SciExtract-LLaMa: A Robust SLM Framework for Automated Structured Data Extraction from Scientific Abstracts via Multi-Stage Prompting, RAG, and Self-Correction

The vast and unstructured scientific literature hinders quantitative analysis and trend discovery. While Small Language Models (SLMs) with Prompt Engineering can structure abstracts, existing methods often face challenges in accuracy, output stability, and extensive manual post-processing. We introduce SciExtract-LLaMa (SEL), a novel framework for robust, automated structured data extraction from scientific abstracts. SEL leverages a Phi-4 SLM, combining multi-stage Prompt Engineering, Retrieval-Augmented Generation with domainspecific knowledge, self-correction, and automated entity standardization. Operating entirely in an inference-only mode, SEL requires no model training or fine-tuning. Evaluated on a comprehensive dataset of scientific abstracts, SEL consistently surpasses conventional SLMbased methods. Our results demonstrate high extraction accuracy, completeness, and format consistency. Critically, SEL significantly reduces human post-processing time compared to baseline methods. These findings confirm SEL's efficacy in generating high-fidelity structured data, streamlining academic discovery and facilitating large-scale scientific trend analysis.