Named Entity Recognition for Ancient Egyptian Transliteration

This preprint presents a named entity recognition model that tags deities, personal names, and places directly in ancient Egyptian Leiden Unified Transliteration, trained on the Thesaurus Linguae Aegyptiae Earlier Egyptian dataset (12,773 sentences, predominantly Old Kingdom). Entity spans come from the corpus's expert proper-noun tags and entity types from its expert glossing, so no manual annotation is needed. After exact duplicate sentences are removed and the data are split into training, development, and test sets, the model reaches 90.9 ± 0.6 percent F1 on held-out test sentences that share no text with its training data, averaged over five random splits (DEITY 90.1, PERSON 93.4, PLACE 68.7). It separates the god Horus from the royal Horus title in context and finds 62.8 percent of names never seen as entities in training. Version 2 corrects version 1. Version 1 reported 95.6 percent F1, measured on a validation set with substantial duplicate overlap with the training data and with entity types from a hand-typed deity list; re-measured without that overlap, the version 1 labelling scores 93.6 ± 0.4. Model, code, and full results: https://github.com/Juhij2/egyptian-transliteration-ner