Οπτική αναγνώριση και γλωσσική αποκατάσταση ελληνικών παπύρων με χρήση τεχνητής νοημοσύνης
Optical character recognition and textual restoration of Greek papyri using artificial intelligence
View/
Keywords
Ψηφιακή παπυρολογία ; Υπολογιστική όραση ; Μεγάλα γλωσσικά μοντέλα ; RAG ; Vision transformers ; Οπτική αναγνώριση χαρακτήρων ; Αποκατάσταση κειμένου ; OCR ; LLMAbstract
The digitization and reading of ancient Greek papyri is one of the most formidable challenges in Computational Humanities, primarily due to severe material degradation and the historical use of continuous script (scriptio continua). This diploma thesis proposes an innovative, two-stage hybrid pipeline for the optical recognition and semantic restoration of papyrological texts. In the first stage (Computer Vision), the YOLO11s algorithm, augmented with the SAHI slicing technique, is employed for character bounding box detection on ultra-high-resolution scans, followed by a Vision Transformer (DeiT) for classification. Given the inherently noisy raw OCR output (characterized by high CER and WER), the second stage incorporates Large Language Models (LLMs) for textual restoration. To mitigate model hallucinations and overcome the tokenization failures of traditional dense embeddings on misspelled continuous text, a custom Retrieval-Augmented Generation (RAG) architecture was developed, tailored to the morphological patterns of ancient Greek. Comprehensive benchmarking across local and commercial state-of-the-art models demonstrated that this specialized RAG approach acts as a crucial grounding mechanism, measurably improving the Character Error Rate (CER). Finally, while the rigid syntactical requirements of translation caused traditional evaluation metrics (BLEU, METEOR) to collapse, evaluation via BERTScore revealed strong semantic resilience. The findings establish that while LLMs may struggle with publication-ready syntax, they excel at extracting the core conceptual meaning, positioning the proposed pipeline as a highly valuable sense-making tool for papyrologists.
Postgraduate Studies Programme
Κυβερνοασφάλεια και Επιστήμη ΔεδομένωνDepartment
Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα ΠληροφορικήςNumber of pages
93Language
GreekCollections
Except where otherwise noted, this item's license is described as
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα