Automated Generation of International Trade Documents via Fusion of Large Language Models and Deterministic Rules

Abstract

In cross-border e-commerce and foreign trade, documents such as commercial invoices, packing lists, and customs declarations rely heavily on manual synthesis from unstructured texts (e.g., chat logs and emails), resulting in low efficiency and frequent
numerical or formatting errors. While Large Language Models have demonstrated exceptional capabilities in natural language understanding, their direct application in high-compliance scenarios for end-to-end generation remains hindered by numerical inconsistency,
logical instability, and uncontrollable outputs. This paper proposes an automated generation method that integrates LLMs with deterministic rules and algorithms. By constructing a hybrid pipeline consisting of "Semantic ParsingRule ConstraintTemplate Generation, " the method leverages LLMs to handle semantic uncertainty while delegating numerical allocation and rigid logic to deterministic
algorithms, ensuring structural consistency and numerical validity. The proposed method has been deployed in real-world trade scenarios for three years. Experimental results indicate that this approach significantly outperforms both manual processes and pure LLMbased solutions in terms of end-to-end efficiency and reliability, providing a viable engineering path for LLM applications in highreliability document automation.