@inproceedings{koirala-etal-2026-nitm,
title = "{NITM} {AI} Lab at {WMT} 2026 Shared Task: Low-Resource {A}rabic-{A}sian Language Translation",
author = "Koirala, Nabin and
Bhetwal, Rhythm and
Choudhury, Nurul",
editor = "Haddow, Barry and
Kocmi, Tom and
Koehn, Philipp and
Monz, Christof",
booktitle = "Proceedings of the Eleventh Conference on Machine Translation",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.wmt-main.174/",
pages = "2324--2328",
ISBN = "979-8-89176-474-3",
abstract = "This paper describes the NITM AI Lab submissions to the WMT26 Low-Resource Arabic-Asian Translation shared task. We participate in two translation directions: English-to-Arabic (Sub-Task 1A) and Arabic-to-Hindi (Sub-Task 2B). To address the data scarcity inherent to these language pairs, we compare three neural architectures: a domain-specialized bilingual model (OPUS-MT-TC-Big), a distilled massively multilingual model (NLLB-200Distilled-600M), and a generalist multilingual sequence-to-sequence transformer (mBART50-Many-to-Many). Our systems contrast full-parameter fine-tuning with parameter-efficient fine-tuning (PEFT) via Low-Rank Adaptation (LoRA), and we measure the effect of rule-based orthographic and numerical normalization together with targeted data augmentation from OPUS. On the official blind test set, the domain-specialized OPUS-MT model is the strongest system for English-to-Arabic, while LoRA-tuned NLLB-200 performs best for Arabic-to-Hindi, with OPUS augmentation giving a small additional gain in the latter direction."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="koirala-etal-2026-nitm">
<titleInfo>
<title>NITM AI Lab at WMT 2026 Shared Task: Low-Resource Arabic-Asian Language Translation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Nabin</namePart>
<namePart type="family">Koirala</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rhythm</namePart>
<namePart type="family">Bhetwal</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nurul</namePart>
<namePart type="family">Choudhury</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the Eleventh Conference on Machine Translation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Barry</namePart>
<namePart type="family">Haddow</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tom</namePart>
<namePart type="family">Kocmi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Philipp</namePart>
<namePart type="family">Koehn</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Christof</namePart>
<namePart type="family">Monz</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-474-3</identifier>
</relatedItem>
<abstract>This paper describes the NITM AI Lab submissions to the WMT26 Low-Resource Arabic-Asian Translation shared task. We participate in two translation directions: English-to-Arabic (Sub-Task 1A) and Arabic-to-Hindi (Sub-Task 2B). To address the data scarcity inherent to these language pairs, we compare three neural architectures: a domain-specialized bilingual model (OPUS-MT-TC-Big), a distilled massively multilingual model (NLLB-200Distilled-600M), and a generalist multilingual sequence-to-sequence transformer (mBART50-Many-to-Many). Our systems contrast full-parameter fine-tuning with parameter-efficient fine-tuning (PEFT) via Low-Rank Adaptation (LoRA), and we measure the effect of rule-based orthographic and numerical normalization together with targeted data augmentation from OPUS. On the official blind test set, the domain-specialized OPUS-MT model is the strongest system for English-to-Arabic, while LoRA-tuned NLLB-200 performs best for Arabic-to-Hindi, with OPUS augmentation giving a small additional gain in the latter direction.</abstract>
<identifier type="citekey">koirala-etal-2026-nitm</identifier>
<location>
<url>https://aclanthology.org/2026.wmt-main.174/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>2324</start>
<end>2328</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T NITM AI Lab at WMT 2026 Shared Task: Low-Resource Arabic-Asian Language Translation
%A Koirala, Nabin
%A Bhetwal, Rhythm
%A Choudhury, Nurul
%Y Haddow, Barry
%Y Kocmi, Tom
%Y Koehn, Philipp
%Y Monz, Christof
%S Proceedings of the Eleventh Conference on Machine Translation
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-474-3
%F koirala-etal-2026-nitm
%X This paper describes the NITM AI Lab submissions to the WMT26 Low-Resource Arabic-Asian Translation shared task. We participate in two translation directions: English-to-Arabic (Sub-Task 1A) and Arabic-to-Hindi (Sub-Task 2B). To address the data scarcity inherent to these language pairs, we compare three neural architectures: a domain-specialized bilingual model (OPUS-MT-TC-Big), a distilled massively multilingual model (NLLB-200Distilled-600M), and a generalist multilingual sequence-to-sequence transformer (mBART50-Many-to-Many). Our systems contrast full-parameter fine-tuning with parameter-efficient fine-tuning (PEFT) via Low-Rank Adaptation (LoRA), and we measure the effect of rule-based orthographic and numerical normalization together with targeted data augmentation from OPUS. On the official blind test set, the domain-specialized OPUS-MT model is the strongest system for English-to-Arabic, while LoRA-tuned NLLB-200 performs best for Arabic-to-Hindi, with OPUS augmentation giving a small additional gain in the latter direction.
%U https://aclanthology.org/2026.wmt-main.174/
%P 2324-2328
Markdown (Informal)
[NITM AI Lab at WMT 2026 Shared Task: Low-Resource Arabic-Asian Language Translation](https://aclanthology.org/2026.wmt-main.174/) (Koirala et al., WMT 2026)
ACL