In recent years, newspaper digitization has advanced significantly, with many national and international initiatives focusing on content analysis and providing open access outcomes that follow FAIR principles. However, most of this content is in Latin script. The situation changes dramatically when non-Latin script and/or historical material are processed. In case of Republican Chinese newspapers, those processing approaches are rarely published even within China, and results usually get commercialized.

The Early Chinese Periodicals Online (ECPO) project publishes its data open access. The project aims to develop a pipeline for processing printed materials from Republican China, focusing on complex newspaper layouts. ECPO collaborates with Heidelberg University's Scientific Software Center to establish a page segmentation pipeline using existing tools like State Library Berlin developed Eynollah and Baidu’s PaddleOCR.

Our experiments show that tools trained on Western newspapers did not work well for Chinese data. We had to fine-tune the models using our own ground truth dataset. As we will show, best results stem from a combination of pixel-based Eynollah and box-oriented PaddleOCR. Feeding the text crops into an OCR engine, earlier experiments achieved a character error rate (CER) of <3% for material printed in regular grids and additional llm-based error correction. Recent open vision language models like Alibaba’s QWEN3-VL reach a robust accuracy of ca. 96% independent on the printing grid and without requiring any ground truth.

Establishing an effective pipeline for page segmentation and OCR processing will mark a significant breakthrough in East Asian studies, advancing open access to full texts of the Republican press. Our initiative not only enhances the resources within ECPO but also sets the foundation for future DH projects to process material in other historical contexts.