💬 NLP

CoPiT: Cognitive Pivot Translation for Digraphic Low-Resource Mongolian in the Traditional Script

The paper introduces CoPiT, a cognitive pivot translation framework that routes low-resource Traditional Mongolian through its better-resourced Cyrillic script to resolve orthographic ambiguity, significantly improving translation quality and enabling synthetic data generation for underrepresented language pairs.

Original authors: Burte Bayarsaikhan, Serynn Kim, Buru Chang

Published 2026-07-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Burte Bayarsaikhan, Serynn Kim, Buru Chang

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 are trying to translate a story written in an ancient, mysterious code (the Traditional Mongolian script) into modern languages like English, Korean, or Russian. The problem is that this ancient code is like a puzzle with missing pieces: the letters look different depending on where they sit in a word, and the same squiggle could mean three different sounds. Because so few people have written digital books in this ancient code, computers (AI) are terrible at reading it directly. They get confused and produce garbage translations.

However, the same story is also written in a much more common, modern code (the Cyrillic script), which is like a clear, standard alphabet. There are millions of examples of this modern code available for computers to learn from.

The paper introduces a new method called CoPiT (Cognitive Pivot Translation). Instead of forcing the computer to guess the meaning of the ancient code directly, CoPiT acts like a smart translator who knows a secret trick used by fluent Mongolian speakers.

The "Secret Trick" Analogy

Think of a fluent Mongolian speaker reading the ancient script. They don't just look at the squiggles and guess; their brain automatically converts those squiggles into the familiar modern letters in their head, figures out the exact sounds, and then understands the meaning.

CoPiT does exactly this, but in steps:

The "Morphological Segmentation" (Cutting the Cake):
The ancient script is messy. It often puts spaces in weird places, like cutting a cake before the frosting is even on. CoPiT first carefully reassembles the pieces, figuring out where the word actually starts and ends, and attaching the correct "sugar sprinkles" (grammar suffixes) to the main "cake" (the root word).
The "Vowel Harmony" (The Musical Rule):
Mongolian words have a musical rule called "vowel harmony." Vowels in a word must match in "tone" (like high or low notes). In the ancient script, this is often hidden. CoPiT acts like a music conductor, listening to the first note of the word and forcing all the other notes to match the correct harmony, narrowing down the possibilities.
The "Latin Bridge" (The Middleman):
To be extra sure, CoPiT temporarily translates the word into a "Latin" version (like the English alphabet) that spells out the exact sounds. This is like writing the ancient code in a phonetic notebook to make sure no sound is missed.
The "Cyrillic Conversion" (The Final Translation):
Now that the sounds are clear, CoPiT writes the word in the clean, modern Cyrillic script. This is the "pivot" point. It's no longer a mysterious ancient puzzle; it's a standard, well-understood text.
The "Self-Reflection" (The Editor):
Before sending the text to the final translator, CoPiT takes a step back and asks, "Does this whole sentence make sense together?" It checks for grammar and logic errors that might have slipped through the individual word checks, acting like a strict editor.
The Final Translation:
Finally, the computer translates this clean, modern Cyrillic text into English, Korean, or Russian. Because the computer is now working with a clear, unambiguous text, the translation is much better.

Why This Matters

The paper shows that this "detour" through the modern script works wonders.

Better Results: Even small, open-source computer models using this method beat the giant, expensive "GPT-4" models that try to translate the ancient script directly.
Creating Data: Because the method is so good at converting the ancient script to the modern one, the researchers used it to create a massive new library of translated sentences (8,000+ pairs). This helps solve the "lack of data" problem for the future, allowing computers to learn better without needing humans to write thousands of new translations from scratch.

In short, CoPiT doesn't try to force the computer to be a genius at the ancient code. Instead, it gives the computer a "cheat sheet" (the modern script) to understand the ancient code first, ensuring the final translation is accurate and natural.

Technical Summary: CoPiT – Cognitive Pivot Translation for Digraphic Low-Resource Mongolian

1. Problem Statement

Machine translation (MT) for low-resource languages (LRLs) remains a significant challenge, particularly for Mongolian, a digraphic language written in two distinct scripts: Cyrillic and Traditional. While the Cyrillic script is relatively well-resourced and phonemically consistent, the Traditional script is severely data-scarce and orthographically ambiguous.

The core difficulties identified are:

Resource Imbalance: The vast majority of textual data and language technologies exist for the Cyrillic script. The Traditional script, despite representing the same language, suffers from a "hierarchical resource imbalance," leading to poor performance in direct translation tasks.
Orthographic Ambiguity: The Traditional script is written vertically with letter forms varying by positional context. Crucially, many phonological and morphological distinctions (such as vowel harmony) are not explicitly encoded. A single written form can correspond to multiple plausible interpretations, causing substantial performance degradation in direct translation.
LLM Limitations: Current Large Language Models (LLMs) struggle with the Traditional script, often failing to resolve these ambiguities or even flagging valid content as harmful due to the script's complexity.

2. Methodology: CoPiT Pipeline

The authors propose CoPiT (Cognitive Pivot Translation), a multi-stage pipeline inspired by how fluent Mongolian readers process the Traditional script. Instead of translating directly from the ambiguous Traditional script to a target language, CoPiT routes translation through the Cyrillic script as an intermediate representation. This approach explicitly resolves script-induced ambiguity before semantic translation occurs.

