are carried out in this paper.The experimental results show that the classification performance of the TSSINN model is significantly better than baseline models,and the optimized HScode Net model further improves the classification effect on the basis of the well performance of the TSSINN model.Overall,the HS code classification methods proposed in this paper can effectively combine the structural characteristics of the commodity description text and the hierarchical structure of the label space,capture the sequential information and spatial information in the text,and finally perform accurate HS code classification.