A fine-tuned large language model approach for compound information extraction

With the increasing complexity of global trade, the rapid and accurate identification of chemical substances has become a significant obstacle for customs inspection. Traditional laboratory-based methods such as GC–MS and HPLC offer high precision and remain gold-standard confirmatory tools; however, the practical bottleneck in large-scale customs operations lies in rapidly interpreting massive textual materials to triage suspicious targets before laboratory verification. To address this limitation, we propose a fine-tuning framework for Large Language Models (LLMs) using Low-Rank Adaptation (LoRA) to automatically extract structured chemical information from unstructured text. The framework enables the identification of substances and their functional descriptions from vast and up-to-date textual data, allowing customs authorities to narrow potential targets based on product claims before laboratory verification. Specifically, ERNIE 4.0 was selected as the base model and fine-tuned on a domain-specific corpus annotated with compound names, molecular formulas, functions, and hazards. Experimental results show that the LoRA-based methodology markedly improves extraction accuracy and fluency, achieving BLEU-4 and ROUGE-L scores of 74.64 and 85.97, respectively, outperforming both non-fine-tuned baselines and other pre-trained models. These results demonstrate that parameter-efficient fine-tuning provides an effective and scalable solution for building rapid detection databases to support intelligent customs supervision.