Using Distilbert To Assign Hs Codes To International Trading Transactions

One significant source of national revenue for many countries is the tax levied on international trade. Tax collection can be achieved by accurately classifying international trading commodities according to Harmonised System (HS) Codes, which later can be used to impose duty/tax rates. The current approach to assigning HS codes to transactions relies on HS codes filled out by international traders and being manually inspected by customs officers. This approach is tedious and prone to error and fraud. Our research aims to determine the HS Codes automatically from commodity description texts in transactions using text classification techniques. However, commodity texts are hard to classify because of their short length, noise, ambiguity, and use of a lot of technical terms. To address these challenges, this paper proposes utilising transformers models, BERT and its variants, DistilBERT, which is claimed to be lighter and faster than the BERT model and has the advantage of being deployed in computational resource-constrained environments. The proposed approach adopts a transfer learning procedure to perform fine-tuning hyperparameters of BERT and DistilBERT. It is evaluated using real-world customs data for multi-class classification of commodity transactions in international trading. Experimental results demonstrate that both models achieve a comparable performance result.