💬 NLP

Neural Machine Translation for Coptic-French: Strategies for Low-Resource Ancient Languages

This paper presents the first systematic study on translating Coptic into French, demonstrating that fine-tuning with a stylistically varied and noise-aware corpus significantly improves translation quality and offers valuable insights for low-resource ancient languages.

Original authors: Nasma Chaoui, Richard Khoury

Published 2026-05-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nasma Chaoui, Richard Khoury

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very old, dusty, and slightly damaged book written in an ancient language called Coptic. You want to translate it into modern French so more people can read it. The problem? There are very few people who speak Coptic today, and there aren't many existing computer programs (AI) that know how to do this translation.

This paper is like a recipe book where the authors, Nasma Chaoui and Richard Khoury, test different "cooking methods" to see how to build the best possible translator for this specific task. They aren't just guessing; they are running experiments to find the most reliable way to turn ancient text into modern French.

Here is a breakdown of their journey, using simple analogies:

1. The Big Question: How do we start?

The authors asked four main questions, which they treated like four different paths up a mountain:

Path A: The Direct Route (Fine-Tuning). Take a smart AI that knows a little bit about languages and "train" it specifically on Coptic and French. It's like taking a general student and hiring a private tutor just for Coptic.
Path B: The Pivot Route (The Detour). Translate Coptic to English first, then translate that English result into French. It's like trying to get to Paris by first flying to London and then taking a train.
Path C: The "Magic Prompt" Route. Ask a smart AI that knows Coptic and English to just "speak French" without any extra training. It's like asking a chef who only knows how to cook Italian to suddenly make a perfect French dish just by asking nicely.
Path D: The Multilingual Route. Use a giant AI that knows 100 languages but isn't specialized in any of them. It's like using a Swiss Army knife to try to perform surgery; it has the tools, but maybe not the precision.

The Result: The "Direct Route" (Path A) won hands down. Training a model specifically for Coptic-to-French was much better than using detours or hoping a general AI would figure it out.

2. Choosing the Right "Student" (The Model)

Once they decided to train a model, they had to pick which one to start with. They tested four different "students":

The Specialist: An AI that already knew Coptic but only spoke English.
The Polyglot: An AI that knew Coptic and French but was part of a huge group of 100 languages.
The Generalist: An AI that knew French but didn't know Coptic at all.
The Cousin: An AI trained on Hieroglyphs (the ancient ancestor of Coptic).

The Result: The Polyglot (the one that already knew both Coptic and French) was the best student. However, the Cousin (the Hieroglyph expert) did surprisingly well! This is like finding out that if you teach someone to read ancient Egyptian hieroglyphs, they pick up Coptic much faster than someone who has never seen an ancient language before. It proves that knowledge of a "cousin" language helps a lot.

3. The Power of Variety (Multiple Translations)

Ancient texts are tricky because one sentence in Coptic might be translated into French in three different ways, all of which are correct. The authors wondered: Should we teach the AI with just one translation, or all three?

They tested this by training the AI on just one version of the Bible translation, and then on all three versions mixed together.

The Result: The AI trained on all three versions was the most flexible and human-like. It learned that there isn't just one "right" way to say something. It became better at capturing the meaning rather than just copying words. It's like teaching a student by showing them three different ways to write an essay; they learn the core idea better than if you only showed them one rigid template.

4. Making the AI "Tough" (Handling Noise)

Old manuscripts are often damaged. Letters might be missing (holes in the page), smudged, or misread by scanners. The authors wanted to know: Can we make the AI tough enough to handle these mistakes?

They simulated this by intentionally "breaking" their training data. They randomly deleted letters, swapped them, or replaced them with similar-looking ones, just like a damaged manuscript.

The Result:

If you train the AI on perfect, clean text, it fails miserably when it sees a damaged page.
If you train the AI on 50% damaged text, it becomes a superhero. It learns to guess the missing pieces and ignore the smudges.
The sweet spot was training with 50% noise. This created a translator that was robust enough to handle real-world, dusty, damaged ancient books without falling apart.

The Final Takeaway

The authors successfully built the first reliable translator for Coptic to French. Their secret sauce wasn't just one thing; it was a combination of:

Training a model specifically for this pair (not using a detour).
Starting with a model that already knew the ancient language or its "cousin" (Hieroglyphs).
Feeding it many different ways to say the same thing (stylistic variety).
Deliberately "ruining" half of the training data so the AI learned to be tough against damage.

They released these two best models (one based on the Polyglot and one on the Hieroglyph expert) so that Egyptologists and historians can finally use AI to unlock the secrets of ancient Coptic manuscripts, even if those manuscripts are a bit beat up.

Technical Summary: Neural Machine Translation for Coptic-French

Problem Statement

Neural Machine Translation (NMT) has advanced significantly for modern languages, yet historical and ancient languages remain underrepresented due to data scarcity. Coptic, the final stage of the Egyptian language family, is critical for Egyptology and Early-Christian studies but lacks robust NMT tools, particularly for translation into French. Existing resources are limited primarily to English or Arabic translation. This paper addresses the gap by systematically investigating strategies to translate Coptic into French, focusing on four core challenges inherent to low-resource ancient language translation:

Translation Strategy: Whether to translate directly or use a pivot language (e.g., English).
Model Selection: How to choose the optimal pre-trained model for fine-tuning.
Data Diversity: The impact of training on multiple stylistic versions of the same text.
Robustness: How to handle noise and corruption common in ancient manuscripts (e.g., missing characters, OCR errors, spelling variations).

