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

# CoPiT: Cognitive Pivot Translation for Digraphic Low-Resource Mongolian in the Traditional Script
Source: [https://arxiv.org/html/2607.05849](https://arxiv.org/html/2607.05849)
Burte Bayarsaikhan1Serynn Kim11footnotemark:12Buru Chang1 1Korea University2Hankuk University of Foreign Studies \{burtebay,buru\_chang\}@korea\.ac\.kr serynn@hufs\.ac\.kr

###### Abstract

Low\-resource languages remain challenging for machine translation, and Mongolian is a representative case\. As a digraphic language, Mongolian is written in both Cyrillic and Traditional scripts, which exhibit a severe imbalance in data availability\. While the Cyrillic script is relatively well\-resourced, the Traditional script remains extremely data\-scarce and orthographically ambiguous, leading to substantial performance degradation in direct translation\. We proposeCoPiT, a cognitively motivated pivot\-based translation pipeline that exploits this internal resource hierarchy by routing translation through the Cyrillic script\. The pipeline explicitly resolves script\-induced ambiguity in the Traditional script before translation, enabling more stable and accurate meaning transfer\. Across multiple backbone models and target languages,CoPiTconsistently outperforms direct translation, achieving substantial absolute BLEU improvements together with consistent 1\.5–1\.6×\\timesCOMET gains\. These gains allow strong open\-source models to match or outperform GPT\-4\.1 under comparable evaluation settings\. Beyond inference\-time improvements,CoPiTenables the construction of synthetic parallel data directly from Traditional\-script text, mitigating data scarcity in realistic low\-resource scenarios\. We release a new multi\-script parallel dataset covering Mongolian in both scripts alongside English, Korean, and Russian\. All datasets and code are publicly available at[https://anonymous\.4open\.science/r/anonymous\_project\-76C7](https://anonymous.4open.science/r/anonymous_project-76C7)\.

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

Burte Bayarsaikhan††thanks:Equal contribution\.1Serynn Kim11footnotemark:12Buru Chang††thanks:Corresponding author\.11Korea University2Hankuk University of Foreign Studies\{burtebay,buru\_chang\}@korea\.ac\.krserynn@hufs\.ac\.kr

## 1Introduction

Large language models \(LLMs\) have evolved from primarily monolingual systems to multilingual systems through large\-scale multilingual pre\-training\. Despite this progress, LLMs continue to perform poorly on low\-resource languages \(LRLs\), largely due to limited training data and sparse linguistic resources\. This limitation is particularly evident in machine translation \(MT\), which relies heavily on large\-scale parallel corpora that are often unavailable for LRLs\(Raja and Vats,[2025](https://arxiv.org/html/2607.05849#bib.bib1)\)\.

![Refer to caption](https://arxiv.org/html/2607.05849v1/x1.png)Figure 1:Mongolian digraphia\. The same content can be written in Traditional script \(left\), with multiple surface forms and limited resources, and in better\-resourced Cyrillic script \(right\)\.These challenges are particularly pronounced in Mongolian, a low\-resource language with a uniquedigraphic characteristic, as it is written in bothCyrillicandTraditionalscripts \(see Figure[1](https://arxiv.org/html/2607.05849#S1.F1)\)\. The Cyrillic script, which serves as the dominant modern form, is largely phonemic and exhibits a high degree of consistency in the sound representation\. In contrast, the Traditional script, an archaic writing system, is written vertically, and its letter forms vary by positional context\. Many phonological and morphological distinctions are not explicitly encoded in the script, such that a single written form may correspond to multiple plausible interpretations\. This intrinsic ambiguity distinguishes the Traditional script structurally from its Cyrillic counterpart\. For example, the Gemini\-3\-Pro\-Preview API flags content written in the Traditional Mongolian script as harmful, as shown in Appendix[A\.1](https://arxiv.org/html/2607.05849#A1.SS1)\.

Beyond these structural differences, the two Mongolian scripts are associated with markedly different levels of linguistic and computational resources\. The Cyrillic script has long served as the dominant writing system in modern Mongolian society following its institutional adoption through state language policies\. As a result, the vast majority of textual data and language technologies are developed for the Cyrillic script\. In contrast, the Traditional script, despite representing the same language, is reintroduced into official use through a policy\-driven effort to promote traditional writing practices111The Law on the Mongolian Language, adopted on February 12, 2015, and amended with effect from January 1, 2025, mandates the use of both Cyrillic and Traditional scripts in official state and local government affairs; Official text available at[https://legalinfo\.mn/mn/detail/10932](https://legalinfo.mn/mn/detail/10932)\.\. Owing to this delayed revival, the Traditional script remains severely under\-resourced, giving rise to a hierarchical resource imbalance within a single LRL\. This imbalance leads current NLP systems to exhibit pronounced performance asymmetry between the two scripts, with poorer performance on the Traditional script\.

To address this gap, we introduce a novel pipeline,CoPiT\(CognitivePivotTranslation\), for translating Mongolian written in the Traditional script into other languages\. Our pipeline is motivated by how fluent Mongolian readers process the Traditional script in practice\(Altangerel and Togtokh,[2024](https://arxiv.org/html/2607.05849#bib.bib2)\)\. Due to the dominance of the Cyrillic script in modern usage, many native speakers do not read the Traditional script in a purely direct manner\. Instead, they implicitly map Traditional script forms to their Cyrillic counterparts and interpret them through this familiar representation, resolving script\-dependent ambiguities in a phonologically informed way\. Accordingly, we adopt Mongolian written in the Cyrillic script as an intermediate representation\. While it remains low\-resource, it offers a more phonemically transparent and comparatively better\-resourced representation than the Traditional script, and thus aligns more effectively with both human reading strategies and the strengths of current LLMs\.

Concretely, we decompose the MT task into two sequential stages: \(1\) conversion from Mongolian written in the Traditional script to the Cyrillic script, followed by \(2\) translation from the Cyrillic script to the target language\. This design isolates the script\-specific challenges of the Traditional script from downstream semantic translation while leveraging the relatively greater, though still limited, resources available for the Cyrillic script\.

The Traditional\-to\-Cyrillic conversion is further factorized into three linguistically motivated steps that align with disambiguation processes used by fluent Mongolian readers: vowel harmony recovery, Latin\-assisted normalization, and Cyrillic normalization\. Collectively, these steps transform the underspecified Traditional script into a phonologically specified Cyrillic representation, thereby facilitating subsequent translation into the target language\. This explicit resolution of script\-level ambiguity enablesCoPiTto translate Mongolian written in the Traditional script without relying on large\-scale parallel corpora\. Instead, it requires only a small amount of data for learning script conversion\.

Across reference\-based, reference\-free, and human evaluations,CoPiTconsistently improves Traditional Mongolian translation into English, Korean, and Russian\. These improvements are robust across proprietary and open\-source backbones, with fine\-tuned open\-sourceCoPiTmodels reaching or surpassing GPT\-4\.1 direct\-translation performance\. Beyond inference\-time translation,CoPiTcan be used to construct synthetic parallel data from real Traditional\-script sources, yielding 8,034 sentence pairs aligned with English, Korean, and Russian\. Experiments show that this synthetic data enables Target→\\rightarrowTraditional Mongolian translation while improving Traditional Mongolian→\\rightarrowTarget translation, demonstrating the broader utility ofCoPiTunder realistic low\-resource constraints\.

![Refer to caption](https://arxiv.org/html/2607.05849v1/x2.png)Figure 2:Overview ofCoPiT, a pipeline translating Mongolian from the Traditional script via the Cyrillic script\.Our contributions are as follows:

- •We proposeCoPiT, a cognitively motivated multi\-stage pipeline that translates Traditional\-script Mongolian by resolving script\-induced ambiguity through Cyrillic pivoting\.
- •We release a multi\-script parallel corpus aligning Traditional and Cyrillic Mongolian with English, Korean, and Russian, addressing a resource gap for digraphic low\-resource MT\.
- •We show thatCoPiTconsistently outperforms direct translation across backbones and target languages, with fine\-tuned open\-source models matching or exceeding GPT\-4\.1 direct\-translation performance\.
- •We show thatCoPiT\-generated synthetic data enables bidirectional translation, improving both Target→\\rightarrowTraditional Mongolian and Traditional Mongolian→\\rightarrowTarget directions\.

## 2Related Work

### 2\.1Low\-Resource Language Processing

Prior work on low\-resource language processing can be broadly categorized into three lines of research\.Data\-centricmethods aim to expand parallel data through web mining\(Schwenket al\.,[2021a](https://arxiv.org/html/2607.05849#bib.bib5),[b](https://arxiv.org/html/2607.05849#bib.bib6)\), back\-translation\(Sennrichet al\.,[2016](https://arxiv.org/html/2607.05849#bib.bib7); Edunovet al\.,[2018](https://arxiv.org/html/2607.05849#bib.bib8)\), or LLM\-generated synthetic corpora\(de Gibertet al\.,[2025](https://arxiv.org/html/2607.05849#bib.bib9); Waldendorfet al\.,[2025](https://arxiv.org/html/2607.05849#bib.bib10)\)\.Model\-centricmethods leverage multilingual pretrained models\(Conneauet al\.,[2020](https://arxiv.org/html/2607.05849#bib.bib11); Xueet al\.,[2021](https://arxiv.org/html/2607.05849#bib.bib12)\), but their shared representation assumptions often break down for languages with divergent or multiple writing systems\(Liuet al\.,[2024](https://arxiv.org/html/2607.05849#bib.bib13)\)\.Linguistically motivatedapproaches incorporate explicit knowledge such as transliteration\(Maet al\.,[2025](https://arxiv.org/html/2607.05849#bib.bib14)\)or morphology\-aware tokenization\(Chaudharyet al\.,[2018](https://arxiv.org/html/2607.05849#bib.bib15); Nzeyimana,[2024](https://arxiv.org/html/2607.05849#bib.bib16)\)\. While effective for reducing surface\-level variation, these methods do not explicitly model script\-level ambiguity arising from digraphy\.

Our work lies at the intersection of linguistically motivated modeling and data\-centric augmentation, as the proposed pivot not only resolves script\-level ambiguity but also enables the creation of synthetic parallel data for low\-resource translation\.

