Extract and Analyze the Performance Data from Research Papers using Generative AI

Abstract

This study explores automated extraction and analysis of tabular data from research papers to streamline researchers’ workflows. It integrates Generative AI and Optical Character Recognition within an end-to-end pipeline applied to over 1,200 open-access Artificial Intelligence and Machine Learning papers. PDF files were first converted into high-resolution images, after which tables were detected using a fine-tuned YOLOv8 model. Text from the detected tables was extracted using Tesseract OCR, and performance-related data was filtered and analyzed using Retrieval-Augmented Generation methods. The analysis identified top-performing models, such as BERT and CODEX, and widely used datasets including SQuAD and GSM8K, enabling automated meta-analysis. The results demonstrate the scalability and effectiveness of combining computer vision and NLP for high-quality data extraction.Models such as Llama3-8B and deepseek-r1:8b 0528-quwen3-q8_0 provided domain-specific insights. The study also suggests improved table detection without relying on keyword searches by leveraging advanced AI, NLP, and ensemble learning techniques.