Miktub: A Manuscript Dataset of Historical Maltese for Handwritten Text Recognition

Abstract

The digitisation of handwritten historical material is essential for preserving cultural heritage and enabling search and computational analysis. For Maltese, historical handwritten resources are scarce, and, to the best of current knowledge, no public handwritten text recognition (HTR) dataset for historical Maltese exists. We introduce a Manuscript Dataset of Historical Maltese (Miktub), collected from the Data Provider: 35 scanned pages transcribed by specialists and converted into a line-level HTR dataset. A key challenge was robust line extraction from heterogeneous pages; fully automatic line segmentation was insufficient, so we developed a semi-automatic pipeline combining horizontal projection profiling with lightweight post-processing and manual refinement to maximise line fidelity. We provide two annotation variants, including a corrected/standardised version (Miktub-COR) designed to improve consistency, accessibility, and downstream learning stability. We benchmark two strong public HTR models, HTR-VT and VAN, and report the best test performance of 4.68% character error rate (CER) and 13.59% word error rate (WER) on Miktub-COR with VAN. We will release Miktub publicly upon acceptance, along with scripts and splits, to support historical Maltese-language technology research.

Resources

Details