Intelligent Nanoparticle Literature Extraction System: From Unstructured Text to Structured Knowledge

This article introduces the RL-Agent-Extraction-PLGANPs project—an automated literature mining system based on reinforcement learning and large language models, aimed at extracting nanoparticle formulation data from scientific literature and building a structured knowledge base. The system adopts a phased pipeline design, emphasizing data lineage tracking and result auditability, and combines LLM semantic understanding with rule-based processing to ensure the extraction process is transparent and reliable.

In the fields of materials science and nanotechnology, researchers need to extract experimental data such as nanoparticle formulations and preparation methods from massive literature. Traditional manual research is inefficient, and simple keyword searches struggle to handle complex contexts and implicit information. The RL-Agent-Extraction-PLGANPs project builds an end-to-end automated literature mining pipeline, using LLM's semantic understanding capabilities to extract structured formulation records, emphasizing data lineage tracking and result auditability to ensure transparency and reliability.

The system adopts a phased pipeline design, with clear input/output and quality control for each stage:

• Stage0: Literature collection and relevance screening. Import raw corpus from Zotero and screen relevant literature to ensure data source quality.
• Stage1: Content cleaning and preprocessing. Perform format standardization, encoding conversion, and noise removal, outputting cleaned content and lists.
• Stage2 (Core): Semantic discovery and information extraction. Use LLM semantic discovery to identify formulation information, complement with rule engine for standardization, expand multiple experimental conditions via DOE rows, and follow the "LLM-first" principle.
• Stage3: Relationship construction and entity association. Identify formulation component entities, establish relationships between components and formulations, and output standardized relationship records.
• Stage4: Evaluation and diagnosis. Conduct field completeness checks, numerical rationality checks, and comparisons with manual benchmarks, supporting error analysis and iteration.
• Stage5: Baseline output and table construction. Materialize relationship records into formulation tables, avoid new inferences to ensure reproducibility, and finally output main tables, variant records, comparative statistics, and evaluation reports.

The project's underlying design embodies the idea of automated prompt optimization:

• Closed-loop feedback: Quality evaluation feedback of extraction results is used to optimize LLM prompt templates, forming a continuous improvement cycle.
• Dynamic curriculum learning: Prioritize processing literature with clear structures, then gradually challenge complex expressions.
• Boundary governance: Define governance categories such as internal intermediate results and diagnostic boundaries, supporting pause, branching, and replay for easy debugging and auditing.

The project practices rigorous data governance:

• Authoritative source contract: Clearly define current operational data sources through documents to avoid implicit inferences.
• Lineage tracking: Each result can be traced back to the original literature, processing stages, and LLM configurations, supporting full traceability.
• Frozen baseline: Raw outputs of key stages are frozen and saved, allowing reproducibility and comparison of historical results.
• Risk stratification: Generate Layer2 risk stratification reports to provide metadata support for downstream audits.

In terms of technical implementation, the project is developed in Python, with code organized by stages in the src directory, along with directories for governance documents, data, and supporting documents. Its application value is not limited to the nanoparticle field; it demonstrates a reusable methodology for literature knowledge extraction:

• Human-machine collaboration paradigm: Combine LLM semantic understanding with deterministic rule processing to leverage their respective advantages.
• Quality-first design: Multi-stage verification, benchmark comparison, and lineage tracking ensure data reliability.
• Evolvable architecture: Clear stage division and boundary definition support gradual system upgrades.

This project provides a reference implementation for research teams needing to build structured knowledge bases.