Named Entity Recognition for Ancient Egyptian Transliteration

Named entity recognition (NER) for ancient Egyptian texts has been limited to work on modern translations, leaving the source language largely unaddressed by computational methods. This paper presents the first NER model trained directly on ancient Egyptian Leiden Unified Transliteration, identifying divine and human entities without reliance on English translation. Using the Thesaurus Linguae Aegyptiae (TLA) corpus of 12,773 Middle Egyptian sentences, training examples are automatically generated by leveraging existing UPOS part-of-speech annotations, eliminating the need for manual entity labeling. A language-agnostic spaCy NER model trained on 7,059 annotated sentences achieves 95.6% F1 across DEITY and PERSON entity classes. The model is demonstrated on Book of the Dead Spell 125 and implications for automated entity extraction pipelines connecting transliteration text to structured demonological knowledge bases such as DemonThings/DemonBase are discussed.