DH-LAB and Qulto RI

DH-LAB and Qulto RI:A novel approach to the manual and AI-based transcription of Ottoman Turkish manuscripts Tamás Kiss (bio) KEYWORDS Ottoman HTR, Ottoman OCR, manual transcription, TEI Viewer, TEI Search, NER Hungarian software company Qulto,1 in collaboration with the National Laboratory for Digital Heritage2 hosted by Eötvös Loránd University's Department of Digital Humanities,3 has created DH-LAB,4 a software infrastructure that enables the AI-based processing and research of digital cultural heritage. Qulto's extensive software development has resulted in a suite of research tools, collectively called Qulto RI,5 which, integrated with DH-LAB, enable complex workflows for research, especially in the humanities and social sciences. One such workflow addresses the challenges of processing, annotating, and presenting Ottoman Turkish manuscripts, with the goal of offering a solution to the automatic text recognition (handwritten text recognition; HTR) of Ottoman Turkish manuscripts. The manual transcription of Ottoman Turkish manuscripts written in the Arabo-Persian script to the Latin script poses several challenges that remain unresolved. Transcribing a right-to-left script to left-to-right, adjusting page segmentation to the specificities of Ottoman Turkish manuscripts (e.g., curved [End Page 49] lines, line ending words, and syllables written on top of lines, etc.), the data enrichment of transcribed documents, and the simultaneous presentation of the resultant image and text layers of the documents to a wider audience are currently not supported by any online tool, especially not as parts of a continuous workflow in an integrated software environment. Ottoman Turkish HTR presents even greater challenges than manual transcription. Currently, no software offers seamless ways to create training sets constituted by cropped images of right-to-left scripts and text files of left-to-right scripts for AI-based recognition. Researchers often transcribe Ottoman Turkish documents in reverse Latin script to create training data, but this is a laborious process and makes the intelligible presentation of the transcribed documents impossible. Additionally, building a training model in an iterative manner (i.e., correcting faulty HTR results to fine-tune training sets in order to decrease character and word error rates) is difficult due to the lack of fit-for-purpose tools. The DH-LAB-Qulto RI production team has been working to offer solutions to these challenges by considering them as interdependent stages of the same workflow. For example, a comprehensive solution to the problem of the manual transcription of right-to-left to left-to-right scripts provides a solution to creating meaningful training sets for the future HTR of Ottoman Turkish documents and, at the same time, to presenting simultaneously the image and text layers of the same document. Converting the ALTO XML result files to TEI XML files enables the data enrichment of the textual layer, which will support further processing and research as well as improving user experience in reading, searching, and researching these documents. And a microservice that supports the correction of future HTR results will create better training sets in each iteration. DH-LAB and Qulto RI's shared software components currently allow users to follow through the transcription (in the future: HTR), processing, and presentation workflow in two phases: Phase 1 1. Depositing (scanned or photographed) images of Ottoman Turkish manuscripts into an InvenioRDM repository, which supports version management. 2. Generating an empty OCR result file (i.e., ALTO XML) in Qulto's OCR module. (Tesseract, Google's open-source OCR engine, is used for OCR). 3. Transcribing documents from scratch (left-to-right and right-to-left text orientation is defined at line level, and, conclusively, transcribing "mixed" documents is supported) in DH-LAB's and Qulto RI's shared transcription/proofreading module (fig. 1).6 [End Page 50] 4. Saving training data into separate models: In the transcription/proofreading module, users can define the language and model into which they wish to save their training data in a file system. In case of reverse transcription (i.e., a right-to-left script transcribed left-to-right), the training data will consist of a cropped image file of a word and a reversed character chain of the transcript entered by the user. (The

Linguistics · Programming language · Transcription (linguistics · Turkish · Workflow · World Wide Web · Computer Science · Handwritten Text Recognition Techniques · Image Processing and 3D Reconstruction · Natural Language Processing Techniques · Software