Lutfiy: MT for Southern Uzbek

• Lutfiy is a neural machine translation model fine-tuned for Southern Uzbek, distinct in phonology, lexicon, and Arabic-script orthography.
• It utilizes a 39,994-sentence parallel corpus and a FLORES+ dev set, incorporating a post-processing half-space restoration for improved textual accuracy.
• Evaluation results demonstrate enhanced translation quality between Southern Uzbek, English, and Northern Uzbek, marking a significant advance for low-resource Turkic languages.

Searching arXiv for the specified paper and closely related context. Lutfiy is a neural machine translation model specifically fine-tuned for Southern Uzbek (ISO 639-3: uzs), a Turkic language variety spoken by around 5 million people in Afghanistan and differing significantly from Northern Uzbek (uzn) in phonology, lexicon, and orthography. Technically, it is a fine-tuned version of Meta’s NLLB-200 distilled 600M model (“nllb-200-distilled-600M”), adapted to handle translation to and from Southern Uzbek, especially in the Arabic script used in Afghanistan. The model and its associated resources were introduced in “Filling the Gap for Uzbek: Creating Translation Resources for Southern Uzbek” (Mamasaidov et al., 20 Aug 2025).

1. Linguistic setting and rationale

Lutfiy was built to address a specific gap in machine translation for Southern Uzbek. Southern Uzbek is spoken across 14 provinces in Afghanistan and is recognized in Afghanistan’s constitution as a potential regional official language, yet it is severely underrepresented in NLP and in major platforms such as Google Translate, generic NLLB, and MADLAD-400 (Mamasaidov et al., 20 Aug 2025). Generic multilingual MT models treat Uzbek as a single macrolanguage (uzb/uzn) and do not model the script-specific and dialectal features of uzs, leading to poor performance on Southern Uzbek.

The distinctiveness of Southern Uzbek is central to the motivation for Lutfiy. The variety differs from Northern Uzbek in phonology, lexicon, and orthography. Southern Uzbek uses Arabic script with only three main vowel letters, ا, و, and ی, plus diacritics that are often omitted, whereas Northern Uzbek primarily uses Latin script and sometimes Cyrillic. Southern Uzbek also uses half-space, or zero-width non-joiner (U+200C), systematically in morphology, including suffixes after vowel-final stems and prefixes from Persian and Arabic (Mamasaidov et al., 20 Aug 2025).

The model name references Lutfi, a 15th-century Central Asian poet whose language forms part of the historical foundation of modern Southern Uzbek. This naming situates the system within a longer literary and linguistic genealogy, while the engineering objective remains narrowly defined: translation to and from Southern Uzbek, especially in Arabic script.

A common misconception is that support for “Uzbek” in a multilingual model is sufficient for Southern Uzbek. The reported results and the linguistic analysis reject that assumption. The paper frames Southern Uzbek not as a marginal orthographic variant of Northern Uzbek, but as a variety whose phonological, lexical, and scriptal properties materially affect MT quality (Mamasaidov et al., 20 Aug 2025).

2. Data resources and corpus construction

The work introduces a set of new resources for Southern Uzbek machine translation. These include a FLORES+ dev set for Southern Uzbek with $N = 997$ sentences and a 39,994-sentence parallel corpus used for fine-tuning (Mamasaidov et al., 20 Aug 2025). The FLORES+ dev set consists of English source sentences from FLORES+ translated by a native Southern Uzbek linguist and then post-reviewed. It was not used for training; instead, it functions as a development and evaluation set for sacreBLEU and chrF++ reporting.

The orthographic standard of the dev set is an important design choice. Its translations strictly adhere to Arabic script conventions, including correct half-space usage for suffixes and prefixes, as well as careful handling of loanwords and standardized affix attachment rules. This matters because the evaluation target is not only semantic adequacy but also orthographic correctness in standardized literary Southern Uzbek (Mamasaidov et al., 20 Aug 2025).

The training corpus aggregates dictionary, literary, and web-sourced material:

Source	Size	Role
Dictionary entries	1,550 pairs	uzn ↔ uzs lexical alignment
Literary corpus	35,865 pairs	sentence-aligned uzn ↔ uzs from 27 books
Web-sourced content	2,579 pairs	en ↔ uzs contemporary and domain-specific vocabulary

The dictionary entries come from Aral’s multilingual phrasebook/dictionary and consist of Northern Uzbek–Southern Uzbek lexical items. The literary corpus, the dominant component, was extracted from 27 books available in both Northern and Southern Uzbek versions and aligned at sentence level. It is rich in literary register, complex syntax, and cultural and historical terminology, but the paper notes a potential bias toward formal, literary style. The web-sourced component was mined from official government websites and reliable online resources, contributing contemporary terms and domain-specific vocabulary related to governance, public services, and institutions (Mamasaidov et al., 20 Aug 2025).