### 2\.2Mongolian Language Processing

Prior work on Mongolian language processing can be broadly grouped by the scripts they address: the Cyrillic script, the Traditional script, and conversion between the two\. Most studies focus on Mongolian written in the Cyrillic script, including benchmark construction\(Zhanget al\.,[2024](https://arxiv.org/html/2607.05849#bib.bib17);[Tumur\-Ochiret al\.,](https://arxiv.org/html/2607.05849#bib.bib18)\)and language model pre\-training\(Naet al\.,[2024b](https://arxiv.org/html/2607.05849#bib.bib19)\)\. These works typically assume input text written in a single standardized script and do not account for challenges arising from digraphy\. Research on Mongolian written in the Traditional script has largely focused on classical NLP tasks such as morphological analysis\(Liuet al\.,[2020](https://arxiv.org/html/2607.05849#bib.bib20); Zhanget al\.,[2025](https://arxiv.org/html/2607.05849#bib.bib21)\), with limited connection to downstream applications or LLM\-based frameworks\. Work addressing both scripts has primarily framed the problem as script conversion\. Early approaches relied on rule\-based methods\(Li and Sarina,[2011](https://arxiv.org/html/2607.05849#bib.bib22)\), followed by hybrid systems combining linguistic rules with neural models\(Naet al\.,[2023](https://arxiv.org/html/2607.05849#bib.bib23),[2024a](https://arxiv.org/html/2607.05849#bib.bib24)\), encoder–decoder architectures\(Naet al\.,[2022](https://arxiv.org/html/2607.05849#bib.bib25)\), and more recent neural machine translation formulations\(Dulamragchaaet al\.,[2024](https://arxiv.org/html/2607.05849#bib.bib26)\)\.

In this work, we treat script conversion as a structured inference step that resolves script\-level ambiguity and produces a meaning\-consistent intermediate representation, enabling more accurate translation from the Traditional script under digraphic low\-resource conditions\.

## 3CoPiT: Cognitive Pivot Translation

### 3\.1Overview

Our task is to translate Mongolian text written in the Traditional script into a target language\. This is challenging for current LLMs because Mongolian is a digraphic low\-resource language with a strong resource imbalance across scripts\. The Traditional script is severely under\-resourced and orthographically ambiguous, leading to significant performance degradation compared to the Cyrillic script and making direct translation difficult\.

To address this challenge, we proposeCoPiT, a translation pipeline for Mongolian text written in the Traditional script\. As illustrated in Figure[2](https://arxiv.org/html/2607.05849#S1.F2),CoPiTis cognitively motivated by how fluent Mongolian readers process the Traditional script: rather than interpreting it directly, they implicitly map it to the Cyrillic script to resolve script\-induced ambiguities\.CoPiTconsists of two stages: \(1\) Traditional\-to\-Cyrillic pivoting, which includes morphological segmentation, multi\-step script disambiguation, and sentence\-level self\-reflection; and \(2\) translation from the disambiguated Cyrillic representation to the target language\. More details, including the prompt design, are provided in Appendix[A\.11](https://arxiv.org/html/2607.05849#A1.SS11)\.

### 3\.2Traditional\-to\-Cyrillic Pivoting

Morphological Segmentation\.

![Refer to caption](https://arxiv.org/html/2607.05849v1/x3.png)Figure 3:Comparison between whitespace\-based segmentation and morphological segmentation for Mongolian sentence written in the Traditional script\.As illustrated in Figure[3](https://arxiv.org/html/2607.05849#S3.F3), the Traditional script frequently inserts spaces before suffixes rather than attaching them to stems, a convention that makes surface whitespace an unreliable indicator of true morphological boundaries\(Janhunen,[2012](https://arxiv.org/html/2607.05849#bib.bib28)\)\. Morphological segmentation thus requires distinguishing true lexical boundaries from suffix attachment\.

The process begins with a simple whitespace\-based split to obtain candidate tokens, followed by dictionary\-guided merging of stems and suffixes\. A curated suffix dictionary is used to identify suffixes that should be combined with the preceding token to form a single morphological segment\. These suffixes encode essential grammatical information, such as case and verbal categories\. This step produces the morphologically coherent segments required for subsequent script disambiguation steps\.

Multi\-Step Script Disambiguation\.Segmentation yields well\-formed morphological units, but script\-inherent ambiguity may still remain at the segment level\. We address this through a structured multi\-step script disambiguation process illustrated in Figure[4](https://arxiv.org/html/2607.05849#S3.F4)\. The process is implemented using an LLM and progressively incorporates explicit phonological and orthographic constraints\.

Vowel Harmony Recovery Step\.Vowel harmony is a core phonological constraint in Mongolian, requiring vowels within a word to align with either a masculine or feminine harmony class\(Svantesson, Jan Olof,[2024](https://arxiv.org/html/2607.05849#bib.bib27)\)\. In the Traditional script, medial vowel shapes often underspecify this distinction, making harmony class identification nontrivial\. At this step, the harmony class of each segmented unit is inferred\. The LLM determines the harmony class primarily based on the first vowel in the segment, following standard phonological rules, while accounting for neutral vowels that do not independently determine harmony\. Enforcing vowel harmony eliminates interpretations that violate phonological constraints and substantially narrows the space of plausible analyses, providing a strong constraint for subsequent normalization steps\.

Latin\-Assisted Normalization Step\.This step performs explicit phonological normalization by mapping each segmented unit into an intermediate Latin\-based representation that approximates its phonological realization in modern Mongolian\. Rather than direct transliteration, the goal is to make phonological distinctions that are implicit in the Traditional script explicit\. The mapping follows standard phonetic correspondences for vowels, diphthongs, and consonants, while leveraging the recovered vowel harmony class to disambiguate visually ambiguous medial forms\. By rendering underspecified phonological structure explicit, the Latin representation substantially reduces the space of plausible interpretations and provides a clearer foundation for subsequent Cyrillic normalization\.

Cyrillic Normalization Step\.In the final script disambiguation step, each Latin\-normalized segment is mapped to a canonical Cyrillic form\. The Cyrillic script serves as an intermediate normalization space rather than the final translation target\. Guided by the phonologically clarified Latin representation and previously inferred constraints, the LLM resolves remaining orthographic variation and selects a single, well\-formed Cyrillic realization\. This normalized representation provides a stable foundation for downstream translation by enabling reliable comparison among alternatives and filtering out morphologically or semantically incompatible candidates\.

![Refer to caption](https://arxiv.org/html/2607.05849v1/x4.png)Figure 4:Multi\-step script disambiguation via vowel harmony recovery and Latin\-to\-Cyrillic normalization\.Sentence Reconstruction with Self\-Reflection\.Once local ambiguities are resolved, we reconstruct the full sentence in the Cyrillic script by assembling the previously disambiguated segments\. This is necessary because translation requires a globally consistent, sentence\-level representation rather than independently resolved segments\.

We then incorporate a self\-reflection component, following prior work on reflection\-based grammatical correction and machine translation\(Wanget al\.,[2024](https://arxiv.org/html/2607.05849#bib.bib40); Chenet al\.,[2024](https://arxiv.org/html/2607.05849#bib.bib41)\)\. As shown in Figure[5](https://arxiv.org/html/2607.05849#S3.F5), the model reviews locally disambiguated segments and revises choices that create sentence\-level semantic or grammatical inconsistencies\. This allows the model to select a globally coherent Cyrillic representation, resolving distinctions such as lemma choice, case marking, and tense that are difficult to determine from the Traditional\-script surface form alone\. By addressing such non\-local ambiguities, self\-reflection produces a more consistent representation for downstream translation\.

### 3\.3Translation from the Disambiguated Representation

Finally, the disambiguated Cyrillic representation is translated into the target languages—English, Korean, and Russian—using a prompt\-tuned LLM\. Crucially, the MT system no longer operates under script\-induced uncertainty: the ambiguities inherent in the Traditional script have been resolved upstream, reducing error propagation into the translation output\. As a result, the translation model can focus on cross\-lingual meaning transfer rather than script interpretation, yielding translations that better reflect the intended semantics of the original Traditional\-script sentence\. Compared to direct translation, this pipeline yields more stable translation behavior without requiring additional training, by explicitly resolving script\-level ambiguity and enabling intermediate consistency checks\.

![Refer to caption](https://arxiv.org/html/2607.05849v1/x5.png)Figure 5:Sentence\-level self\-reflection for global semantic and grammatical consistency\.
### 3\.4Component\-wise Training

We train each LLM\-based component independently using supervised fine\-tuning with parameter\-efficient adaptation\(Huet al\.,[2022](https://arxiv.org/html/2607.05849#bib.bib39)\)\. Each component is trained with dedicated prompts and objectives, enabling effective specialization in its intended function within the overall pipeline\.

The Vowel Harmony Recovery, Latin\-Assisted Normalization, and Cyrillic Normalization components are trained using*word\-level*supervision from a Traditional–Cyrillic parallel lexical dataset comprising 14,125 entries\. This dataset is derived from a dictionary that aligns Traditional and Cyrillic word forms with phonetic transliterations, with vowel harmony labels automatically assigned and verified by a native Mongolian speaker\. Because the data consist of dictionary\-level word pairs rather than sentence\-level parallel corpora, they can be collected with substantially lower cost and effort, which is particularly advantageous in low\-resource settings\. The proposed components are explicitly designed to leverage this word\-level lexical supervision and apply it compositionally to longer input sequences in the end\-to\-end pipeline\.

The sentence\-level self\-reflection component is trained separately using 2,061 synthetic Cyrillic revision pairs\. To construct this data, sentences written in the Traditional script are first converted into Cyrillic by a base LLM to produce imperfect or noisy hypotheses, which are then paired with their original Cyrillic counterparts\. This data construction process is relatively straightforward in Mongolian due to its digraphic nature, where equivalent content is commonly available in both scripts\. Training on such revision pairs enables the model to correct locally plausible but globally inconsistent analyses at the sentence level\. All datasets used for component\-wise training are publicly released to support reproducibility and future research\.

