--- title: "Ensemble Methods for Low-Resource — Metinov 2026" description: "This paper presents the author's entry in the LoResMT 2026 Shared Task, which targets translation from Russian into Kyrgyz." authors: ["Adilet Metinov"] year: 2026 venue: "Proceedings for the Ninth Workshop on Technologies for Machine Translation of Low Resource Languages (LoResMT 2026)" source: "https://aclanthology.org/2026.loresmt-1.24/" doi: "10.18653/v1/2026.loresmt-1.24" licence: "CC BY 4.0 (ACL Anthology)" licence_url: "https://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/papers/metinov-2026-ensemble-methods-low-resource-russian/" --- [Elamur Research](https://md.research.elamur.ai/) · [this page in HTML](https://research.elamur.ai/free/papers/metinov-2026-ensemble-methods-low-resource-russian/) [home](https://md.research.elamur.ai/) / [papers](https://md.research.elamur.ai/free/papers/) / [2026](https://md.research.elamur.ai/free/papers/2026/page/2/) / Ensemble Methods for Low-Resource Russian-Kyrgyz Machine Translation: When Diverse Models Beat Better Models # Ensemble Methods for Low-Resource Russian-Kyrgyz Machine Translation: When Diverse Models Beat Better Models **Authors:** Adilet Metinov · **Year:** 2026 · **Venue:** Proceedings for the Ninth Workshop on Technologies for Machine Translation of Low Resource Languages (LoResMT 2026) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/2026.loresmt-1.24/) · [PDF](https://aclanthology.org/2026.loresmt-1.24.pdf) · [DOI 10.18653/v1/2026.loresmt-1.24](https://doi.org/10.18653/v1/2026.loresmt-1.24) · [OpenAlex](https://openalex.org/W7140099592) **Citations:** 0 citations (openalex-snapshot · 2026-10-14) ## Summary This paper presents the author's entry in the LoResMT 2026 Shared Task, which targets translation from Russian into Kyrgyz. It asks whether a strong system can be built for this low-resource pair without fine-tuning or extra parallel data. The author ran several open-weight models: three NLLB-200 variants and DeepSeek outputs, including DeepSeek-R1 run locally. For each sentence, the system picks the candidate that agrees most with the other candidates. Agreement is measured as the average Jaccard similarity over character trigrams. Voting over five systems reached 49.31 CHRF++, which placed third in the track and beat the best single model by 1.37 points. The study also reports several surprising results. Adding individually weaker DeepSeek outputs to the NLLB-only ensemble raised the score. Heuristic quality filtering and output cleanup both lowered results slightly. The author attributes this to the reference translations containing similar artifacts. For under-resourced languages like Kyrgyz, the work suggests that diverse models combined by plain consensus can be a cheap and competitive baseline. ## Key points - Individual test-set CHRF++ scores were 47.94 for NLLB-200 3.3B, 47.34 for 1.3B, 46.59 for 600M (distilled), 46.73 for DeepSeek (basic) and 45.80 for DeepSeek-R1. - Ensemble size mattered. Three NLLB models scored 48.22, five models scored 49.31 (the best result) and seven models scored 49.17. - Filtering candidates before voting reached at most 49.23, at thresholds t=40 and t=50. The filters flagged English words, untranslated Russian, encoding artifacts and odd length ratios. Punctuation and spacing cleanup reduced the score from 49.31 to 49.27. - Every one of the five systems was picked for some sentences. NLLB 1.3B was chosen 31.8% of the time and DeepSeek (basic) 5.9% (136 sentences). - The two leading teams scored 51.03 and 51.02 CHRF++, about 1.7 points higher. Stated limitations are: - only one language pair was tested; - there was no human evaluation; - running several models raises inference cost. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* Fields: [Linguistics & languages](https://md.research.elamur.ai/free/shelves/linguistics/) · [Machine translation](https://md.research.elamur.ai/free/shelves/machine-translation/) Topics: [Machine translation](https://md.research.elamur.ai/free/topics/machine-translation/) ## Cite The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{metinov2026ensemble, title = {{Ensemble Methods for Low-Resource Russian-Kyrgyz Machine Translation: When Diverse Models Beat Better Models}}, author = {Metinov, Adilet}, booktitle = {Proceedings for the Ninth Workshop on Technologies for Machine Translation of Low Resource Languages (LoResMT 2026)}, year = {2026}, doi = {10.18653/v1/2026.loresmt-1.24}, url = {https://doi.org/10.18653/v1/2026.loresmt-1.24} } ``` - [← previous English Word Sense Disambiguation in 2026: When the Labels Become the Bottleneck](https://md.research.elamur.ai/free/papers/lex-2026-philippov-english-wsd-labels/) - [next → Estonian WinoGrande Dataset: Comparative Analysis of LLM Performance on Human and Machine Translation](https://md.research.elamur.ai/free/papers/ojastu-2026-estonian-winogrande-dataset-comparative-analysis/) - [↑ back to papers of 2026](https://md.research.elamur.ai/free/papers/2026/page/2/) ## References 2 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. NLLB Team, Marta R. Costa-jussà, James Cross, and 1 others. 2022. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672. [find](https://scholar.google.com/scholar?q=NLLB+Team%2C+Marta+R%2E+Costa%2Djuss%C3%A0%2C+James+Cross%2C+and+1+others%2E+2022%2E+No+language+left+behind%3A+Scaling+human%2Dcentered+machine+translation%2E+arXiv+preprint+arXiv%3A2207%2E04672%2E) 2. Different model architectures better suited for Turkic languages • Access to additional training resources Our approach, using only off-the-shelf open-weight models without fine-tuning, represents a strong baseline for resource-constrained settings. [find](https://scholar.google.com/scholar?q=Different+model+architectures+better+suited+for+Turkic+languages+%E2%80%A2+Access+to+additional+training+resources+Our+approach%2C+using+only+off%2Dthe%2Dshelf+open%2Dweight+models+without+fine%2Dtuning%2C+represents+a+s)