The alignment methodology for the literary corpus used LaBSE embeddings, but with an important adjustment. Because LaBSE does not explicitly include uzs, the authors transliterated Southern Uzbek texts from Arabic to Latin using rule-based scripts before applying LaBSE. The result was approximately 40% more successfully aligned sentence pairs than with Arabic script alone. The alignment procedure combined cosine similarity with length-ratio adjustment and filtered pairs for high similarity and reasonable length alignment (Mamasaidov et al., 20 Aug 2025).

For English–Southern Uzbek web mining, the procedure was reverse-translation-based: Southern Uzbek sentences were translated to English using Gemini-2.0-Flash; LaBSE embeddings were then used to align original English web text with back-translated English; and manual review was applied to ensure semantic fidelity. No explicit train/dev/test split is given for the 39,994 training sentences; all were used for fine-tuning (Mamasaidov et al., 20 Aug 2025).

3. Architecture and fine-tuning procedure

Lutfiy is a fine-tuned instance of NLLB-200-distilled-600M, a Transformer-based sequence-to-sequence MT model with 600M parameters. The base model is multilingual, trained on around 200 languages, and uses a shared SentencePiece tokenizer and language codes. Southern Uzbek is not included in NLLB as a dedicated language, but Northern Uzbek or Uzbek is included; Lutfiy reuses the same tokenizer and vocabulary, leveraging shared Turkic and Arabic/Persian tokens to model uzs (Mamasaidov et al., 20 Aug 2025).

Fine-tuning was performed on all 39,994 sentence pairs in a multidirectional setup covering four translation directions: uzs $\rightarrow$ en, en $\rightarrow$ uzs, uzs $\rightarrow$ uzn, and uzn $\rightarrow$ uzs. The training objective is standard MT cross-entropy over target tokens,

$$\mathcal{L} = -\sum_{t=1}^{T} \log p_\theta\left(y_t \mid y_{<t}, x\right),$$

where $x$ is the source sentence, $y_1,\dots,y_T$ the target tokens, and $\theta$ the model parameters. No additional loss is reported; regularization is via weight decay (Mamasaidov et al., 20 Aug 2025).

The optimization configuration is explicitly specified: Adafactor was used as optimizer, with learning rate $1 \times 10^{-4}$, warmup steps $\rightarrow$0, weight decay $\rightarrow$1, batch size 32, maximum sequence length 128 tokens, and 5000 training steps, corresponding to approximately 2–3 epochs over the dataset. Training ran on a single NVIDIA A100 40GB GPU. No explicit early stopping is reported; the run was fixed at 5000 steps (Mamasaidov et al., 20 Aug 2025).

The paper does not spell out the exact NLLB language tag syntax. It notes, however, that given NLLB conventions, training samples likely include language tags to indicate target direction. This suggests that Lutfiy remains structurally compatible with the NLLB inference framework while specializing its parameters for Southern Uzbek translation.

4. Arabic-script orthography and half-space restoration

A defining technical feature of Lutfiy is its handling of Arabic-script half-space orthography. Southern Uzbek morphology is agglutinative, and orthographic rules depend on stem-final phonology. For suffixes such as $\rightarrow$2, $\rightarrow$3, $\rightarrow$4, and $\rightarrow$5, vowel-final stems ending in ه, و, or ا require separation by half-space (U+200C), whereas consonant-final stems attach suffixes directly. Prefixes, especially from Persian and Arabic, including be-, no-, xoʻsh-, ser-, ba-, ham-, and bad-, are written with a half-space before the stem (Mamasaidov et al., 20 Aug 2025).

Correct half-space placement signals morphological boundaries, affects tokenization and n-gram matching in BLEU and chrF, and is essential for standardized literary Southern Uzbek orthography. The difficulty is that NLLB’s SentencePiece tokenizer normalizes zero-width non-joiner to regular spaces. As a result, half-spaces are lost during tokenization, and the model cannot learn their positions directly; generated text therefore lacks them or places spacing incorrectly (Mamasaidov et al., 20 Aug 2025).

To address this, the authors introduce a separate post-processing half-space restoration model. It is a character-level n-gram model trained on a small corpus of Southern Uzbek text in which half-spaces were manually corrected. The model scans character sequences, identifies stems ending with vowel letters, recognizes suffix strings and common prefixes, and inserts U+200C at positions where half-spaces should appear. The paper characterizes this as a statistical rule-based corrector for half-space placement (Mamasaidov et al., 20 Aug 2025).

This solution is a pragmatic workaround rather than a tokenizer redesign. It preserves compatibility with NLLB infrastructure while recovering orthographic detail after generation. The paper explicitly notes that a fully integrated solution would require changing or extending SentencePiece behavior.

5. Evaluation and comparative performance

Evaluation used sacreBLEU and chrF++ on the FLORES+ Southern Uzbek dev set of 997 sentences. The reported systems include GPT-4.1, Gemini-2.0-Flash-001, Claude-Sonnet-4, NLLB-200-600M, Google Translate, MADLAD400-3B-MT, and Lutfiy with and without half-space restoration (Mamasaidov et al., 20 Aug 2025).

