TL;DR
This paper introduces the challenge of deciphering the ancient Meroitic language as a computational problem, presenting the first machine-readable Meroitic corpus and training word embeddings. While initial cross-lingual alignment experiments with Late Egyptian yielded limited success, the work lays crucial groundwork for future computational approaches to understanding this historically significant but poorly understood language.
What Problem Does This Paper Address?
The Meroitic language, spoken in the ancient Kingdom of Kush (modern-day Sudan) from roughly 270 BC to 330 AD, remains largely undeciphered. Unlike languages like Ancient Egyptian, which benefited from bilingual texts like the Rosetta Stone, Meroitic lacks such a key. This linguistic barrier has significantly hindered our understanding of the Meroitic civilization's history, culture, and broader context within the ancient world. While the Meroitic script itself has been deciphered, its grammar and vocabulary remain a mystery. This paper aims to bridge this gap by applying modern computational linguistics techniques, specifically focusing on machine translation and natural language processing (NLP), to the task of Meroitic decipherment.
Methods
The authors employ a multi-pronged computational approach:
- Corpus Creation: They construct the first machine-readable Meroitic corpus by manually transcribing and digitizing existing Meroitic texts from various sources, including vocabulary lists and royal narratives. This corpus is made publicly available.
- Word Embeddings: They train Word2Vec embeddings on both the newly created Meroitic corpus and a cleaned version of the Late Egyptian corpus. This allows for the representation of words as dense vectors in a semantic space.
- Intrinsic Evaluation: The quality of these embeddings is evaluated intrinsically using semantic similarity tests and word analogy tasks. This helps assess how well the embeddings capture relationships between known words within each language.
- Cross-Lingual Alignment: The core of their decipherment attempt involves aligning the Meroitic and Late Egyptian embedding spaces. This is done using VecMap, a method suitable for distant languages, with the goal of inducing translations for unknown Meroitic words based on their proximity to known Egyptian words. They explore unsupervised alignment, alignment based on numerals, and alignment based on a dictionary of nouns.
Datasets
- Meroitic Corpus: A newly created machine-readable corpus of Meroitic texts, including translated words, phrases, and royal narratives. This corpus is made publicly available.
- Late Egyptian Corpus: A cleaned version of the Ramses Online Corpus of Late Egyptian texts, scraped and processed into JSON format.
- Alignment Dictionaries: Seven small dictionaries were compiled, pairing known Meroitic word forms with their hypothesized Egyptian counterparts. These serve as training and evaluation sets for the lexicon induction task.
Key Insights and Ideas
- Meroitic as a Computational Challenge: The paper frames Meroitic decipherment not just as a philological problem but as a tractable computational task, opening new avenues for research.
- The Value of Even Limited Data: Despite the "extremely limited nature" of the Meroitic corpus, the authors demonstrate that even this data can capture some semantic information, particularly in the Egyptian embeddings.
- The Difficulty of Low-Resource Alignment: The results highlight the significant challenges in cross-lingual alignment when dealing with extremely low-resource languages like Meroitic, even when paired with a more established language like Late Egyptian.
- Orthographic Variation as a Major Hurdle: The inherent variability in Meroitic spelling is identified as a significant obstacle that can erroneously create distinct word forms, complicating embedding and alignment tasks.
Key Results
- Intrinsic Evaluation: The Egyptian embeddings performed well on semantic similarity tests, showing high cosine similarities for related words (e.g., "king" and "Ra"). The Meroitic embeddings also showed promise, capturing some semantic relationships, though less consistently than Egyptian.
- Lexicon Induction: The cross-lingual alignment experiments yielded "abysmal" results. The best performance on the test dictionary was only 20% accuracy for numerals, and even the training dictionary saw a maximum of 20% accuracy. This suggests that the current embeddings and alignment methods are insufficient for reliable lexicon induction between Meroitic and Late Egyptian.
- Preliminary Cognate Study: Hand-identified possible cognates for the words /seb/ and /kek/ were found, suggesting potential thematic links to "earth," "water," or "darkness," but these did not directly support existing scholarly translations and remain speculative.
Limitations and Future Work
- Data Scarcity: The most significant limitation is the extremely small size and potential outdatedness of the Meroitic corpus. The authors acknowledge that more machine-readable data is crucial.
- Orthographic Variation: The inconsistent spelling of Meroitic words poses a challenge for capturing semantic meaning accurately.
- Limited Alignment Success: The poor performance of lexicon induction indicates that current methods are not yet effective for this specific language pair.
Future work suggested by the authors includes:
- Expanding the Corpus: Curating and digitizing more Meroitic texts, potentially from recent publications like Hallof (2024), is a top priority.
- Exploring Other Languages: Aligning Meroitic with other related languages like Coptic or Old Nubian, or even modern languages, could yield better results.
- Handling Orthographic Variation: Developing methods to normalize or account for spelling variations before training embeddings.
- Advanced NLP Tasks: Utilizing the new corpus for cognate detection, Named Entity Recognition (NER), and POS-tagging.
- Genre-Specific Analysis: Studying lexical elements common to different Meroitic genres to better understand vocabulary usage.
Summary updated