Abstract

The automation of Harmonized Tariff Schedule (HTS) code classification - a crucial task in international trade that entails assigning standardized codes to goods for tariff and regulatory purposes - is examined in this paper using GPT-3, a cutting-edge Large Language Model (LLM). The focus on creative prompt-engineering techniques and the usage a frozen model approach eliminates the need for further model fine-tuning. The suggested approach makes use of GPT-3’s built-in capabilities to decipher intricate product descriptions and reliably assign them to the appropriate HTS categories. We show through an array of experiments that strategic prompt adjustments including few-shot prompting allowed the model's initial top-1 accuracy of 23 percent to be significantly increased to an average of 73 percent. The main strengths of the approach are the explainability for the chosen category as well as the fully generalizable nature of the methodology.

‹ 1 / ? › − 100% + Open in new tab Download Cite

Loading the full text…