BackboneMethodEnglishKoreanRussianBLEU\-3/4↑\\uparrowchrF/chrF\+\+↑\\uparrowCOMET↑\\uparrowBLEU\-3/4↑\\uparrowchrF/chrF\+\+↑\\uparrowCOMET↑\\uparrowBLEU\-3/4↑\\uparrowchrF/chrF\+\+↑\\uparrowCOMET↑\\uparrowQwen\-3 \(4B\)Direct0\.54/0\.2113\.34/11\.730\.3730\.74/0\.352\.88/2\.750\.3510\.35/0\.105\.03/4\.650\.266CoPiT \(Zero\-shot\)0\.94/0\.4615\.57/13\.730\.4220\.92/0\.413\.50/3\.360\.3720\.59/0\.3111\.42/9\.370\.312CoPiT \(Fine\-tuned\)9\.40/5\.7029\.70/27\.550\.62810\.78/6\.6511\.15/10\.910\.6393\.45/1\.9918\.60/16\.240\.544Qwen\-3 \(30B\)Direct1\.25/0\.4819\.73/17\.510\.4431\.28/0\.593\.94/3\.850\.4210\.33/0\.129\.55/8\.330\.331CoPiT \(Zero\-shot\)1\.86/0\.9221\.75/19\.280\.4932\.95/1\.436\.08/5\.870\.4980\.83/0\.3913\.41/11\.540\.388CoPiT \(Fine\-tuned\)7\.61/4\.5629\.25/26\.940\.5939\.31/5\.8110\.91/10\.650\.5952\.83/1\.5519\.51/17\.000\.487Ministral\-3 \(3B\)Direct1\.21/0\.4623\.19/20\.210\.5000\.39/0\.250\.65/0\.680\.2360\.60/0\.2014\.08/12\.030\.404CoPiT \(Zero\-shot\)1\.09/0\.2222\.66/19\.820\.4970\.81/0\.303\.61/3\.420\.3930\.69/0\.2616\.34/14\.150\.439CoPiT \(Fine\-tuned\)4\.81/2\.6629\.38/26\.210\.6195\.59/3\.179\.46/8\.880\.5852\.21/1\.0021\.92/19\.080\.547Ministral\-3 \(14B\)Direct1\.79/0\.7721\.44/19\.130\.4671\.44/0\.654\.53/4\.300\.4350\.92/0\.3416\.70/14\.440\.473CoPiT \(Zero\-shot\)3\.02/1\.4026\.02/23\.300\.5493\.37/1\.806\.93/6\.720\.5282\.31/1\.1221\.70/19\.050\.558CoPiT \(Fine\-tuned\)14\.16/9\.5838\.24/35\.410\.70714\.64/9\.8216\.01/15\.300\.7149\.04/5\.8432\.49/29\.120\.728GPT\-4\.1Direct2\.75/1\.3519\.59/17\.660\.4682\.67/1\.395\.30/5\.040\.4541\.57/0\.7315\.86/14\.000\.516CoPiT \(Zero\-shot\)14\.07/9\.8337\.29/24\.590\.66313\.37/9\.0414\.59/14\.050\.6678\.71/5\.6331\.84/28\.530\.694

Table 1:Machine translation performance across languages under reference\-based evaluation\.BackboneMethodEnglishKoreanRussianCOMETKiwi↑\\uparrowAdeq\.↑\\uparrowFluen\.↑\\uparrowCOMETKiwi↑\\uparrowAdeq\.↑\\uparrowFluen\.↑\\uparrowCOMETKiwi↑\\uparrowAdeq\.↑\\uparrowFluen\.↑\\uparrowQwen\-3 \(4B\)Direct0\.0781\.331\.890\.2041\.001\.330\.1671\.001\.00CoPiT \(Zero\-shot\)0\.1831\.002\.000\.1601\.001\.440\.1241\.001\.00CoPiT \(Fine\-tuned\)0\.3522\.443\.000\.4112\.563\.000\.2801\.221\.33Qwen\-3 \(30B\)Direct0\.2041\.112\.330\.3221\.001\.780\.1561\.001\.00CoPiT \(Zero\-shot\)0\.2502\.113\.000\.3762\.003\.000\.2111\.562\.67CoPiT \(Fine\-tuned\)0\.3602\.333\.440\.4192\.672\.890\.3011\.892\.22Ministral\-3 \(3B\)Direct0\.3151\.222\.670\.3441\.001\.440\.2281\.111\.33CoPiT \(Zero\-shot\)0\.2911\.222\.890\.2721\.001\.220\.2371\.001\.00CoPiT \(Fine\-tuned\)0\.4022\.783\.220\.4132\.332\.560\.3381\.892\.22Ministral\-3 \(14B\)Direct0\.3601\.562\.440\.3911\.001\.440\.3601\.001\.33CoPiT \(Zero\-shot\)0\.3092\.112\.330\.4001\.891\.890\.3491\.892\.11CoPiT \(Fine\-tuned\)0\.4614\.224\.440\.4903\.563\.330\.4493\.673\.89GPT\-4\.1Direct0\.2701\.442\.670\.2491\.441\.560\.4291\.442\.67CoPiT \(Zero\-shot\)0\.4704\.334\.560\.4723\.443\.330\.3994\.004\.44

Table 2:Machine translation performance across languages under reference\-free evaluation\.

## 4Experiments

We conduct our experiments to address the following three research questions, which collectively examine translation quality, component\-level effectiveness, and the potential of the proposed pipeline to alleviate low\-resource challenges\.

BackboneMethodEnglishKoreanRussianBLEU\-3/4↑\\uparrowchrF/chrF\+\+↑\\uparrowCOMET↑\\uparrowBLEU\-3/4↑\\uparrowchrF/chrF\+\+↑\\uparrowCOMET↑\\uparrowBLEU\-3/4↑\\uparrowchrF/chrF\+\+↑\\uparrowCOMET↑\\uparrowQwen\-3\(4B\)Direct0\.54/0\.2113\.34/11\.730\.3730\.74/0\.352\.88/2\.750\.3510\.35/0\.105\.03/4\.650\.266CoPiT w/o VHR8\.08/5\.0029\.57/29\.370\.6336\.79/4\.1710\.50/10\.230\.6382\.00/1\.1217\.31/15\.170\.546CoPiT w/o LAN3\.06/1\.5523\.34/21\.280\.5683\.42/1\.937\.25/7\.090\.5580\.74/0\.3913\.04/11\.200\.456CoPiT w/o SR1\.44/0\.7418\.22/16\.360\.5501\.60/0\.885\.19/4\.990\.5400\.44/0\.2010\.50/8\.940\.415CoPiT9\.40/5\.7029\.70/27\.550\.62810\.78/6\.6511\.15/10\.910\.6393\.45/1\.9918\.60/16\.240\.544

Table 3:Ablation study of individual components in the proposedCoPiTacross target languages under reference\-based evaluation, whereVHR,LAN, andSRdenote Vowel Harmony Recovery, Latin\-Assisted Normalization, and Self\-Reflection, respectively\.RQ1\.Does our pipeline improve end\-to\-end machine translation performance over baselines?

RQ2\.How does each component affect overall translation quality?

RQ3\.Can the pipeline contribute to mitigating low\-resource data challenges?

### 4\.1Experimental Setups

Datasets\.Due to the severe data scarcity of Mongolian, it is impractical to collect parallel corpora that are well aligned across languages or to obtain fully human written reference translations\. We therefore construct two datasets for reference\-based and reference\-free evaluation, respectively\.

The first dataset contains 1,031 sentence\-level web\-crawled Traditional–Cyrillic parallel pairs with human\-verified reference translations\. English references are translated from the Cyrillic side and subsequently validated by bilingual speakers\. Korean and Russian references are then generated from the verified English sentences and independently reviewed by fluent speakers of each target language\. This dataset enables reliable reference\-based evaluation across multiple target languages\.

The second dataset comprises 380 news articles, each containing about five sentences on average, in the Traditional and Cyrillic scripts collected from the official Mongolian president’s website\.222[https://president\.mn/en/](https://president.mn/en/)As no reference translations are available, this dataset is used exclusively for reference\-free evaluation\.

Backbone Models\.We adopt Qwen\-3 \(4B, 30B\)\(Yanget al\.,[2025](https://arxiv.org/html/2607.05849#bib.bib29)\)and Ministral\-3 \(3B, 14B\)\(Mistral AI,[2025](https://arxiv.org/html/2607.05849#bib.bib30)\)as state\-of\-the\-art multilingual open\-source backbones, and GPT\-4\.1\(Achiamet al\.,[2023](https://arxiv.org/html/2607.05849#bib.bib32); OpenAI,[2024](https://arxiv.org/html/2607.05849#bib.bib31)\)as representative commercial LLMs, enabling a comprehensive evaluation across model scale, and real\-world deployment settings\.

Evaluation Metrics\.We employ different evaluation metrics depending on the availability of reference translations\. For datasets with references, we use BLEU\-3/4\(Papineniet al\.,[2002](https://arxiv.org/html/2607.05849#bib.bib34)\)and chrF/chrF\+\+\(Popović,[2015](https://arxiv.org/html/2607.05849#bib.bib35),[2017](https://arxiv.org/html/2607.05849#bib.bib36)\)to measure lexical overlap, along with COMET\(Reiet al\.,[2022](https://arxiv.org/html/2607.05849#bib.bib37)\)to assess semantic adequacy\.

For datasets without reference translations, we adopt COMETKiwi\(Reiet al\.,[2023](https://arxiv.org/html/2607.05849#bib.bib38)\), a reference\-free quality estimation metric that scores translation quality from source–hypothesis pairs\. We additionally conduct human evaluation with six fluent bilingual speakers of Mongolian and each target language\. Translations are evaluated along two criteria:Adequacy \(Adeq\.\): Faithful meaning preservation with respect to the source text\.Fluency \(Fluen\.\): Grammaticality and overall naturalness of the translation in the target language\.

Implementation Details\.Open\-source backbone models are fine\-tuned with LoRA, as described in Section[3\.4](https://arxiv.org/html/2607.05849#S3.SS4)\. GPT\-4\.1 is accessed via its APIs without additional fine\-tuning\. Additional implementation details, including hyperparameters, train/validation/test dataset split information, and inference settings, are provided in Appendix[A\.3](https://arxiv.org/html/2607.05849#A1.SS3)\.

### 4\.2Experimental Results

#### 4\.2\.1Machine Translation Evaluation\.

Table[1](https://arxiv.org/html/2607.05849#S3.T1)presents reference\-based evaluation results across English, Korean, and Russian\. Across all backbones,CoPiToutperforms direct translation\. Even without training, theCoPiTpipeline yields measurable gains, particularly for Korean and Russian, indicating the effectiveness of script\-level pivoting for typologically distant languages\. For example, on Ministral\-3 \(3B\), zero\-shotCoPiTimproves COMET by \+0\.157 for Korean and \+0\.035 for Russian, while yielding no improvement for English\. With component fine\-tuning, gains become substantial across all languages; on Ministral\-3 \(14B\), COMET increases by \+0\.240 \(English\), \+0\.279 \(Korean\), and \+0\.255 \(Russian\)\. Similar trends are observed for BLEU and chrF/chrF\+\+, where fine\-tunedCoPiTconsistently achieves higher scores across languages\. Overall, larger backbones attain higher absolute performance, and fine\-tunedCoPiTon open\-source models matches or exceeds GPT\-4\.1 under direct translation\.

