Name in original language
LibrAIfy: Digitalizacija tekstualnog kulturnog nasleđa - povećanje kvaliteta optički prepoznatog teksta
Initiative overview
Our initiative was oriented towards the application of modern AI technologies to the digitisation of Serbian textual cultural heritage. This process poses significant challenges due to the variable quality of scanned material, historical language use, and authentic printing practices often reflected in distinctive typography, decorations, and complex layouts. Although AI tools offer vast potential, they are rarely suitable for off-the-shelf use and require careful domain and language adaptation.
Building on open-source tools rooted in AI subfields such as computer vision and natural language processing, and informed by recent academic advances, our team developed a technology pipeline suitable for image enhancement, layout segmentation, optical character recognition (OCR), and post-OCR text correction and improvement. Thanks to the partnership with the National Library of Serbia, the solution was tested and evaluated on a representative collection of historical periodicals, resulting in the digitisation of more than 16,000 documents and over 400 GB of material.
The initiative was important for two main reasons: (1) it supports the preservation of culture and language while providing access to a vast body of historical, cultural, and social content important for the education of the general public and for research; (2) it contributes to the creation of resources essential for the progress of local language technologies and the development of AI-based tools for text representation and understanding. As such, it brought valuable and extensive experience working with historical documents and provided deeper insight into the state of the technology, particularly in OCR and post-OCR processing for low-resource languages.
Accordingly, the objectives of the initiative were:
- Establishing a suite suitable for technology benchmarking;
- Developing a Serbian-focused post-OCR tool to mitigate inaccuracies in modern OCR solutions;
- Creating additional high-quality, diverse resources to support local language technology development and language model training.
Given that our partner, the National Library of Serbia, is continuously engaged in digitisation activities, diverse metrics and user feedback can be systematically collected and leveraged for ongoing software improvement and adaptation to broader, more varied scenarios, enhancing its applicability across other local digitisation centres. The initiative was also recognised and strongly supported by the UNDP Digital, AI, and Innovation Hub, bringing global visibility and opening opportunities for collaboration with communities worldwide.
