Ancient texts are important because they connect us with ancient civilizations, through which we gain cultural, religious and scientific knowledge. Ancient texts, whether on papyrus, parchment, or other substrates, are often fragmented, degraded, or partially erased due to the passage of time. Restoration of these texts presents a significant challenge to historians and scholars, requiring meticulous manual effort and expertise.
Ancient text restoration is a specialized branch of the text restoration that focuses on recovering and preserving textual content from historical or ancient documents.
Traditional restoration methods rely heavily on manual intervention by experts, which is time-consuming and often subjective. In recent years, the application of machine learning (ML) and artificial intelligence (AI) techniques has shown promise in automating and enhancing the restoration process.
Deep learning techniques have shown remarkable success in various domains, including image processing and natural language processing. In this thesis, different models were proposed for the restoration of ancient texts by using deep neural networks.
Two datasets used for training and testing the models the first dataset being “Codex Sinaiticus” a manuscript dating back to the fourth century, it is a significant artifact as it provides the earliest extant complete copy of the New Testament in the Christian Bible. The handwritten material is written in the Greek language.
The second dataset being “Argonautica 3” which refers to an epic poem written by the ancient Greek poet Apollonius of Rhodes in the 3rd century BCE which is written in the Greek language too.
The dataset has been preprocessed by encoding dataset. New lines, numbers, symbols and special characters have been removed. After that the result text has been tokenized, generate missing character, class label obtained, augmentation performed to support dataset, and normalization process performed.
Three prediction models were used as proposed models for retrieving missing ancient texts, Long Short-Term Memory (LSTM), Recurrent Neural Networks (RNN), and Generative Adversarial Networks (GAN) and the results were testing accuracy 86%, 92% and 98.3% according to the first dataset and 94%, 88% 98.7% according to the second dataset respectively.
Comparing the performance of each model, GAN gave the best accuracy results, and thus it proved its effectiveness in the field of restoring missing text. The results of the proposed system were also compared with other restoration techniques, where the results showed that the proposed technique had higher accuracy results than others.
Overall, this work contributes to the interdisciplinary intersection of deep learning and digital humanities, offering a promising solution for the restoration and preservation of ancient textual artifacts.