Table[2](https://arxiv.org/html/2607.05849#S3.T2)reports reference\-free evaluation results using COMETKiwi and human adequacy and fluency ratings\. Consistent with reference\-based findings,CoPiTimproves translation quality across all backbones\. On Ministral\-3 \(14B\), fine\-tunedCoPiTimproves COMETKiwi by \+0\.101 \(English\), \+0\.099 \(Korean\), and \+0\.089 \(Russian\), accompanied by gains in human adequacy and fluency across most conditions\. When applied to GPT\-4\.1, zero\-shotCoPiTalso yields strong improvements, increasing COMETKiwi by \+0\.200 \(English\) and \+0\.223 \(Korean\), highlighting the effectiveness of pivot\-based inference in reference\-scarce settings\. Across both reference\-based and reference\-free evaluations,CoPiTconsistently improves translation quality across model backbones and target languages\. Applying theCoPiTpipeline alone already provides gains in most cases, while fine\-tuning the pipeline components yields the best overall performance\. These results demonstrate that script\-level pivoting is an effective and practical approach for improving low\-resource digraphic machine translation and for substantially reducing the performance gap between open\-source and proprietary models\. Detailed annotation protocols with inter\-annotator agreement analyses and few\-shot prompting baselines are provided in Appendix[A\.2](https://arxiv.org/html/2607.05849#A1.SS2)and[A\.7](https://arxiv.org/html/2607.05849#A1.SS7), respectively\.

BackboneMethodEnglish→TMKorean→TMRussian→TMBLEU\-3/4↑\\uparrowchrF/chrF\+\+↑\\uparrowBLEU\-3/4↑\\uparrowchrF/chrF\+\+↑\\uparrowBLEU\-3/4↑\\uparrowchrF/chrF\+\+↑\\uparrowQwen\-3 \(4B\)Base0\.07/0\.030\.10/0\.120\.10/0\.040\.12/0\.140\.07/0\.020\.10/0\.12Fine\-tuned5\.04/3\.2226\.03/22\.616\.68/4\.2629\.94/26\.214\.43/2\.6825\.08/21\.78Ministral\-3 \(3B\)Base0\.05/0\.021\.02/0\.800\.04/0\.013\.18/2\.420\.04/0\.020\.19/0\.18Fine\-tuned5\.31/3\.2823\.81/20\.646\.99/4\.3928\.38/24\.885\.42/3\.3624\.34/21\.11

Table 4:Reverse\-direction translation performance across languages usingCoPiT\-generated synthetic parallel data\.
#### 4\.2\.2Single\-Component Ablation\.

Table[3](https://arxiv.org/html/2607.05849#S4.T3)presents an ablation study of individual components of the fullCoPiTpipeline across languages under reference\-based metrics for Qwen\-3 \(4B\); full results for Ministral\-3 \(3B\) and pairwise ablations are provided in Appendix[A\.5](https://arxiv.org/html/2607.05849#A1.SS5)\. The full pipeline generally achieves the strongest overall performance, attaining the best or second\-best results in most settings, suggesting that the backbone benefits from interactions among multiple components rather than from any single component in isolation\. Vowel Harmony Recovery contributes to overall performance, with its effect varying across languages and metrics\. Latin\-Assisted Normalization provides complementary but moderate gains, indicating a supporting role rather than a dominant contribution\. Self\-Reflection emerges as the most consistently important component, as its removal leads to substantial performance degradation across all target languages and evaluation metrics\.

#### 4\.2\.3Low\-Resource Mitigation

Synthetic Data Generation\.Using the proposedCoPiTpipeline with Ministral\-3 \(14B\), we construct a new parallel corpus of8,034 sentence pairsaligning Mongolian written in the Traditional script with English, Korean, and Russian\. We clarify that this dataset is not purely synthetic; while the target\-language translations are generated by our pipeline, the Traditional Mongolian source texts are curated from real\-world corpora\. This resource addresses a critical data gap for digraphic low\-resource languages\. We validate its usefulness by reversing the dataset for reverse\-direction translation and training the translation module separately\. All datasets are publicly released to facilitate future research\. A human evaluation and failure mode analysis of theCoPiT\-generated synthetic corpus, along with a data scaling analysis, are provided in Appendix[A\.8](https://arxiv.org/html/2607.05849#A1.SS8),[A\.9](https://arxiv.org/html/2607.05849#A1.SS9)and[A\.10](https://arxiv.org/html/2607.05849#A1.SS10)\.

![Refer to caption](https://arxiv.org/html/2607.05849v1/Figures/11_trft_half.png)Figure 6:COMET scores under pipeline configurations with and without translation module fine\-tuning\.Improving Forward Translation\.To test whether CoPiT\-generated synthetic data improves forward translation, we fine\-tune the translation module on the Cyrillic–Target portion of the corpus and combine it with the pivot mechanism \(CoPiT\+TrFT\)\. We compare it withDirect\(no fine\-tuning\),FT\(direct fine\-tuning on the same synthetic data without pivoting\), andCoPiT \(FT\)\(component\-fine\-tuned CoPiT without translation\-module fine\-tuning\)\. Since FT and CoPiT\+TrFT are trained on identical synthetic data, performance differences generally reflect the structural contribution of theCoPiTpipeline rather than data exposure\.

As shown in Figure[6](https://arxiv.org/html/2607.05849#S4.F6), CoPiT\+TrFT tends to achieve strong COMET scores across most languages and backbones, suggesting thatCoPiT\-generated synthetic data can provide effective supervision for translation; full results across BLEU\-4 and chrF\+\+ are provided in Appendix[A\.6](https://arxiv.org/html/2607.05849#A1.SS6)\. In most settings, CoPiT\+TrFT outperforms FT despite using the same training data, offering evidence that the pivot mechanism contributes independently of the data; CoPiT \(FT\) also performs comparably to FT in several settings, suggesting that the structural benefits of pivoting are substantial even without translation module fine\-tuning with parallel data\. We note that this comparison represents an upper\-bound analysis, as CoPiT is by design trained using only Traditional Mongolian–Cyrillic data without access to target\-language references\.

Reverse\-Direction Translation\.To further demonstrate the practical utility ofCoPiT, we examine whether the synthetic parallel data it generates can enableTarget → Traditional Mongoliantranslation, a direction that is otherwise infeasible or limited due to the lack of parallel resources\. We construct reverse\-direction training data by flipping each pair in theCoPiT\-generated synthetic corpus and fine\-tune the translation module on this reversed data\.

The fine\-tuned models yield generally consistent improvements over base models across languages and backbones\. As shown in Table[4](https://arxiv.org/html/2607.05849#S4.T4), BLEU and chrF scores increase steadily for ENG → TM, KOR → TM, and RUS → TM, suggesting that theCoPiT\-generated synthetic data can provide effective supervision for generating Traditional Mongolian for both backbones\.

We note that evaluation for Traditional Mongolian remains challenging due to the absence of specialized metrics and the limited representation of the script in existing multilingual models\. As a result, semantic metrics such as COMET are unreliable for this language\. Following prior work on low\-resource and morphologically rich languages, we therefore report BLEU and chrF/chrF\+\+ scores as reference\-based evaluation metrics\.

## 5Contribution

In this paper, we introduceCoPiT, a cognitively motivated translation pipeline for digraphic low\-resource languages, focusing on Mongolian in the Traditional script\. Rather than treating Traditional\-to\-Cyrillic conversion as standalone transcription,CoPiTformulates it as structured disambiguation, using Cyrillic as a meaning\-consistent intermediate representation for downstream translation\. This design mirrors native reading practices and decouples script\-level ambiguity from semantic transfer\. Our pipeline integrates linguistically grounded components, vowel harmony recovery, intermediate normalization, and sentence\-level reflection, each targeting a distinct source of underspecification in the Traditional script\. Extensive experiments across multiple backbones and target languages show consistent improvements, including in zero\-shot settings\. Beyond forward translation,CoPiT\-generated synthetic data enables reverse\-direction translation and further improves forward translation quality via translation module fine\-tuning, demonstrating the practical utility of our approach for digraphic low\-resource machine translation\.

## Limitations

The effectiveness of the proposedCoPiTpipeline depends in part on the quality of the downstream Cyrillic\-to\-target translation stage\. This reliance is partly mitigated by the relatively higher availability of Traditional–Cyrillic parallel data and by the steady increase in resources, which support continued improvement of the pivoting and synthetic data generation process\. As shown in Section[4\.2](https://arxiv.org/html/2607.05849#S4.SS2), the resultingCoPiT\-generated synthetic data can further improve downstream translation quality through fine\-tuning\. In addition, the multi\-stage design introduces higher inference latency compared to single\-stage models, as multiple conversion and translation components are sequentially applied\. Detailed latency analysis is provided in Appendix[A\.4](https://arxiv.org/html/2607.05849#A1.SS4)\. Although this limits applicability to strict real\-time settings, the pipeline is more suitable for offline document translation and batch\-oriented low\-resource MT scenarios\. Finally, the coreCoPiTinference pipeline is designed for outbound translation from Traditional Mongolian into non\-Mongolian target languages\. While our experiments show thatCoPiT\-generated synthetic data can support preliminary Target→Traditional Mongolian translation across multiple language pairs and backbones, a fully integrated bidirectional pipeline with dedicated evaluation metrics for Traditional Mongolian remains future work\.

