Transferring old-letter ottoman Turkish scripts to latin-based turkish using natural language processing techniques

The history of the Republic of Turkey is essentially a continuation of the Ottoman Empire. Therefore, modern Turkish inherits many elements from Ottoman Turkish. The Ottoman Empire holds an important place not only for its political impact on world and Turkish history but also for its social and cultural influence. As many of the regions it once governed are now home to various modern states, research in this field is significant both from the perspective of national history and international studies. Ottoman Turkish, which blended concepts primarily from Arabic and Persian as well as from Western languages, needs to be understood—especially in historical studies but also in many other areas. The use of the Arabic script, along with modifications specific to Ottoman Turkish, has made learning this script and the language itself essential for researchers today. The fact that tens of millions of documents in the Presidential State Archives of the Republic of Turkey have yet to be fully utilized in academic research highlights the necessity of not only human resources but also advanced technological methods. Accordingly, a technical solution has been developed to digitize and interpret Ottoman Turkish.The foundation of these projects lies in Ottoman-Turkish to modern Turkish texts sourced from books. During the digitization process, pages were scanned at 300 DPI to preserve clarity and detail, resulting in a dataset of approximately 800 pages. The data was categorized based on page structure and linguistic features into short, medium, and long sentence samples, creating a total of 1,250 sentences. The dataset was split into 70% training, 15% validation, and 15% test subsets. The system was built on Tesseract OCR 5.0 and supported by deep learning models such as LSTM, CNN, RNN, Seq2Seq, and Word2Vec. The architecture, which integrates Python and ASP.NET Core projects, includes customized algorithms for both text recognition and translation processes. As a result of testing, the system achieved a 4-gram BLEU score of 0.8381, ROUGE-1 F1 score of 0.8674, ROUGE-L F1 score of 0.8352, an average TER of 0.2516, WER of 0.2489, and CER of 0.1762. These values demonstrate the system's high performance at the character, word, and sentence structure levels. Moreover, translations were rated acceptable at 88.24% by TER, 88.40% by WER, and 95.04% by CER. Higher accuracy was observed in short and medium-length sentences, while longer and syntactically complex structures slightly challenged system performance. The high ROUGE-L F1 score indicates the system's strong ability to preserve sentence integrity. With this robust design, this thesis demonstrates the effectiveness of deep learning-based NLP techniques in translating Ottoman documents into modern Turkish and offers both an academic and practical framework for similar historical language transformations.