Neural Models for Lemmatization and POS-Tagging of Earlier and Late Egyptian (Supporting Hieroglyphic Input) and Demotic

Aleksi Sahala, Eliese-Sophia LinckeNAACL 2025

TL;DR

This paper introduces updated neural models for lemmatization and Part-of-Speech (POS) tagging of earlier and later stages of the Egyptian language, including Demotic, with the novel capability of processing hieroglyphic input. The models achieve high accuracy across these ancient languages, even on datasets with damaged text, paving the way for more robust computational analysis of Egyptian historical texts.

What Problem Does This Paper Address?

Lemmatization and POS-tagging are crucial for making large collections of historical texts searchable and analyzable. However, ancient languages like Egyptian present significant challenges due to complex morphology, extensive spelling variations, and the existence of multiple distinct language phases (Earlier Egyptian, Late Egyptian, Demotic). Furthermore, the input data itself can be fragmented, containing "breakages" or lacunae. Existing tools often struggle with these complexities, especially when dealing with the original hieroglyphic script, which adds another layer of difficulty compared to transliterated text. This paper aims to bridge this gap by developing and evaluating neural models that can handle these challenges, including the use of hieroglyphic input and less-than-perfectly preserved texts.

Methods

The core of this work relies on the BabyLemmatizer, a system originally designed for cuneiform languages but adapted here for Egyptian. BabyLemmatizer treats lemmatization and POS-tagging as machine translation tasks, mapping character sequences to their corresponding linguistic labels. It employs a deep attentional encoder-decoder network with a BiLSTM encoder and a unidirectional LSTM decoder. A key aspect of their approach is the use of a dictionary-based post-corrector to refine lemmatizations and verify vocabulary entries, which also provides confidence scores for potentially incorrect lemmata. For this study, they specifically adapted the system to handle two input types: Unicode-encoded hieroglyphs and standard Egyptological transcription. They also experimented with data augmentation and a secondary model for post-correcting lemma identifiers, though the latter did not yield improvements.

Datasets

The study utilizes datasets derived from the Thesaurus Linguae Aegyptiae (TLA), specifically corpus v18. These datasets cover three language stages:

  • Earlier Egyptian: Texts predating the Egyptian New Kingdom (c. 1550-1070 BC).
  • Late Egyptian: Texts explicitly labeled as such in the TLA metadata.
  • Demotic: The entire Demotic text corpus in the TLA.

Crucially, these datasets are not filtered for "premium" sentences and include damaged text (broken or destroyed individual hieroglyphs or word forms). This represents a significant departure from previous work that often relied on cleaned, "perfect" data.

Key Insights and Ideas

  • Hieroglyphic Input is Challenging but Achievable: While hieroglyphic input generally leads to lower accuracy than transcription, the models still achieve respectable results, demonstrating the feasibility of processing this complex script directly.
  • Damaged Text Adds Complexity: The inclusion of texts with breakages, a common issue in ancient documents, highlights the practical utility of the developed models, as they perform reasonably well even on imperfect data.
  • Sub-lemma Disambiguation is a Major Hurdle: The high ambiguity of Egyptian lemmata, often requiring disambiguation between multiple sub-lemmata (e.g., the preposition 'm'), significantly impacts lemmatization accuracy. This is a key area where current models struggle.
  • OOV Words are a Significant Factor: The Out-of-Vocabulary (OOV) rate, particularly high for hieroglyphic input, directly correlates with lower accuracy, underscoring the challenge of unseen word forms.

Key Results

The models achieved the following accuracies (with confidence intervals):

  • UPOS Tagging: Consistently high, ranging from 93.64% (Earlier Egyptian, hieroglyphic) to 97.45% (Demotic, transcription).
  • XPOS Tagging: Also strong, from 92.97% (Earlier Egyptian, hieroglyphic) to 97.13% (Demotic, transcription).
  • Lemmatization: This proved to be the most challenging task, with accuracies ranging from 76.59% (Late Egyptian, hieroglyphic) to 92.15% (Demotic, transcription).

The OOV-rate for hieroglyphic input was notably higher (up to 16.75%) compared to transcription (around 2.5-5.5%), directly impacting lemmatization performance.

Limitations and Future Work

The authors acknowledge that disambiguating highly ambiguous Egyptian lemmata is beyond the current model's capabilities. They plan to address this in the future by exploring more context-aware approaches, such as transformers and LLMs, which could be fine-tuned for disambiguation. Additionally, incorporating dependency parsing and other annotation layers is suggested as a potential avenue to improve lemmatization quality, as syntactic and morphological labels have previously aided lemma disambiguation. They also propose organizing a shared task for Egyptian lemmatization to foster community involvement in solving these unique challenges.

Summary updated

BibTeX
@inproceedings{sahala2025neural,
  author = {Aleksi Sahala and Eliese-Sophia Lincke},
  title = {{Neural Models for Lemmatization and POS-Tagging of Earlier and Late Egyptian (Supporting Hieroglyphic Input) and Demotic}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
}
Figure