AI Commodity Classification: The Core of Trade Data Automation
In trade, HS code (Harmonized System Code) classification is the foundation for determining tariff rates, producing trade statistics, and determining FTA rules of origin. Accurately mapping the world's 5,000+ six-digit HS codes and each country's ten-digit sub-classifications is among the most specialized areas of trade practice, and AI-powered automatic classification has emerged as a flagship challenge in DX innovation.
KOTRA's DX Innovation Lab has developed an NLP (Natural Language Processing)-based automatic commodity classification model that estimates HS codes from product description text alone, and has built a system that also supports cross-conversion with MTI (Korea Trade Statistics Classification) and SITC (Standard International Trade Classification).
Understanding Trade Commodity Classification Systems
Several commodity classification systems are used in global trade, each serving different purposes. AI automatic classification models learn the mapping relationships between these systems, enabling simultaneous estimation of multiple codes from a single input.
| System | Governing Body | Code Structure | Primary Use |
|---|---|---|---|
| HS (Harmonized System) | WCO (World Customs Organization) | 6-digit (10-digit nationally) | Tariff imposition, customs clearance, rules of origin |
| MTI (Korea Trade Statistics Classification) | KITA (Korea International Trade Association) | 6-digit | Korean trade statistics and analysis |
| SITC (Standard International Trade Classification) | UN (United Nations) | 5-digit | International comparative statistics |
| BEC (Broad Economic Categories) | UN (United Nations) | 3-digit | Economic analysis, input-output relationships |
| CPC (Central Product Classification) | UN (United Nations) | 5-digit | Integrated services and goods classification |
AI Classification Model Architecture
The AI commodity classification model adopts a three-stage hierarchical classification approach. By reflecting the hierarchical structure of HS codes (chapter → heading → subheading), the model progressively narrows classification from broad to specific, improving overall accuracy.
Training Data and Model Performance
The accuracy of an AI commodity classification model depends heavily on the quality and volume of training data. KOTRA's DX Innovation Lab secured 5 million+ actual customs clearance records from the Korea Customs Service to train the model, and built industry-specialized glossaries and synonym databases to improve domain-specific accuracy.
| Classification Stage | Code Count | Top-1 Accuracy | Top-3 Accuracy | Main Error Sources |
|---|---|---|---|---|
| Chapter (2-digit) | 97 | 96% | 99% | Boundary between similar chapters (e.g., Ch.84 vs Ch.85) |
| Heading (4-digit) | 1,200+ | 89% | 95% | Ambiguous product descriptions |
| Subheading (6-digit) | 5,300+ | 82% | 91% | Insufficient specification detail |
| 10-digit (Korea) | 12,000+ | 75% | 88% | Differences in national sub-classifications |
Practical Application: Bangladesh Trade Commodity Classification
AI commodity classification is especially valuable in Korea-Bangladesh trade. Bangladesh's tariff system uses HS code sub-classifications that differ from Korea's, and the multi-tier additional duty structure (CD, SD, VAT, AIT, AT) is complex — meaning accurate HS code identification has a direct impact on total landed costs.