Harmonized System Code Classification Using Transfer Learning with Pre-Trained Weights

Abstract

The Harmonized System (HS) was developed as a multipurpose international product nomenclature that describes the type of good that is shipped. It allows customs authorities to identify and clear every commodity that enters or crosses any international borders. HS classification is to identify the HS code of a commodity according to its description information in a trade manifest. Compared with general text classification the challenge of this task is that commodity description texts are often short, unstructured and extremely noisy. HS misclassification can lead to penalties, fines and delays upon import. We first propose novel approaches for extracting and filtering relevant commodity information from a trade document. Then our HS classification methodology utilizes pre-trained STS models via deep transfer learning using sentence-level transfer. We also introduce a new evaluation method to properly evaluate our approach based on real-world applications. Extensive experiments and model comparisons show the superiority of our approach.