## References

- J\. Achiam, S\. Adler, S\. Agarwal, L\. Ahmad, I\. Akkaya, F\. L\. Aleman, D\. Almeida, J\. Altenschmidt, S\. Altman, S\. Anadkat,et al\.\(2023\)Gpt\-4 technical report\.arXiv preprint arXiv:2303\.08774\.Cited by:[§4\.1](https://arxiv.org/html/2607.05849#S4.SS1.p4.1)\.
- G\. Altangerel and O\. Togtokh \(2024\)Classical mongolian: a textbook for students, scholars, and everyone interested in mongolian\.Reichert Verlag,Wiesbaden, Germany\.External Links:ISBN 9783752008203Cited by:[§1](https://arxiv.org/html/2607.05849#S1.p4.1)\.
- A\. Chaudhary, C\. Zhou, L\. Levin, G\. Neubig, D\. R\. Mortensen, and J\. Carbonell \(2018\)Adapting word embeddings to new languages with morphological and phonological subword representations\.InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing,E\. Riloff, D\. Chiang, J\. Hockenmaier, and J\. Tsujii \(Eds\.\),Brussels, Belgium,pp\. 3285–3295\.External Links:[Link](https://aclanthology.org/D18-1366/),[Document](https://dx.doi.org/10.18653/v1/D18-1366)Cited by:[§2\.1](https://arxiv.org/html/2607.05849#S2.SS1.p1.1)\.
- A\. Chen, L\. Lou, K\. Chen, X\. Bai, Y\. Xiang, M\. Yang, T\. Zhao, and M\. Zhang \(2024\)DUAL\-reflect: enhancing large language models for reflective translation through dual learning feedback mechanisms\.arXiv preprint arXiv:2406\.07232\.Cited by:[§3\.2](https://arxiv.org/html/2607.05849#S3.SS2.p9.1)\.
- A\. Conneau, K\. Khandelwal, N\. Goyal, V\. Chaudhary, G\. Wenzek, F\. Guzmán, E\. Grave, M\. Ott, L\. Zettlemoyer, and V\. Stoyanov \(2020\)Unsupervised cross\-lingual representation learning at scale\.InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics,D\. Jurafsky, J\. Chai, N\. Schluter, and J\. Tetreault \(Eds\.\),Online,pp\. 8440–8451\.External Links:[Link](https://aclanthology.org/2020.acl-main.747/),[Document](https://dx.doi.org/10.18653/v1/2020.acl-main.747)Cited by:[§2\.1](https://arxiv.org/html/2607.05849#S2.SS1.p1.1)\.
- O\. de Gibert, J\. Attieh, T\. Vahtola, M\. Aulamo, Z\. Li, R\. Vázquez, T\. Hu, and J\. Tiedemann \(2025\)Scaling low\-resource MT via synthetic data generation with LLMs\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,C\. Christodoulopoulos, T\. Chakraborty, C\. Rose, and V\. Peng \(Eds\.\),Suzhou, China,pp\. 27662–27680\.External Links:[Link](https://aclanthology.org/2025.emnlp-main.1408/),[Document](https://dx.doi.org/10.18653/v1/2025.emnlp-main.1408),ISBN 979\-8\-89176\-332\-6Cited by:[§2\.1](https://arxiv.org/html/2607.05849#S2.SS1.p1.1)\.
- U\. Dulamragchaa, N\. Batbileg, S\. Batbaatar, P\. Tumurbaatar, B\. Narantuya, B\. Nyamkhuu, T\. Bold, and B\. Enkhtaivan \(2024\)Joint nmt models for text conversion between traditional mongolian script and cyrillic mongolian: a comparative study\.Journal of Institute of Mathematics and Digital Technology6\(1\),pp\. 99–108\.Cited by:[§2\.2](https://arxiv.org/html/2607.05849#S2.SS2.p1.1)\.
- S\. Edunov, M\. Ott, M\. Auli, and D\. Grangier \(2018\)Understanding back\-translation at scale\.InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing,E\. Riloff, D\. Chiang, J\. Hockenmaier, and J\. Tsujii \(Eds\.\),Brussels, Belgium,pp\. 489–500\.External Links:[Link](https://aclanthology.org/D18-1045/),[Document](https://dx.doi.org/10.18653/v1/D18-1045)Cited by:[§2\.1](https://arxiv.org/html/2607.05849#S2.SS1.p1.1)\.
- E\. J\. Hu, Y\. Shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, and W\. Chen \(2022\)LoRA: low\-rank adaptation of large language models\.InInternational Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=nZeVKeeFYf9)Cited by:[§3\.4](https://arxiv.org/html/2607.05849#S3.SS4.p1.1)\.
- J\. A\. Janhunen \(2012\)Mongolian\.London Oriental and African Language Library, Vol\.19,pp\. 320\.External Links:ISBN 9789027202588,ISSN 1382\-3485Cited by:[§3\.2](https://arxiv.org/html/2607.05849#S3.SS2.p2.1)\.
- K\. Krippendorff \(2019\)Content analysis: an introduction to its methodology\.Fourth edition,SAGE Publications, Inc\.,Thousand Oaks, CA\.External Links:[Document](https://dx.doi.org/10.4135/9781071878781),[Link](https://methods.sagepub.com/book/mono/content-analysis-4e/toc)Cited by:[§A\.2](https://arxiv.org/html/2607.05849#A1.SS2.p4.6)\.
- H\. Li and B\. Sarina \(2011\)The study of comparison and conversion about traditional mongolian and cyrillic mongolian\.In2011 4th International Conference on Intelligent Networks and Intelligent Systems,Vol\.,pp\. 199–202\.External Links:[Document](https://dx.doi.org/10.1109/ICINIS.2011.64)Cited by:[§2\.2](https://arxiv.org/html/2607.05849#S2.SS2.p1.1)\.
- N\. Liu, X\. Su, H\. Zhang, G\. Gao, and F\. Bao \(2020\)Incorporating inner\-word and out\-word features for Mongolian morphological segmentation\.InProceedings of the 28th International Conference on Computational Linguistics,D\. Scott, N\. Bel, and C\. Zong \(Eds\.\),Barcelona, Spain \(Online\),pp\. 4638–4648\.External Links:[Link](https://aclanthology.org/2020.coling-main.408/),[Document](https://dx.doi.org/10.18653/v1/2020.coling-main.408)Cited by:[§2\.2](https://arxiv.org/html/2607.05849#S2.SS2.p1.1)\.
- Y\. Liu, C\. Ma, H\. Ye, and H\. Schuetze \(2024\)TransliCo: a contrastive learning framework to address the script barrier in multilingual pretrained language models\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 2476–2499\.External Links:[Link](https://aclanthology.org/2024.acl-long.136/),[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.136)Cited by:[§2\.1](https://arxiv.org/html/2607.05849#S2.SS1.p1.1)\.
- C\. Ma, Y\. Liu, H\. Ye, and H\. Schütze \(2025\)Exploring the role of transliteration in in\-context learning for low\-resource languages written in non\-latin scripts\.InProceedings of the 5th Workshop on Multilingual Representation Learning \(MRL 2025\),pp\. 397–410\.Cited by:[§2\.1](https://arxiv.org/html/2607.05849#S2.SS1.p1.1)\.
- Mistral AI \(2025\)Introducing mistral 3\.Note:[https://mistral\.ai/news/mistral\-3](https://mistral.ai/news/mistral-3)Accessed: 2025\-12\-26Cited by:[§4\.1](https://arxiv.org/html/2607.05849#S4.SS1.p4.1)\.
- M\. Na, F\. Bao, W\. Wang, G\. Gao, and U\. Dulamragchaa \(2023\)Traditional mongolian\-to\-cyrillic mongolian conversion method based on the combination of rules and transformer\.In2023 IEEE 9th International Conference on Cloud Computing and Intelligent Systems \(CCIS\),pp\. 440–445\.Cited by:[§2\.2](https://arxiv.org/html/2607.05849#S2.SS2.p1.1)\.
- M\. Na, F\. Bao, W\. Wang, G\. Gao, and U\. Dulamragchaa \(2024a\)Cyrillic mongolian\-to\-traditional mongolian conversion method based on the transformer\.Journal of Institute of Mathematics and Digital Technology6\(1\),pp\. 120–129\.External Links:[Link](https://mongoliajol.info/index.php/JIMDT/article/view/3599),[Document](https://dx.doi.org/10.5564/jimdt.v6i1.3599)Cited by:[§2\.2](https://arxiv.org/html/2607.05849#S2.SS2.p1.1)\.
- M\. Na, R\. Liu, F\. Bao, and G\. Gao \(2022\)A deep investigation of rnn and self\-attention for the cyrillic\-traditional mongolian bidirectional conversion\.InInternational Conference on Neural Information Processing,pp\. 612–625\.Cited by:[§2\.2](https://arxiv.org/html/2607.05849#S2.SS2.p1.1)\.
- M\. Na, R\. Liu, F\. Bao, and G\. Gao \(2024b\)Pre\-training language model for mongolian with agglutinative linguistic knowledge injection\.In2024 International Joint Conference on Neural Networks \(IJCNN\),pp\. 1–8\.Cited by:[§2\.2](https://arxiv.org/html/2607.05849#S2.SS2.p1.1)\.
- A\. Nzeyimana \(2024\)Low\-resource neural machine translation with morphological modeling\.arXiv preprint arXiv:2404\.02392\.Cited by:[§2\.1](https://arxiv.org/html/2607.05849#S2.SS1.p1.1)\.
- OpenAI \(2024\)GPT\-4\.1\.Note:[https://openai\.com/index/gpt\-4\-1/](https://openai.com/index/gpt-4-1/)Accessed: 2025\-01\-10Cited by:[§4\.1](https://arxiv.org/html/2607.05849#S4.SS1.p4.1)\.
- K\. Papineni, S\. Roukos, T\. Ward, and W\. Zhu \(2002\)Bleu: a method for automatic evaluation of machine translation\.InProceedings of the 40th annual meeting of the Association for Computational Linguistics,pp\. 311–318\.Cited by:[§4\.1](https://arxiv.org/html/2607.05849#S4.SS1.p5.1)\.
- M\. Popović \(2015\)ChrF: character n\-gram f\-score for automatic mt evaluation\.InProceedings of the tenth workshop on statistical machine translation,pp\. 392–395\.Cited by:[§4\.1](https://arxiv.org/html/2607.05849#S4.SS1.p5.1)\.
- M\. Popović \(2017\)ChrF\+\+: words helping character n\-grams\.InProceedings of the second conference on machine translation,pp\. 612–618\.Cited by:[§4\.1](https://arxiv.org/html/2607.05849#S4.SS1.p5.1)\.
- R\. Raja and A\. Vats \(2025\)Parallel corpora for machine translation in low\-resource Indic languages: a comprehensive review\.InProceedings of the Eighth Workshop on Technologies for Machine Translation of Low\-Resource Languages \(LoResMT 2025\),A\. Kr\. Ojha, C\. Liu, E\. Vylomova, F\. Pirinen, J\. Washington, N\. Oco, and X\. Zhao \(Eds\.\),Albuquerque, New Mexico, U\.S\.A\.,pp\. 129–143\.External Links:[Link](https://aclanthology.org/2025.loresmt-1.12/),[Document](https://dx.doi.org/10.18653/v1/2025.loresmt-1.12),ISBN 979\-8\-89176\-230\-5Cited by:[§1](https://arxiv.org/html/2607.05849#S1.p1.1)\.
- R\. Rei, J\. G\. De Souza, D\. Alves, C\. Zerva, A\. C\. Farinha, T\. Glushkova, A\. Lavie, L\. Coheur, and A\. F\. Martins \(2022\)COMET\-22: unbabel\-ist 2022 submission for the metrics shared task\.InProceedings of the Seventh Conference on Machine Translation \(WMT\),pp\. 578–585\.Cited by:[§4\.1](https://arxiv.org/html/2607.05849#S4.SS1.p5.1)\.
- R\. Rei, N\. M\. Guerreiro, D\. Van Stigt, M\. Treviso, L\. Coheur, J\. G\. de Souza, A\. F\. Martins,et al\.\(2023\)Scaling up cometkiwi: unbabel\-ist 2023 submission for the quality estimation shared task\.InProceedings of the Eighth Conference on Machine Translation,pp\. 841–848\.Cited by:[§4\.1](https://arxiv.org/html/2607.05849#S4.SS1.p6.1)\.
- H\. Schwenk, V\. Chaudhary, S\. Sun, H\. Gong, and F\. Guzmán \(2021a\)WikiMatrix: mining 135M parallel sentences in 1620 language pairs from Wikipedia\.InProceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume,P\. Merlo, J\. Tiedemann, and R\. Tsarfaty \(Eds\.\),Online,pp\. 1351–1361\.External Links:[Link](https://aclanthology.org/2021.eacl-main.115/),[Document](https://dx.doi.org/10.18653/v1/2021.eacl-main.115)Cited by:[§2\.1](https://arxiv.org/html/2607.05849#S2.SS1.p1.1)\.
- H\. Schwenk, G\. Wenzek, S\. Edunov, E\. Grave, A\. Joulin, and A\. Fan \(2021b\)CCMatrix: mining billions of high\-quality parallel sentences on the web\.InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing \(Volume 1: Long Papers\),C\. Zong, F\. Xia, W\. Li, and R\. Navigli \(Eds\.\),Online,pp\. 6490–6500\.External Links:[Link](https://aclanthology.org/2021.acl-long.507/),[Document](https://dx.doi.org/10.18653/v1/2021.acl-long.507)Cited by:[§2\.1](https://arxiv.org/html/2607.05849#S2.SS1.p1.1)\.
- R\. Sennrich, B\. Haddow, and A\. Birch \(2016\)Improving neural machine translation models with monolingual data\.InProceedings of the 54th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),K\. Erk and N\. A\. Smith \(Eds\.\),Berlin, Germany,pp\. 86–96\.External Links:[Link](https://aclanthology.org/P16-1009/),[Document](https://dx.doi.org/10.18653/v1/P16-1009)Cited by:[§2\.1](https://arxiv.org/html/2607.05849#S2.SS1.p1.1)\.
- R\. Sennrich and B\. Zhang \(2019\)Revisiting low\-resource neural machine translation: a case study\.arXiv preprint arXiv:1905\.11901\.Cited by:[§A\.10](https://arxiv.org/html/2607.05849#A1.SS10.p2.4)\.
- Svantesson, Jan Olof \(2024\)Vowel harmony in Mongolic languages\.InThe Oxford Handbook of Vowel Harmony,pp\. 774–779\(eng\)\.External Links:ISBN 9780198826804Cited by:[§3\.2](https://arxiv.org/html/2607.05849#S3.SS2.p5.1)\.
- \[34\]D\. Tumur\-Ochir, F\. Cheng, Y\. Murawaki, and C\. ChuEvaluating large language models in mongolian\.Cited by:[§2\.2](https://arxiv.org/html/2607.05849#S2.SS2.p1.1)\.
- J\. Waldendorf, B\. Haddow, A\. Birch, and M\. Klimaszewski \(2025\)Multilingual data filtering using synthetic data from large language models\.InFindings of the Association for Computational Linguistics: EMNLP 2025,C\. Christodoulopoulos, T\. Chakraborty, C\. Rose, and V\. Peng \(Eds\.\),Suzhou, China,pp\. 9317–9334\.External Links:[Link](https://aclanthology.org/2025.findings-emnlp.495/),[Document](https://dx.doi.org/10.18653/v1/2025.findings-emnlp.495),ISBN 979\-8\-89176\-335\-7Cited by:[§2\.1](https://arxiv.org/html/2607.05849#S2.SS1.p1.1)\.
- Y\. Wang, J\. Zeng, X\. Liu, F\. Meng, J\. Zhou, and M\. Zhang \(2024\)Taste: teaching large language models to translate through self\-reflection\.arXiv preprint arXiv:2406\.08434\.Cited by:[§3\.2](https://arxiv.org/html/2607.05849#S3.SS2.p9.1)\.
- L\. Xue, N\. Constant, A\. Roberts, M\. Kale, R\. Al\-Rfou, A\. Siddhant, A\. Barua, and C\. Raffel \(2021\)MT5: a massively multilingual pre\-trained text\-to\-text transformer\.InProceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,K\. Toutanova, A\. Rumshisky, L\. Zettlemoyer, D\. Hakkani\-Tur, I\. Beltagy, S\. Bethard, R\. Cotterell, T\. Chakraborty, and Y\. Zhou \(Eds\.\),Online,pp\. 483–498\.External Links:[Link](https://aclanthology.org/2021.naacl-main.41/),[Document](https://dx.doi.org/10.18653/v1/2021.naacl-main.41)Cited by:[§2\.1](https://arxiv.org/html/2607.05849#S2.SS1.p1.1)\.
- A\. Yang, A\. Li, B\. Yang, B\. Zhang, B\. Hui, B\. Zheng, B\. Yu, C\. Gao, C\. Huang, C\. Lv,et al\.\(2025\)Qwen3 technical report\.arXiv preprint arXiv:2505\.09388\.Cited by:[§4\.1](https://arxiv.org/html/2607.05849#S4.SS1.p4.1)\.
- M\. Zhang, R\. Wang, B\. Xia, Y\. Sun, and X\. Zhao \(2024\)MM\-eval: a hierarchical benchmark for modern mongolian evaluation in llms\.InInternational Conference on Intelligent Multilingual Information Processing,pp\. 352–363\.Cited by:[§2\.2](https://arxiv.org/html/2607.05849#S2.SS2.p1.1)\.
- Z\. Zhang, J\. Ma, N\. Liu, N\. Wu, Y\. Ji, and G\. Liu \(2025\)Dependent syntactic analysis of mongolian based on semi\-supervised self\-training\.In2025 International Joint Conference on Neural Networks \(IJCNN\),Vol\.,pp\. 1–8\.External Links:[Document](https://dx.doi.org/10.1109/IJCNN64981.2025.11228752)Cited by:[§2\.2](https://arxiv.org/html/2607.05849#S2.SS2.p1.1)\.

