Abstract:
This paper explores the application of Natural Language Processing (NLP) techniques to automate Harmonized System (HS) tariff line transposition, employing a three-stage process: unique 1:1 tariff code matching (Round 1), exact description matching (Round 2), and "smart" description matching (Round 3) using Artificial Intelligence (AI) and lexical similarity methods paired with harmonized 6- digit concordance and cosine similarity. Similarity is calculated using either Term Frequency Inverse Document Frequency (TF-IDF) vectors or Sentence-BERT (SBERT) embeddings, comparing two scenarios: a straightforward case (Economy A) with standardized descriptions, and a complex case (Economy B), with more detailed technical descriptions. Results indicate that automated HS transposition can significantly augment the efficiency of traditionally manual methods, reducing processing time from two to three weeks to approximately half a day (up to 30 times faster). The overall accuracy rate is 99.6% for the simpler scenario and 98.8% for the complex one, for a standard set of approximately 10,000 HS codes. While non-AI techniques cover most of the accurate matches, AI-based Round 3 techniques address cases requiring the most manual effort. SBERT generally outperforms TF-IDF, however including subheadings tends to reduce its accuracy. In certain cases, particularly for highly technical tariffs, TF-IDF's straightforward approach provides an advantage over SBERT. Overall, NLP techniques hold significant potential for improving HS transposition methods and facilitating the development of richer tariffs and trade datasets to enable more in-depth analyses. Future research should focus on refining these techniques across diverse datasets to optimize their broader application in tariff and trade data analysis.