Enhancing Research Synthesis Efficiency through Transparent AI-Powered Data Extraction

• In Event: AERA Poster Session Wednesday 3:45pm
  In Poster Session: Systematic Review and Meta-Analysis: Methods, Contexts, and Impacts

Abstract

Full-text data extraction remains a time-consuming and error-prone step in systematic reviews, often requiring manual effort to locate and interpret information scattered across sections. This study introduces TraceSynthesis, a web-based extraction tool powered by generative AI to support both structured and context-dependent information retrieval while improving transparency for human verification. The tool accepts full-text PDFs, a coding sheet, and optional background materials. To handle long-form inputs, the system applies a sliding-window chunking strategy. For each coding item, it then uses standardized prompting to generate an extracted value, its extraction type (Direct, Judgment-Based, or Indirect), and a transparent justification. Preliminary validation on 20 studies of generative AI and student creativity showed strong performance and enabled hallucination detection through traceable, interpretable outputs.

Authors

• Xiyu Wang, Purdue University

• Yukiko Maeda, Purdue University

• Hua-Hua Chang, Purdue University