## Appendix AAppendix

### A\.1Content Filtering in Traditional Script

Figure[7](https://arxiv.org/html/2607.05849#A1.F7)illustrates cases in which the Gemini\-3\-Pro\-Preview API classifies non\-offensive inputs written in the Traditional Mongolian script as harmful, resulting in blocked translation outputs\. This behavior highlights limitations in script\-level content filtering, as discussed in Section[1](https://arxiv.org/html/2607.05849#S1)\.

### A\.2Human Evaluation

We conduct a human evaluation on three articles, comprising a total of 12 sentences\. The evaluation results are reported in Table[2](https://arxiv.org/html/2607.05849#S3.T2)\. For each target language, six annotators are assigned\. All annotators are native Mongolian speakers who are proficient in both Traditional Mongolian and Cyrillic Mongolian scripts, as well as in the corresponding target languages\.

Annotators were provided with written instructions describing the evaluation task and rating criteria\. They were informed that the study evaluates machine translation quality for research purposes only and that no personally identifiable information would be collected\. Participation was voluntary\.

Annotators evaluate translations using a 5\-point Likert scale along two dimensions: Adequacy and Fluency\. The detailed criteria for each score are defined as follows\.Adequacy \(Adeq\.\)measures how well the translation preserves the meaning of the source text\.

- •5 \(Excellent\): The source meaning is fully preserved with no omissions or distortions\.
- •4 \(Good\): The main meaning is preserved with only minor, non\-critical errors\.
- •3 \(Fair\): The translation partially preserves the meaning but contains noticeable errors or omissions\.
- •2 \(Poor\): Only a limited portion of the source meaning is preserved\.
- •1 \(Very Poor\): The translation fails to convey the source meaning or is largely unrelated\.

Fluency \(Fluen\.\)assesses the grammatical correctness and naturalness of the translated text\.

- •5 \(Excellent\): The translation is fully fluent and natural, resembling native\-written text\.
- •4 \(Good\): The translation is mostly fluent with minor grammatical or stylistic issues\.
- •3 \(Fair\): The translation is generally understandable but contains noticeable grammatical or phrasing errors\.
- •2 \(Poor\): The translation is disfluent with frequent errors that hinder readability\.
- •1 \(Very Poor\): The translation is incomprehensible or severely ungrammatical\.

![Refer to caption](https://arxiv.org/html/2607.05849v1/x6.png)Figure 7:Example of a translation request flagged by the safety system\.Language\#Annotators\#SystemsKrippendorff’sα\\alphaAdeq\./Fluen\.English6140\.630 / 0\.263Korean6140\.639 / 0\.274Russian6140\.618 / 0\.635

Table 5:Inter\-annotator agreement \(Krippendorff’sα\\alpha\)LanguageCombinedα\\alphaB1α\\alpha/ B2α\\alphaPairwise\|Δ\|≤1\|\\Delta\|\\leq 1English0\.2630\.378 / 0\.20073%Korean0\.2740\.578 / 0\.13578%Russian0\.6350\.553 / 0\.73881%

Table 6:Fluency agreement statistics\.PhaseData SourceTrain / Val / TestPurposeSec\.TrainingWord\-level lexical pairs14,125 / 3,532 / –VHR, LAN training3\.4TrainingSentence\-level revision pairs2,061 / 343 / –SR training3\.4EvaluationSentence\-level parallel pairs– / – / 1,031Reference\-based evaluation4\.1EvaluationNews articles– / – / 380Reference\-free evaluation4\.1

Table 7:Data splits used for training and evaluation\.Inter\-annotator agreement is measured using Krippendorff’sα\\alpha\(Krippendorff,[2019](https://arxiv.org/html/2607.05849#bib.bib44)\), computed separately for adequacy and fluency \(see Table[5](https://arxiv.org/html/2607.05849#A1.T5)\)\. Adequacy shows moderate agreement across languages \(α=0\.618\\alpha=0\.618–0\.6390\.639\), indicating reasonable consistency in meaning preservation judgments\. Fluency exhibits greater variability across languages \(English:α=0\.263\\alpha=0\.263; Korean:α=0\.274\\alpha=0\.274; Russian:α=0\.635\\alpha=0\.635\), reflecting the subjective nature of fluency assessment, as perceptions of naturalness vary across annotators\. To further examine this fluency pattern, we conducted a per\-batch analysis\. Agreement shown in Table[6](https://arxiv.org/html/2607.05849#A1.T6)varies across groups: some batches show moderate consistency \(e\.g\., Korean B1; Russian B1 and B2\), while others are lower \(e\.g\., English B2; Korean B2\)\. This variability suggests differences in scoring tendencies, where certain annotators systematically use lower or higher portions of the rating scale\.

Importantly, absolute difference analysis indicates that disagreements are predominantly small: 73% \(English\), 78% \(Korean\), and 81% \(Russian\) of pairwise ratings are exact matches or differ by only one level, while extreme disagreements are rare\. This pattern suggests boundary\-level variation in fluency judgments\.

Overall, although fluency assessment involves inherent subjectivity and scale variation, the structured disagreement pattern and limited extreme divergence support the use of the annotations for comparative system analysis\.

