Enhanced Chemical Reaction Mining through Combined Model Adaptation and Prompt Engineering

Area of Science:

• Chemical Informatics
• Natural Language Processing
• Computational Chemistry

Background:

• Extracting structured chemical reaction data from scientific literature is complex due to varied language and formats.
• Existing methods struggle with the nuances of chemical terminology and reporting styles.

Purpose of the Study:

• To develop a novel framework for enhanced chemical reaction information extraction.
• To improve the accuracy and efficiency of identifying chemical entities like reactants, products, solvents, and reagents from unstructured text.

Main Methods:

• A three-stage framework combining quantized low-rank adaptation fine-tuning, entity self-grounding, and dynamic prompt engineering.
• Progressive enhancement stages for domain adaptation, accuracy refinement, and task optimization.
• Evaluation on a dataset of 26,785 reaction descriptions.

Main Results:

• Achieved high F1 scores: 92.62% for reactants, 90.72% for products, 89.64% for solvents, and 88.83% for reagents.
• Demonstrated strong chemical identifier recognition, with 93.88% recall for IUPAC names.
• Outperformed specialized systems (ChemRxnBERT at 78.7% F1) and recent LLM approaches (83.0% F1).

Conclusions:

• The proposed framework offers a significant advancement in chemical reaction information extraction.
• It requires less training data and computational resources compared to traditional domain-adaptive pretraining.
• Enables applications in automated literature mining, synthesis planning, and chemical informatics.