For Southern Uzbek to English, Gemini-2.0-Flash achieved $\rightarrow$6, GPT-4.1 $\rightarrow$7, Claude-Sonnet-4 $\rightarrow$8, NLLB-200-600M $\rightarrow$9, Google Translate $\rightarrow$0, MADLAD400-3B-MT $\rightarrow$1, and Lutfiy $\rightarrow$2. For Southern Uzbek to Northern Uzbek, Gemini-2.0-Flash achieved $\rightarrow$3, Claude-Sonnet-4 $\rightarrow$4, NLLB-200-600M $\rightarrow$5, Google Translate $\rightarrow$6, MADLAD400-3B-MT $\rightarrow$7, and Lutfiy $\rightarrow$8 (Mamasaidov et al., 20 Aug 2025).

For English to Southern Uzbek, the paper reports GPT-4.1 at $\rightarrow$9, Gemini-2.0-Flash at $\rightarrow$0, Claude-Sonnet-4 at $\rightarrow$1, and Lutfiy at $\rightarrow$2 without half-space restoration and $\rightarrow$3 with restoration. For Northern Uzbek to Southern Uzbek, GPT-4.1 achieved $\rightarrow$4, Gemini-2.0-Flash $\rightarrow$5, Claude-Sonnet-4 $\rightarrow$6, and Lutfiy $\rightarrow$7 without half-space restoration and $\rightarrow$8 with restoration (Mamasaidov et al., 20 Aug 2025).

Two findings structure the evaluation. First, LLMs are strongest when translating from Southern Uzbek, especially on uzs $\rightarrow$9 en and uzs $\rightarrow$0 uzn. Second, Lutfiy is particularly strong when generating Southern Uzbek, especially on uzn $\rightarrow$1 uzs, where it outperforms all baselines by a large margin in both BLEU and chrF++ (Mamasaidov et al., 20 Aug 2025).

The half-space post-processing materially changes evaluation outcomes. On uzn $\rightarrow$2 uzs, the shift from $\rightarrow$3 to $\rightarrow$4 corresponds to $\rightarrow$5 and $\rightarrow$6. For en $\rightarrow$7 uzs, the gain is smaller but consistent, from $\rightarrow$8 to $\rightarrow$9. This suggests that much of the BLEU penalty in Arabic-script Southern Uzbek generation arises from orthographic boundary mismatches rather than gross semantic failure (Mamasaidov et al., 20 Aug 2025).

6. Significance, limitations, and released resources

Lutfiy addresses a conjunction of linguistic challenges: agglutinative morphology, Arabic script with limited overt vowels, dual-function letters such as و and ی, heavy Arabic and Persian lexical influence, and divergence from Northern Uzbek in both form and usage. By training on parallel uzs–uzn and uzs–en data and then post-correcting half-spaces, the system directly targets mapping literary Southern Uzbek to Northern Uzbek and English and producing grammatically and orthographically correct Southern Uzbek text in Arabic script (Mamasaidov et al., 20 Aug 2025).

Its practical applications are correspondingly concrete. For Afghan Southern Uzbek speakers, Lutfiy supports translation of Northern Uzbek content into Southern Uzbek orthography and vocabulary and translation of English content into the native script and variety. The paper also identifies use cases for media and educational institutions, including official websites, schools, and TV/radio. For NLP researchers and developers, the model and datasets constitute the first neural MT resources for Southern Uzbek and provide a benchmark for low-resource, script-diverse Turkic varieties (Mamasaidov et al., 20 Aug 2025).

The limitations are also explicit. The training corpus of about 40K sentences remains small, and generalization across conversational or technical domains may therefore be limited. The heavy reliance on literary sources introduces a formal-register bias. The half-space restoration component is a workaround rather than a solution to tokenizer normalization. Evaluation is mainly automatic, with no human evaluation reported. The paper also notes likely weaknesses involving named entities, code-switching with Dari or Pashto, informal or colloquial Southern Uzbek, and ambiguous vowel representation in Arabic script (Mamasaidov et al., 20 Aug 2025).

All datasets, models, and tools are released publicly. The released components include the 997-sentence FLORES+ dev set, the 39,994-sentence parallel corpus, the fine-tuned Lutfiy model, the half-space post-processing scripts or model, and transliteration scripts for Arabic–Latin conversion. The model is hosted on Hugging Face at tahrirchi/lutfiy, and the transliteration scripts are hosted at https://github.com/tahrirchi/uzs-scripts (Mamasaidov et al., 20 Aug 2025).

The broader significance lies in the paper’s positioning of Lutfiy as a first MT baseline for Southern Uzbek and as a methodological example for low-resource varieties embedded within larger macrolanguage categories. The authors connect the work to the Open Language Data Initiative and to the practical “No Language Left Behind” objective. They also outline future directions: expanding corpora, using Gemini, GPT-4, or Claude for synthetic data generation, improving orthographic normalization beyond post-processing, and extending similar techniques to other underrepresented Turkic varieties (Mamasaidov et al., 20 Aug 2025).