### A\.3Implementation and Training Details

All experiments are conducted on two NVIDIA RTX PRO 6000 GPUs using CUDA 13\.0\. For open\-source backbone models, we apply LoRA\-based parameter\-efficient fine\-tuning with a batch size of 4, gradient accumulation of 2–4 steps, a learning rate of1×10−41\\times 10^\{\-4\}with a cosine scheduler, and train the models for 2–3 epochs\. Minor variations within these ranges are introduced only to accommodate differences in model scale, while the overall optimization strategy and evaluation protocol remain identical across models\. For inference, we consistently use greedy decoding for all models to ensure deterministic and reproducible outputs\.

Table[7](https://arxiv.org/html/2607.05849#A1.T7)summarizes the train/validation/test splits, including their data sources, purposes, and relevant sections\. Each component is trained independently on its designated training set, with model selection based on the corresponding validation set\.

### A\.4Inference Latency Analysis

AlthoughCoPiTimproves translation quality, it incurs additional inference latency due to its multi\-stage pipeline\. Table[8](https://arxiv.org/html/2607.05849#A1.T8)shows thatCoPiTrequires approximately 1\.9 seconds more per sample than the direct baseline, primarily due to the Vowel Harmony Recovery, Latin\-Assisted Normalization, and Self\-Reflection components\. In contrast, the direct approach often produces substantially lower\-quality translations, whereasCoPiTconsistently generates higher\-quality outputs across model backbones and target languages, as reflected in Table[1](https://arxiv.org/html/2607.05849#S3.T1)\. This additional latency may be acceptable in offline or batch\-oriented translation settings where translation quality is prioritized\.

MethodVHR, LANSelf\-ReflectionTranslationTotalDirect––3\.68 s3\.68 sCoPiT4\.29 s0\.67 s0\.64 s5\.60 s

Table 8:Inference latency comparison betweenDirectandCoPiT\.VHR, LANstand for Vowel Harmony Recovery, Latin\-Assisted Normalization, respectively\.
### A\.5Component Ablation Analysis

We present two complementary ablation analyses targeting different stages of theCoPiTpipeline\. The first evaluates the contribution of individual components to end\-to\-end translation quality, covering the full pipeline including the translation stage\. The second focuses exclusively on the Traditional\-to\-Cyrillic pivoting stage, examining how component interactions affect script conversion quality prior to translation\.

Single\-Component Ablation\.

BackboneMethodEnglishKoreanRussianBLEU\-3/4↑\\uparrowchrF/chrF\+\+↑\\uparrowCOMET↑\\uparrowBLEU\-3/4↑\\uparrowchrF/chrF\+\+↑\\uparrowCOMET↑\\uparrowBLEU\-3/4↑\\uparrowchrF/chrF\+\+↑\\uparrowCOMET↑\\uparrowQwen\-3\(4B\)Direct0\.54/0\.2113\.34/11\.730\.3730\.74/0\.352\.88/2\.750\.3510\.35/0\.105\.03/4\.650\.266CoPiT w/o VHR8\.08/5\.0029\.57/29\.370\.6336\.79/4\.1710\.50/10\.230\.6382\.00/1\.1217\.31/15\.170\.546CoPiT w/o LAN3\.06/1\.5523\.34/21\.280\.5683\.42/1\.937\.25/7\.090\.5580\.74/0\.3913\.04/11\.200\.456CoPiT w/o SR1\.44/0\.7418\.22/16\.360\.5501\.60/0\.885\.19/4\.990\.5400\.44/0\.2010\.50/8\.940\.415CoPiT9\.40/5\.7029\.70/27\.550\.62810\.78/6\.6511\.15/10\.910\.6393\.45/1\.9918\.60/16\.240\.544Ministral\-3\(3B\)Direct1\.21/0\.4623\.19/20\.210\.5000\.39/0\.250\.65/0\.680\.2360\.60/0\.2014\.08/12\.030\.404CoPiT w/o VHR5\.02/2\.7129\.92/26\.720\.6254\.57/2\.669\.17/8\.590\.5942\.03/1\.0720\.30/17\.620\.542CoPiT w/o LAN3\.86/1\.9528\.64/25\.470\.6073\.40/1\.897\.78/7\.300\.5631\.60/0\.8418\.60/16\.120\.507CoPiT w/o SR2\.46/1\.1626\.41/23\.210\.5771\.68/0\.825\.84/5\.410\.5100\.58/0\.2215\.17/13\.040\.442CoPiT4\.81/2\.6629\.38/26\.210\.6195\.59/3\.179\.46/8\.880\.5852\.21/1\.0021\.92/19\.080\.547

Table 9:Ablation study of individual components in the proposedCoPiTacross target languages under reference\-based evaluation, whereVHR,LAN, andSRdenote Vowel Harmony Recovery, Latin\-Assisted Normalization, and Self\-Reflection, respectively\.Table[9](https://arxiv.org/html/2607.05849#A1.T9)reports full ablation results across both Qwen\-3 \(4B\) and Ministral\-3 \(3B\)\. The trends observed in the main paper are largely consistent across both backbones: the full pipeline generally achieves the strongest overall performance, and Self\-Reflection remains the most consistently important component across all languages and metrics\. For Ministral\-3 \(3B\), the impact of Vowel Harmony Recovery shows backbone\-dependent behavior: its removal yields competitive or higher COMET scores for English and Korean, suggesting that the contribution of explicit phonological regularization depends on backbone\-specific representations and downstream interactions\. Latin\-Assisted Normalization provides complementary but moderate gains across both backbones, indicating a supporting role rather than a dominant contribution\. Overall, these results suggest that the fullCoPiTpipeline benefits from the complementary contributions of all components, with Self\-Reflection playing the most critical role and the impact of individual components varying across backbones and languages\.

ModelMethodchrF / chrF\+\+↑\\uparrowCER↓\\downarrowQwen\-3 \(4B\)\+CoPiTw/o VHR81\.01/77\.200\.10w/o LAT61\.85/55\.710\.23w/o SR51\.20/43\.710\.27Ministral\-3 \(3B\)\+CoPiTw/o VHR78\.98/74\.660\.11w/o LAT72\.50/ 67\.580\.16w/o SR51\.92/ 44\.310\.26

Table 10:Ablation study of single components for Traditional\-to\-Cyrillic script pivoting, whereVHR,LAN, andSRdenote Vowel Harmony Recovery, Latin\-Assisted Normalization, and Self\-Reflection, respectively\.ModelMethodchrF / chrF\+\+↑\\uparrowCER↓\\downarrowQwen\-3 \(4B\)\+CoPiTw/o VHR & LAT60\.87/54\.750\.24w/o VH & SR50\.94/43\.400\.28w/o LAT & SR44\.57/37\.200\.33Ministral\-3 \(3B\)\+CoPiTw/o VHR & LAT70\.09/65\.140\.18w/o VHR & SR51\.91/44\.260\.27w/o LAT & SR49\.33/41\.960\.29

Table 11:Ablation study of pairwise components for Traditional\-to\-Cyrillic script pivoting, whereVHR,LAN, andSRdenote Vowel Harmony Recovery, Latin\-Assisted Normalization, and Self\-Reflection, respectively\.Pairwise Component Ablation\.This analysis isolates the Traditional\-to\-Cyrillic pivoting stage\. Tables[10](https://arxiv.org/html/2607.05849#A1.T10)and[11](https://arxiv.org/html/2607.05849#A1.T11)present single\- and pairwise\-component ablations, evaluated using chrF/chrF\+\+ and Character Error Rate \(CER\)\.

Across both backbones, jointly removing Latin\-Assisted Normalization and Self\-Reflection results in the worst performance, yielding the lowest chrF/chrF\+\+ scores and the highest CER\. This suggests that these two components provide complementary contributions to pivoting performance\. In contrast, removing Vowel Harmony Recovery together with Latin\-Assisted Normalization leads to comparatively better scores, suggesting that Latin\-Assisted Normalization provides limited additional benefit when phonological regularization from Vowel Harmony Recovery is already absent\.

Compared with single\-component ablations in Table[10](https://arxiv.org/html/2607.05849#A1.T10), pairwise removals in Table[11](https://arxiv.org/html/2607.05849#A1.T11)consistently exhibit substantially lower performance, confirming that Traditional\-to\-Cyrillic pivoting quality depends on multiple components operating jointly rather than independently\.

### A\.6Improving Forward Translation\.

![Refer to caption](https://arxiv.org/html/2607.05849v1/Figures/11.5_trft_full.png)Figure 8:Full results across BLEU\-4, chrF\+\+, and COMET under pipeline configurations with and without translation module fine\-tuning, where TrFT denotes fine\-tuning of the translation module on synthetic parallel data\.Figure[8](https://arxiv.org/html/2607.05849#A1.F8)reports full results across BLEU\-4, chrF\+\+, and COMET for both backbones\. The trends observed in the main paper are largely consistent across all metrics: CoPiT\+TrFT achieves the strongest results in most settings, with particularly pronounced gains in BLEU\-4 and chrF\+\+ for English and Korean across both backbones\. FT and CoPiT \(FT\) show mixed relative performance depending on the language and metric; however, we note that this comparison is not data\-matched, as FT fine\-tunes the translation module directly withCoPiT\-generated synthetic data while CoPiT \(FT\) applies only component\-level fine\-tuning without translation module fine\-tuning\. The more informative comparison is between FT and CoPiT\+TrFT, which are trained on identical synthetic data but differ in whether the pivot mechanism is applied: CoPiT\+TrFT consistently outperforms FT across most languages and metrics, providing cleaner evidence that the pivot mechanism contributes independently of the data\. We note that this comparison operates outside the original low\-resource problem setting, in which no Traditional Mongolian–Target parallel data is assumed to be available, and therefore represents an upper\-bound analysis of what becomes possible once the pipeline itself is used to generate synthetic resources\.

### A\.7Few\-shot Prompting Evaluation

Table[12](https://arxiv.org/html/2607.05849#A1.T12)presents machine translation performance under few\-shot prompting\. Few\-shot baselines are additionally evaluated for Ministral\-3 \(3B\) and Qwen\-3 \(4B\) under both Direct andCoPiTsettings; results for other conditions are carried over from Table[1](https://arxiv.org/html/2607.05849#S3.T1)for reference\. BothCoPiTand Direct improve over their respective zero\-shot counterparts with few\-shot prompting; however, fine\-tunedCoPiTstill achieves the best performance across languages and backbones, indicating that supervised adaptation drives the primary gains rather than prompt\-level context alone\.

