English-to-Prakrit Machine Translation via Multilingual Transfer Learning
This paper demonstrates that English-to-Prakrit machine translation is feasible in low-resource settings by adapting the IndicTrans2 model to route Prakrit through the Hindi language tag, achieving improved BLEU scores despite challenges from data scarcity and dialect mismatches.
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 master translator who is an expert at speaking 22 modern Indian languages, but they have never heard of Prakrit, an ancient language family used in classical Indian literature and religious texts. You want this translator to learn how to translate English into Prakrit, but you only have a tiny dictionary (about 1,500 sentences) to teach them.
This paper describes an experiment where the researchers taught this "master translator" (a model called IndicTrans2) to speak Prakrit without actually adding a new language to its brain.
Here is how they did it, using simple analogies:
The Problem: The "Missing Menu Item"
The translator's software is like a restaurant with a huge menu of modern Indian dishes (languages like Hindi, Marathi, Gujarati). Prakrit isn't on the menu. The kitchen (the computer model) doesn't know how to cook it, and the staff (the tokenizer/vocabulary) doesn't have the right ingredients listed.
Usually, to add a new dish, you'd have to rebuild the kitchen, hire new staff, and rewrite the recipe book. But the researchers wanted to see if they could get the job done without any of that heavy lifting.
The Solution: The "Hindi Disguise"
The researchers used a clever trick: The Script Swap.
Both Hindi and Prakrit are written in the same script (Devanagari), which looks like a shared alphabet. The researchers told the translator: "When you see the request for 'Prakrit', pretend it's 'Hindi'."
- The Analogy: Imagine you are a chef who only knows how to make Italian pasta. You want to make a specific type of ancient Roman pasta, but you don't have the recipe. However, you know that Italian and Roman pasta use the same type of flour and noodles. So, you tell your kitchen, "Treat this order as Italian pasta," and you use the same tools and ingredients you already have.
- The Result: The model didn't learn a new "language" in the traditional sense. Instead, it used its existing knowledge of Hindi (which shares a family tree with Prakrit) and the shared writing system to guess how to construct the ancient sentences.
The Training: A Tiny Library
The researchers didn't have a massive library of books to teach the model. They only had:
- 1,326 sentences to teach it (the training set).
- 148 sentences to check its homework (the validation set).
- 20 sentences for the final exam (the test set).
To make it even harder, the training was done in one dialect of Prakrit (Maharashtri), but the final exam was in a different dialect (Ardhamagadhi). It's like teaching someone to drive in a small, quiet town and then immediately testing them on a busy highway in a different city.
The Results: A Big Leap, But Not Perfect
Before this experiment, the model was terrible at this task, scoring a 1.57 on a scale where higher is better (think of it as a grade of F).
After the "Hindi disguise" training, the score jumped to 14.3.
- What this means: The model went from barely understanding the assignment to producing sentences that actually looked and sounded like the target language. It wasn't perfect, but it was a massive improvement.
The researchers looked at a sample translation and saw that the model could generate words that were structurally similar to the correct answer, even though it sometimes picked the wrong vocabulary or grammar, likely because it was trying to bridge the gap between two different dialects with very little data.
The Bottom Line
The paper concludes that you don't always need to build a new AI from scratch to handle ancient or unsupported languages. If the languages share a writing system and a family history, you can "route" the request through a related, supported language (like Hindi) to get a surprisingly good result.
What the paper does NOT claim:
- It does not say this method is ready for commercial use or that the translations are perfect enough for critical historical analysis.
- It does not claim to have solved the problem for all ancient languages, only that this specific "script-compatible" trick worked for this specific experiment.
- It admits the results are limited by the tiny amount of data and the dialect mismatch, and they did not test human experts to see if the translations were actually meaningful to a scholar.
In short: They proved that with a little creative routing and a shared alphabet, an AI can learn to "speak" an ancient language it was never explicitly taught, using only a handful of examples.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.