Babbage in Babylon: Pioneering Deep Learning Approaches to Sumerian Cuneiform

To date, archaeologists have unearthed more than 120,000 Sumerian texts, a corpus which reflects both the apparent origins of writing itself as well as a rich written tradition spanning three millennia. Reading Sumerian, however, requires years of specialized training. With such an abundance of texts and so few specialists able to read them, many texts have not been read. Few have been translated. This thesis explores the promise of deep learning-based tooling in accelerating the reading of, and facilitating non-specialist engagement with, Sumerian cuneiform. Approaching the reading process as a series of NLP tasks has previously been limited by a lack of well-structured, accessible datasets. To that end, I publicly release three novel datasets: SumTablets, SumTablets_Photos, and SumTablets_English. I then train and evaluate two language models: The first model predicts obfuscated glyphs with an accuracy of 80%; the second model performs automated transliteration, achieving a state-of-the-art average character-level F-score (chrF) of 97.55. In releasing datasets, modeling tasks, and establishing baseline performance, this work lays the foundation for a new approach to Assyriology, where cutting-edge advances in deep learning enable rapid, accurate parsing of the world's most ancient texts.