BackboneMethodEnglishKoreanRussianBLEU\-3/4↑\\uparrowchrF/chrF\+\+↑\\uparrowCOMET↑\\uparrowBLEU\-3/4↑\\uparrowchrF/chrF\+\+↑\\uparrowCOMET↑\\uparrowBLEU\-3/4↑\\uparrowchrF/chrF\+\+↑\\uparrowCOMET↑\\uparrowQwen\-3 \(4B\)Direct0\.54/0\.2113\.34/11\.730\.3730\.74/0\.352\.88/2\.750\.3510\.35/0\.105\.03/4\.650\.266Direct \(Few\-shot\)1\.78/0\.7819\.45/17\.100\.4913\.25/1\.594\.79/4\.660\.5050\.72/0\.3413\.95/11\.650\.486CoPiT \(Zero\-shot\)0\.94/0\.4615\.57/13\.730\.4220\.92/0\.413\.50/3\.360\.3720\.59/0\.3111\.42/9\.370\.312CoPiT \(Few\-shot\)1\.70/0\.7820\.01/17\.580\.4631\.89/0\.844\.78/4\.600\.4340\.91/0\.3915\.42/13\.070\.397CoPiT \(Fine\-tuned\)9\.40/5\.7029\.70/27\.550\.62810\.78/6\.6511\.15/10\.910\.6393\.45/1\.9918\.60/16\.240\.544Ministral\-3 \(3B\)Direct1\.21/0\.4623\.19/20\.210\.5000\.39/0\.250\.65/0\.680\.2360\.60/0\.2014\.08/12\.030\.404Direct \(Few\-shot\)1\.15/0\.4316\.72/14\.390\.4562\.56/1\.054\.65/4\.320\.4740\.96/0\.3917\.77/14\.940\.511CoPiT \(Zero\-shot\)1\.09/0\.2222\.66/19\.820\.4970\.81/0\.303\.61/3\.420\.3930\.69/0\.2616\.34/14\.150\.439CoPiT \(Few\-shot\)1\.12/0\.3823\.54/20\.370\.5142\.33/0\.995\.68/5\.380\.4910\.89/0\.1721\.45/18\.340\.517CoPiT \(Fine\-tuned\)4\.81/2\.6629\.38/26\.210\.6195\.59/3\.179\.46/8\.880\.5852\.21/1\.0021\.92/19\.080\.547

Table 12:Machine translation performance under few\-shot prompting, evaluated under reference\-based metrics\.
### A\.8Human Evaluation of Synthetic Data Quality

To assess the quality ofCoPiT\-generated synthetic data, we conducted a human evaluation in which three bilingual annotators independently rated 20 randomly sampled Cyrillic–Target sentence pairs for each target language\. They evaluate translations using a 3\-point scale along two dimensions\.Adequacy \(Adeq\.\)measures how well the translation preserves the meaning of the source text\.

- •2 \(Good\): The source meaning is fully preserved with no omissions or distortions\.
- •1 \(Acceptable\): The meaning is mostly preserved with minor errors or omissions\.
- •0 \(Poor\): The meaning is largely lost or the translation is unrelated to the source\.

Fluency \(Fluen\.\)assesses the grammatical correctness and naturalness of the translated text\.

- •2 \(Good\): The translation is fully fluent and natural, resembling native\-written text\.
- •1 \(Acceptable\): The translation is mostly fluent with minor grammatical or phrasing issues\.
- •0 \(Poor\): The translation is disfluent with frequent errors that hinder readability\.

This scale was chosen to reduce annotator fatigue while maintaining sufficient granularity for an acceptability judgment, and is consistent with the small but balanced evaluation design across languages, annotators, and dimensions\.

Annotators were provided with written instructions describing the evaluation task and rating criteria\. They were informed that the study evaluates machine translation quality for research purposes only and that no personally identifiable information would be collected\. Participation was voluntary\.

Average adequacy scores are 1\.47 \(English\), 1\.03 \(Korean\), and 1\.12 \(Russian\), while average fluency scores are 1\.75, 1\.22, and 1\.18, respectively, as shown in Table[13](https://arxiv.org/html/2607.05849#A1.T13)\. The consistent pattern of fluency scores exceeding adequacy scores across all three languages suggests that the primary quality limitation of the synthetic data lies in meaning preservation rather than linguistic naturalness, with further statistical failure mode analysis reported in Appendix[A\.9](https://arxiv.org/html/2607.05849#A1.SS9)\.

Inter\-annotator agreement, measured using Krippendorff’sα\\alpha, shows moderate adequacy agreement for English \(α\\alpha=0\.408\) and lower agreement for Korean \(α\\alpha=0\.343\) and Russian \(α\\alpha=0\.335\)\. Fluency agreement is more variable: Korean shows the strongest agreement \(α\\alpha= 0\.511\), Russian shows weak agreement \(α=0\.254\\alpha=0\.254\), and English exhibits agreement \(α=−0\.049\\alpha=\-0\.049\)\. The negative English fluencyα\\alphais attributable to score compression at the upper end of the scale, annotators consistently rate English fluency highly but occasionally differ on boundary cases \(1 vs\. 2\), which inflates observed disagreement relative to chance\. This pattern is consistent with the low English fluency agreement observed in our main human evaluation, suggesting inherent subjectivity in fluency assessment for Mongolian machine translation\. Absolute difference analysis in Table[14](https://arxiv.org/html/2607.05849#A1.T14)confirms that disagreements are predominantly boundary\-level across all languages: 100% \(English\), 95\.0% \(Korean\), and 96\.7% \(Russian\) of pairwise fluency ratings are exact matches or differ by at most one level, while extreme disagreements are negligible \(0%, 5\.0%, and 3\.3% respectively\)\.

LanguageAvg\. ScoreKrippendorff’sα\\alphaAdeq\.Fluen\.Adeq\.Fluen\.English1\.471\.750\.408\-0\.049Korean1\.031\.220\.3430\.511Russian1\.121\.180\.3350\.254

Table 13:Human evaluation results forCoPiT\-generated synthetic data\.
### A\.9Assessment of Synthetic Dataset Noisiness

To assess the noisiness of the synthetic dataset of N=8,034 \(Section[4\.2](https://arxiv.org/html/2607.05849#S4.SS2)\) , we conducted a targeted inspection as an indirect measure of data quality, complementing the downstream performance improvements observed in our experiments\. Specifically, we examined three interpretable failure modes:prompt artifact leakage, sentence\-mode drift, andcode\-mixingacross three target languages\.

LanguageExact \(%\)Within One Level \(%\)Extreme \(%\)English60\.0100\.00\.0Korean55\.095\.05\.0Russian36\.796\.73\.3

Table 14:Pairwise absolute difference analysis of fluency ratings in the synthetic data human evaluation\.Table[15](https://arxiv.org/html/2607.05849#A1.T15)reports the observed error rates\. Overall, artifact leakage occurs most frequently across languages \(4\.10%\), followed by code\-mixing \(3\.67%\) and sentence\-mode drift \(1\.93%\)\. For Korean, code\-mixing appears at a noticeably higher rate than the other error types, while for Russian, artifact leakage constitutes the dominant failure mode\. For English, artifact leakage accounts for 2\.74% of outputs, followed by sentence\-mode drift \(1\.02%\) and code\-mixing \(0\.39%\)\.

Despite these issues, the overall error rates remain relatively low\. Together with the experiments in Section[4\.2](https://arxiv.org/html/2607.05849#S4.SS2)and Appendix[A\.10](https://arxiv.org/html/2607.05849#A1.SS10), this analysis suggests thatCoPiTand its generated datasets are still useful in low\-resource mitigation scenarios despite some noisiness levels\.

LanguageArtifact Leakage \(%\)Sentence\-mode Drift \(%\)Code Mixing \(%\)English2\.741\.020\.39Korean0\.170\.542\.91Russian1\.190\.370\.37Total \(N = 24,102\)4\.101\.933\.67Table 15:Error rates \(%\) across translation outputs in theCoPiT\-generated synthetic dataset\.
### A\.10Synthetic Data Scaling Analysis

We validate the utility of theCoPiT\-generated synthetic corpus by training translation models at increasing data scales \(1K–7K\) using Ministral\-3 \(3B\) and Qwen\-3 \(4B\), and evaluating them on a fixed reference\-based test set\.

Ministral\-3 \(3B\)\.Figure[9\(a\)](https://arxiv.org/html/2607.05849#A1.F9.sf1)presents translation performance as a function of synthetic data scale\. Across all target languages, performance generally improves as more synthetic data is added, with COMET scores increasing by approximately\+\+0\.09 for English,\+\+0\.36 for Korean, and\+\+0\.21 for Russian from the smallest to the largest scale\. At smaller scales, English shows a slight performance decrease from Base to 1K \(approximately−\-0\.03\), while Korean exhibits nearly negligible changes between 1K and 3K\. Such behavior is characteristic of unstable learning dynamics in extremely low\-resource regimes\(Sennrich and Zhang,[2019](https://arxiv.org/html/2607.05849#bib.bib43)\)\. Beyond this range, performance improvements remain stable across languages despite typological differences, suggesting thatCoPiT\-generated data provides reliable supervision as data scale increases\.

Qwen\-3 \(4B\)\. Figure[9\(b\)](https://arxiv.org/html/2607.05849#A1.F9.sf2)presents the corresponding results for Qwen\-3 \(4B\)\. Consistent with Ministral\-3, COMET scores generally improve as synthetic data increases, though the trajectory appears smoother across data scales with fewer fluctuations at lower data regimes\. Together, these results across both backbones suggest thatCoPiT\-generated synthetic data provides consistent and reliable supervision for low\-resource machine translation, regardless of the underlying backbone architecture\.

![Refer to caption](https://arxiv.org/html/2607.05849v1/x7.png)\(a\)Ministral\-3 \(3B\)
![Refer to caption](https://arxiv.org/html/2607.05849v1/x8.png)\(b\)Qwen\-3 \(4B\)

Figure 9:COMET scores at increasingCoPiT\-generated synthetic data scales \(1K–7K\)\.
### A\.11Prompt Templates

This appendix presents the prompts used in our experiments\. Figures[10](https://arxiv.org/html/2607.05849#A1.F10)–[14](https://arxiv.org/html/2607.05849#A1.F14)show the prompts used in each stage of theCoPiT, while Figure[15](https://arxiv.org/html/2607.05849#A1.F15)presents the prompt used for the direct translation baseline\.

![Refer to caption](https://arxiv.org/html/2607.05849v1/x9.png)Figure 10:Prompt used for vowel harmony recovery![Refer to caption](https://arxiv.org/html/2607.05849v1/x10.png)Figure 11:Prompt used for latin normalization![Refer to caption](https://arxiv.org/html/2607.05849v1/x11.png)Figure 12:Prompt used for Cyrillic normalization![Refer to caption](https://arxiv.org/html/2607.05849v1/x12.png)Figure 13:Prompt used for self\-reflection![Refer to caption](https://arxiv.org/html/2607.05849v1/x13.png)Figure 14:Prompt used for translation![Refer to caption](https://arxiv.org/html/2607.05849v1/x14.png)Figure 15:Prompt used for direct translation