TL;DR
This paper surveys the application of Natural Language Processing (NLP) techniques to Ancient Egyptian languages, highlighting the challenges posed by low-resource status and the unique characteristics of these historical scripts. It provides a comprehensive overview of existing corpora, NLP tasks tackled, and the current state of research, aiming to stimulate further interest in the field.
What Problem Does This Paper Address?
Ancient Egypt, a civilization foundational to Western culture, left behind a vast corpus of textual evidence. However, despite its historical significance and popular appeal, the languages of Ancient Egypt have been relatively overlooked in the field of Natural Language Processing (NLP). This is primarily due to several key challenges: the absence of native speakers, making pronunciation and nuanced understanding difficult; the vast timespan of the language (over 3,000 years), leading to significant variations; and the often-damaged state of surviving documents. These factors classify Ancient Egyptian languages as "low-resource," meaning they cannot readily benefit from the cutting-edge, data-hungry NLP models like transformers. This paper aims to bridge this gap by reviewing how NLP has been applied to these languages, identifying common issues, and showcasing existing resources and research.
Methods
The paper provides a literature review of NLP applications to Ancient Egyptian languages, categorizing them by task. The methods discussed include:
- Transliteration: Developing probabilistic automata and alignment models to convert hieratic and hieroglyphic scripts into phonetic values or Latin script. This involves creating rule-based systems and finite-state automatons.
- Translation and Part-of-Speech (POS) Tagging: Exploring joint modeling approaches, such as using encoders and decoders, to simultaneously translate and tag text, assuming that shared learning can compensate for limited data.
- Text Classification: Applying techniques like Naïve Bayes, SVM, and segment-and-combine methods for genre classification of ancient texts.
- Text Retrieval: Developing systems that index and allow querying of digitized documents using Latin script, hieroglyphs, or a combination, employing Boolean models and vector space representations.
- Semantic Representations: Creating "semantic maps" that graph meanings and their connections across languages and time, reflecting expected semantic changes.
- Morphological Analysis: Utilizing transducers and context-free grammars to model the morphology of Coptic, a descendant language.
- Named Entity Recognition (NER): Investigating the adaptation of existing NER systems and transformer-based models for ancient languages, with mixed success.
Datasets
The paper discusses several key corpora and lexical resources:
- Middle Egyptian: A corpus annotated for hieratic transliteration and symbol functions, derived from the St. Andrews corpora. It consists of only two texts, with a recommendation to train and test on separate texts to avoid data leakage.
- Late Egyptian: The Ramses project aims to build a comprehensive annotated corpus of Late Egyptian texts. It has grown significantly, containing thousands of texts and over a million words, with annotations focusing on inflections, lemmata, and spellings.
- Demotic: The Chicago Demotic Dictionary is a significant resource, including words and scans of documents, updated until 2012.
- Coptic: The Coptic Scriptorium is a comprehensive corpus with eleven sub-corpora, featuring tens of thousands of manually annotated words, covering tasks from tokenization to treebanking. Other resources include the Database and Dictionary of Greek Loanwords in Coptic and the Marcion project.
- Thesaurus Linguae Aegyptiae (TLA): A corpus covering Old, Middle, and Late Egyptian, Demotic, and Coptic, spanning from the Old Kingdom to Roman times, with lemmatization and morpho-syntactic annotation.
- Thot Sign List (TSL): A collection of graphemes found in hieroglyphic or hieratic texts.
Key Insights and Ideas
- Low-Resource Challenges are Paramount: The inherent scarcity of data and the absence of native speakers fundamentally shape the NLP approaches applicable to Ancient Egyptian, often precluding state-of-the-art deep learning models.
- Coptic as a Bridge: Coptic, as a later stage of Ancient Egyptian written with a modified Greek alphabet, presents fewer transliteration challenges and benefits from its inclusion in the Universal Dependencies framework, leading to more advanced NLP research compared to other Egyptian stages.
- Interdisciplinary Collaboration is Crucial: The paper emphasizes the need for greater interaction between NLP researchers and Egyptologists to better understand the specific needs and nuances of studying ancient texts.
- Standardization is an Ongoing Issue: The lack of standardized writing systems and the variations across time periods and regions make tasks like transliteration and segmentation complex.
- Semantic Maps Offer Unique Insights: The development of "semantic maps" provides a novel way to visualize and understand how meanings evolve over time within a language.
Key Results
- Transliteration: Probabilistic models for Middle Egyptian hieratic transliteration achieved recall and precision scores of approximately 0.95. A simpler model for transliteration achieved accuracies of 0.98 and 0.97 on different texts.
- Translation and POS Tagging: A joint modeling approach for translating Middle Egyptian to German achieved a BLEU score of 22.79 points, outperforming a direct translation baseline.
- Text Classification: A Naïve Bayes classifier achieved a recall of slightly over 0.84 for genre classification, improving to over 0.97 for letters and monumental inscriptions.
- Coptic NLP: The Coptic Scriptorium's NLP pipeline achieved accuracies of over 0.9 for segmentation, 0.98 for normalization, and 0.95/0.97 for POS tagging and lemmatization.
Limitations and Future Work
- Data Scarcity: The primary limitation across all Ancient Egyptian languages (except perhaps Coptic) is the lack of sufficient annotated data, hindering the application of advanced NLP models.
- Generalization Across Time: Many approaches struggle to generalize across the vast temporal span of Ancient Egyptian, often focusing on specific periods.
- Limited Quantitative Results for Some Tasks: For tasks like transcription and POS tagging in the Wiesenbach and Riezler (2019) paper, specific quantitative results were not reported. Similarly, the paper by Bédi et al. (2022) on multimodal systems did not report quantitative results.
- Future Work Suggestions: The authors suggest increased collaboration between NLP and Egyptology projects, the development of more standardized annotation schemes, and the exploration of comparative NLP studies across different ancient languages. The potential for further development of the Ramses project and the Thesaurus Linguae Aegyptiae corpora is also highlighted as beneficial for text retrieval systems.
Summary updated