Methodology

The authors constructed a comprehensive pipeline using aligned biblical corpora to evaluate four distinct experimental stages:

1. Dataset Construction and Preprocessing

Source: Coptic texts were extracted from the Coptic SCRIPTORIUM project (Sahidic New and Old Testaments).
Target: Three public-domain French Bible translations (Louis Segond, Crampon, and Darby) were selected for their stylistic diversity.
Alignment: Verses were aligned automatically using verse IDs.
Romanization: Since standard Transformer vocabularies do not support original Coptic Unicode, the uroman tool was used to romanize the Coptic text.
Split: A book-level split was employed to prevent data contamination. Books (1 Corinthians, Mark, Galatians, Hebrews) were held out for testing, while the remaining 59 books formed the training set.

2. Experimental Stages

The study evaluated four specific research questions through controlled experiments:

Pertinence of Fine-Tuning: The authors compared four approaches:
1. Fine-tuning a Coptic-specific model (Enis) to French.
2. Using a multilingual model (Helsinki) out-of-the-box.
3. Zero-shot prompting of the Coptic model to output French.
4. A pivot strategy (Coptic $\to$ English $\to$ French).
Pre-Trained Model Choice: Four pre-trained models were fine-tuned on the Coptic-French corpus:
- Enis: A Coptic-to-English model.
- Helsinki: A multilingual Bible-trained model (Coptic/French).
- T5: A general-purpose model (French exposure, no Coptic).
- Hiero: A Hieroglyphic-to-English/German model (related language family).
Fine-Tuning Dataset Choice: The best-performing models (Helsinki and Hiero) were trained on individual French translations versus a combined dataset of all three translations to assess the benefit of stylistic variation.
Robustness to Noise: To simulate manuscript degradation, the authors injected noise into the training data:
- Lacunae: 2% character deletion.
- Spelling Errors: 2% adjacent character swap.
- OCR Errors: 10% replacement with visually similar characters.
  Models were trained with varying noise levels (0%, 10%, 30%, 50%, 100%) and tested on clean and noisy data.

3. Evaluation Metrics

Given that a single Coptic verse can have multiple valid French translations, lexical overlap metrics (like BLEU) were deemed insufficient. The study utilized semantic similarity and learned metrics:

BERTScore: Contextual token similarity.
BLEURT: Learned metric predicting human quality judgments.
COMET: Multilingual BERT-based metric for translation quality.
METEOR: Unigram matching with synonymy and stemming.

Key Results

1. Translation Strategy

Fine-tuning is essential. The fine-tuned Coptic-to-French model significantly outperformed all alternatives.

Fine-tuning achieved the highest scores across all metrics (e.g., BERTScore 0.820).
Pivot strategies (Coptic $\to$ English $\to$ French) and zero-shot prompting performed poorly, often worse than the baseline models, indicating error compounding and mismatched decoding expectations.
Multilingual models used without fine-tuning performed suboptimally for low-resource pairs.

2. Model Selection

Pre-training on related languages or the target pair is crucial.

The Helsinki model (pre-trained on Coptic and French Bible texts) achieved the best overall performance (BERTScore 0.850, COMET 0.731). Its success is attributed to having both source and target languages in its pre-training corpus and the domain alignment (Bible texts).
The Hiero model (trained on Hieroglyphs, a related language) performed as the second-best model, demonstrating that knowledge transfers effectively between related ancient languages.
Models lacking Coptic or related language exposure (T5) performed the worst.

3. Dataset Diversity

Stylistic diversity improves generalization.

Models trained on a single French translation performed best when tested on that same translation (high lexical match).
However, models trained on all three translations consistently achieved the highest BLEURT and COMET scores across all test sets. This indicates that exposure to multiple valid translations allows the model to learn better semantic representations and generate higher-quality translations overall, even if the wording differs slightly from a specific reference.

4. Robustness to Noise

Controlled noise injection enhances resilience.

Models trained on clean data (0% noise) performed best on clean test data but suffered severe performance drops (up to 53% on BLEURT for Hiero) when tested on noisy data.
Models trained with 50% noise achieved the best trade-off. They maintained high performance on clean data while showing significantly less degradation on noisy inputs (dropping less than 15% on average).
This suggests that for ancient language translation, training with simulated corruption is a necessary strategy to handle real-world manuscript degradation.

Significance and Contributions

The paper claims to present the first systematic study of strategies for translating Coptic into French. Its primary contributions are:

Empirical Validation: It demonstrates that task-specific fine-tuning is superior to pivot-based or zero-shot approaches for ancient language translation.
Transfer Learning: It confirms that models pre-trained on related ancient languages (Hieroglyphs) or the specific language pair (Coptic-French) yield the best results.
Data Strategy: It establishes that training on stylistically diverse corpora (multiple Bible translations) improves semantic robustness and overall translation quality.
Noise Robustness: It provides a practical guideline for low-resource ancient language NMT: injecting approximately 50% noise during training creates models that are resilient to the corruption inherent in historical manuscripts without sacrificing too much performance on clean text.

The authors conclude that these findings offer practical guidance for developing NMT tools for other low-resource or ancient language pairs, extending beyond the specific Coptic-French case. The resulting models and code have been made publicly available.