The pipeline consists of two primary stages:

Stage 1: Traditional-to-Cyrillic Pivoting

This stage converts Traditional Mongolian text into a disambiguated Cyrillic representation through a structured, linguistically motivated process:

Morphological Segmentation:
- Traditional script often inserts spaces before suffixes rather than attaching them to stems, making whitespace an unreliable boundary indicator.
- The system performs a whitespace-based split followed by dictionary-guided merging of stems and suffixes to create morphologically coherent segments.
Multi-Step Script Disambiguation:
- Vowel Harmony Recovery: Infers the vowel harmony class (masculine or feminine) for each segment based on the first vowel and standard phonological rules. This eliminates interpretations that violate phonological constraints.
- Latin-Assisted Normalization: Maps segmented units to an intermediate Latin-based representation. This step makes implicit phonological distinctions explicit, leveraging the recovered vowel harmony to disambiguate visually ambiguous medial forms.
- Cyrillic Normalization: Maps the Latin-normalized segments to a canonical Cyrillic form. Guided by the phonologically clarified Latin representation, the model resolves remaining orthographic variations to select a single, well-formed Cyrillic realization.
Sentence Reconstruction with Self-Reflection:
- The disambiguated segments are assembled into a full sentence.
- A self-reflection component reviews the locally disambiguated segments to identify and revise choices that create sentence-level semantic or grammatical inconsistencies (e.g., lemma choice, case marking, tense). This ensures global coherence before translation.

Stage 2: Translation

The resulting disambiguated Cyrillic representation is translated into the target language (English, Korean, or Russian) using a prompt-tuned LLM. By upstream resolution of script-level ambiguity, the translation model focuses on cross-lingual meaning transfer rather than script interpretation.

Training Strategy

The pipeline components are trained independently using Supervised Fine-Tuning (SFT) with parameter-efficient adaptation (LoRA):

Word-Level Components: Vowel Harmony Recovery, Latin-Assisted Normalization, and Cyrillic Normalization are trained on a curated Traditional–Cyrillic parallel lexical dataset (14,125 entries) derived from dictionaries.
Sentence-Level Component: The Self-Reflection module is trained on synthetic Cyrillic revision pairs (2,061 pairs), constructed by converting Traditional text to Cyrillic via a base LLM and pairing imperfect hypotheses with original Cyrillic counterparts.

3. Key Contributions

CoPiT Pipeline: A cognitively motivated, multi-stage translation pipeline that resolves script-induced ambiguity in Traditional Mongolian by pivoting through the Cyrillic script, avoiding the need for large-scale parallel corpora for the Traditional script.
New Dataset: Release of a multi-script parallel corpus aligning Traditional and Cyrillic Mongolian with English, Korean, and Russian, addressing a critical resource gap for digraphic low-resource MT.
Performance Gains: Demonstration that CoPiT consistently outperforms direct translation across various backbone models (Qwen-3, Ministral-3) and target languages.
Synthetic Data Generation: Validation that the pipeline can generate high-quality synthetic parallel data (8,034 sentence pairs) from real Traditional-script sources, enabling bidirectional translation and improving performance in both directions.

4. Experimental Results

Experiments were conducted on English, Korean, and Russian targets using reference-based (BLEU, chrF, COMET) and reference-free (COMETKiwi, human evaluation) metrics.

Performance Improvement: CoPiT consistently outperformed direct translation.
- Zero-shot: Even without fine-tuning, CoPiT showed measurable gains, particularly for typologically distant languages like Korean and Russian.
- Fine-tuned: With component fine-tuning, gains were substantial. For example, on the Ministral-3 (14B) model, COMET scores increased by +0.240 (English), +0.279 (Korean), and +0.255 (Russian) compared to direct translation.
- Benchmark Comparison: Fine-tuned open-source CoPiT models matched or exceeded the direct translation performance of GPT-4.1.
Ablation Study:
- Self-Reflection was identified as the most critical component; its removal led to substantial performance degradation across all languages.
- Vowel Harmony Recovery and Latin-Assisted Normalization provided complementary gains, with the full pipeline benefiting from the interaction of all components.
Synthetic Data Utility: The 8,034 sentence pairs generated by CoPiT were successfully used to train translation models, improving both Traditional Mongolian $\to$ Target and Target $\to$ Traditional Mongolian directions.

5. Significance and Claims

The paper claims that CoPiT offers a practical solution for digraphic low-resource machine translation by leveraging the internal resource hierarchy of a single language. By explicitly resolving script-level ambiguity through a cognitively inspired pivot, the method:

Enables strong open-source models to compete with proprietary models (like GPT-4.1) in low-resource settings.
Mitigates data scarcity by allowing the construction of synthetic parallel data directly from under-resourced Traditional script sources.
Provides a framework that can be adapted to other digraphic languages where one script is significantly better resourced than the other.

The authors emphasize that this approach does not require large-scale parallel corpora for the low-resource script, relying instead on a small amount of data for script conversion and the inherent strengths of the better-resourced script.