Enhanced Chemical Reaction Mining through Combined Model Adaptation and Prompt Engineering

Extracting structured chemical reaction information from the scientific literature remains challenging due to the complexity of chemical language and diverse reporting formats. This study presents a novel three-stage framework that combines quantized low-rank adaptation fine-tuning, entity self-grounding, and dynamic prompt engineering for enhanced chemical reaction extraction. The approach systematically addresses domain adaptation, accuracy refinement, and task optimization through progressive enhancement stages. Evaluation on 26,785 reaction descriptions demonstrates substantial improvements over existing methods, achieving F₁ scores of 92.62% for reactants, 90.72% for products, 89.64% for solvents, and 88.83% for reagents. Chemical identifier recognition shows strong performance, with IUPAC names achieving 93.88% recall. The framework outperforms specialized systems like ChemRxnBERT (78.7% F₁) and recent LLM approaches (83.0% F₁) while requiring significantly less training data compared to domain-adaptive pretraining approaches and computational resources. This methodology establishes an effective framework for chemical reaction information extraction from unstructured text, enabling applications in automated literature mining, synthesis planning, and chemical informatics.