Towards AI-assisted Translation of Hieroglyphic Egyptian: Interpreting Iufaa’s Coffin Texts

Summary Recent advances in deep learning have made automated reading and translation of hieroglyphic Egyptian a realistic, though methodologically demanding, goal. The present study argues that ancient Egyptian must presently be treated as a low-resource language, making end-to-end neural translation impractical. We therefore propose a modular, sevenstage workflow – from wall documentation and hieroglyph detection to transliteration, grammatical analysis, and translation – in which each component can be developed and evaluated independently. Using Late Period shaft tomb inscriptions as a primary test case, we discuss the technical and philological challenges affecting each stage, including surface geometry, paleographic variation, reading-sequence reconstruction, and orthographic ambiguity. Hieroglyph detection using graph neural networks achieves very high accuracy on facsimile data, while reading order and orientation remain major bottlenecks. Sign recognition performs well on synthetic datasets but degrades on wall inscriptions due to domain shift, highlighting the need for transfer learning. Transformer-based transliteration models trained on the Ramses corpus reach accuracy comparable to previous state-of-the-art approaches and reveal systematic divergences that reflect genuine linguistic ambiguity as well as occasional errors in human-annotated datasets. Through detailed case studies drawn from the sarcophagus texts of Iufaa and related Late Period material, we demonstrate how automated methods can help identify variant readings, non-standard spellings, and editorial strategies of Late Period priests, while leaving final interpretation firmly in the hands of the Egyptologist.