TL;DR
This paper introduces THOTH AI, a novel system that significantly improves the translation of ancient Egyptian texts by integrating Retrieval-Augmented Generation (RAG) with a specialized Coptic-Egyptian lexicon. By grounding Large Language Models (LLMs) with domain-specific linguistic knowledge, THOTH AI achieves state-of-the-art translation accuracy, paving the way for advancements in digital humanities and language revitalization.
What Problem Does This Paper Address?
Ancient Egyptian, despite its historical significance, remains a "low-resource language" for Natural Language Processing (NLP) tasks. This scarcity stems from limited digitized parallel corpora and the inherent complexity of its scripts and grammar. While general-purpose LLMs have revolutionized translation for well-resourced languages, they often struggle with the nuances and specialized vocabulary of ancient texts. This gap hinders research in digital humanities, e-learning, and efforts to revitalize ancient languages.
Methods
The core innovation is the application of Retrieval-Augmented Generation (RAG) to ancient Egyptian translation. THOTH AI integrates a powerful LLM (Claude 3.5 Sonnet) with a meticulously curated "Comprehensive Coptic Lexicon." This lexicon contains etymological, morphological, and historical data across various stages of Egyptian, from Old Egyptian to Coptic.
Here's how it works:
- Vectorization: The lexicon's entries are converted into vector embeddings, creating a searchable semantic space.
- Retrieval: When a user inputs a text snippet (in transliterated Middle Egyptian), THOTH AI generates vector embeddings for these tokens and searches the lexicon for the most similar entries.
- Augmentation: The retrieved relevant lexical information (definitions, morphological notes, historical context) is appended directly to the LLM's prompt.
- Generation: The LLM, now "grounded" with this domain-specific knowledge, generates a more accurate and contextually relevant translation.
The system also includes an optional OCR module for handling hieroglyphic and hieratic scripts, though this is noted as experimental.
Datasets
The primary dataset used for evaluation is the "coronation decree of Thutmose I," a classical Middle Egyptian text. This text was chosen for its complexity, containing formal epithets, references to deities, and lineage information, making it a good test case for LLM limitations. A Latin transliteration from de Buck (1948) and an English rendition by Nederhof (2023) served as the reference translations for quantitative scoring. The "Comprehensive Coptic Lexicon" is the key knowledge source, drawing from projects like the Thesaurus Linguae Aegyptiae, Coptic SCRIPTORIUM, and KELLIA.
Key Insights and Ideas
- RAG is Crucial for Low-Resource Languages: The paper strongly demonstrates that RAG is not just an incremental improvement but a critical technique for achieving high-quality translations in domains with limited data.
- Domain Knowledge is King: Explicitly injecting specialized linguistic knowledge (lexicons, morphology) into LLMs is far more effective than relying solely on their implicit, general-purpose training data for specialized tasks.
- Bridging Historical Stages: The Comprehensive Coptic Lexicon's ability to link earlier Egyptian forms to Coptic descendants highlights the potential for RAG to illuminate linguistic evolution and aid language revitalization.
- Beyond Literal Translation: The evaluation metrics (BLEU, METEOR, ROUGE, chrF) show that RAG improves not only literal accuracy but also the coverage of keywords and morphological consistency.
Key Results
THOTH AI, the RAG-enhanced system, outperformed all other tested models, including Claude 3.5 Sonnet (baseline), Gemini 2.0, DeepSeek R1, and various GPT variants, across multiple quantitative metrics.
- Highest BLEU Score: THOTH AI achieved a BLEU score of 0.354, significantly higher than its closest competitor, Claude 3.5 Sonnet (0.325).
- Strong Performance in Other Metrics: It also led in ROUGE and METEOR scores, indicating better coverage of key phrases and semantic fidelity.
- Qualitative Improvements: Manual review revealed that THOTH AI consistently provided correct glosses for rare or archaic terms and accurately handled complex morphological structures, which often confused other models.
Limitations and Future Work
- Lexicon Coverage: While extensive, the Comprehensive Coptic Lexicon may not cover every specialized religious text or local dialect variant, potentially leading RAG to miss crucial definitions.
- OCR Accuracy: OCR for ancient scripts remains experimental and can be visually complex, requiring human oversight for accurate transliteration.
- Cultural Nuance: Deeply contextualized cultural knowledge and interpretations of minor deities or ephemeral socio-historical events cannot be fully encoded in dictionary entries and still require Egyptologist expertise.
Future work could involve expanding the lexicon, improving OCR capabilities, and exploring how to incorporate deeper cultural context. The authors also highlight the potential for THOTH AI in e-learning, digital humanities, and Coptic language revitalization efforts, suggesting its broad applicability.
Summary updated