Knowledge-aware Recommendation Using Language Models

This whitepaper presents a hybrid approach for predicting Harmonized System (HS) codes from product descriptions using Large Language Models (LLMs). It outlines the challenges faced in traditional methods of HS code assignment, which include knowledge base mapping and historical best practice mapping. The document details the development of a composite task model designed to improve the accuracy of HS code assignments by integrating knowledge-driven and data-driven methodologies. The proposed solution aims to enhance precision in HS code assignment, adapt to changes in coding rules, and encapsulate historical best practices while maintaining data privacy. The paper also discusses the implications of these challenges on key performance indicators such as processing time, operational costs, regulatory constraints, and training costs for employees involved in the HS code assignment process. Overall, the approach seeks to streamline the categorization of goods in international trade, addressing the complexities introduced by linguistic variations and regulatory requirements.