@inproceedings{bhaskar-etal-2025-progressive,
title = "Progressive Perturbation with {KTO} for Enhanced Machine Translation of {I}ndian Languages",
author = "Bhaskar, Yash and
Shetye, Ketaki and
Mujadia, Vandan and
Sharma, Dipti Misra and
Krishnamurthy, Parameswari",
editor = "Bouillon, Pierrette and
Gerlach, Johanna and
Girletti, Sabrina and
Volkart, Lise and
Rubino, Raphael and
Sennrich, Rico and
Farinha, Ana C. and
Gaido, Marco and
Daems, Joke and
Kenny, Dorothy and
Moniz, Helena and
Szoc, Sara",
booktitle = "Proceedings of Machine Translation Summit XX: Volume 1",
month = jun,
year = "2025",
address = "Geneva, Switzerland",
publisher = "European Association for Machine Translation",
url = "https://aclanthology.org/2025.mtsummit-1.26/",
pages = "344--352",
ISBN = "978-2-9701897-0-1",
abstract = "This study addresses the critical challenge of data scarcity in machine translation for Indian languages, particularly given their morphological complexity and limited parallel data. We investigate an effective strategy to maximize the utility of existing data by generating negative samples from positive training instances using a progressive perturbation approach. This is used for aligning the model with preferential data using Kahneman-Tversky Optimization (KTO). Comparing it against traditional Supervised Fine-Tuning (SFT), we demonstrate how generating negative samples and leveraging KTO enhances data efficiency. By creating rejected samples through progressively perturbed translations from the available dataset, we fine-tune the Llama 3.1 Instruct 8B model using QLoRA across 16 language directions, including English, Hindi, Bangla, Tamil, Telugu, and Santali. Our results show that KTO-based preference alignment with progressive perturbation consistently outperforms SFT, achieving significant gains in translation quality with an average BLEU increase of 1.84 to 2.47 and CHRF increase of 2.85 to 4.01 compared to SFT for selected languages, while using the same positive training samples and under similar computational constraints. This highlights the potential of our negative sample generation strategy within KTO, especially in low resource scenarios."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="bhaskar-etal-2025-progressive">
<titleInfo>
<title>Progressive Perturbation with KTO for Enhanced Machine Translation of Indian Languages</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yash</namePart>
<namePart type="family">Bhaskar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ketaki</namePart>
<namePart type="family">Shetye</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Vandan</namePart>
<namePart type="family">Mujadia</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dipti</namePart>
<namePart type="given">Misra</namePart>
<namePart type="family">Sharma</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Parameswari</namePart>
<namePart type="family">Krishnamurthy</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-06</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of Machine Translation Summit XX: Volume 1</title>
</titleInfo>
<name type="personal">
<namePart type="given">Pierrette</namePart>
<namePart type="family">Bouillon</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Johanna</namePart>
<namePart type="family">Gerlach</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sabrina</namePart>
<namePart type="family">Girletti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lise</namePart>
<namePart type="family">Volkart</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Raphael</namePart>
<namePart type="family">Rubino</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rico</namePart>
<namePart type="family">Sennrich</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ana</namePart>
<namePart type="given">C</namePart>
<namePart type="family">Farinha</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Marco</namePart>
<namePart type="family">Gaido</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Joke</namePart>
<namePart type="family">Daems</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dorothy</namePart>
<namePart type="family">Kenny</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Helena</namePart>
<namePart type="family">Moniz</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sara</namePart>
<namePart type="family">Szoc</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>European Association for Machine Translation</publisher>
<place>
<placeTerm type="text">Geneva, Switzerland</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">978-2-9701897-0-1</identifier>
</relatedItem>
<abstract>This study addresses the critical challenge of data scarcity in machine translation for Indian languages, particularly given their morphological complexity and limited parallel data. We investigate an effective strategy to maximize the utility of existing data by generating negative samples from positive training instances using a progressive perturbation approach. This is used for aligning the model with preferential data using Kahneman-Tversky Optimization (KTO). Comparing it against traditional Supervised Fine-Tuning (SFT), we demonstrate how generating negative samples and leveraging KTO enhances data efficiency. By creating rejected samples through progressively perturbed translations from the available dataset, we fine-tune the Llama 3.1 Instruct 8B model using QLoRA across 16 language directions, including English, Hindi, Bangla, Tamil, Telugu, and Santali. Our results show that KTO-based preference alignment with progressive perturbation consistently outperforms SFT, achieving significant gains in translation quality with an average BLEU increase of 1.84 to 2.47 and CHRF increase of 2.85 to 4.01 compared to SFT for selected languages, while using the same positive training samples and under similar computational constraints. This highlights the potential of our negative sample generation strategy within KTO, especially in low resource scenarios.</abstract>
<identifier type="citekey">bhaskar-etal-2025-progressive</identifier>
<location>
<url>https://aclanthology.org/2025.mtsummit-1.26/</url>
</location>
<part>
<date>2025-06</date>
<extent unit="page">
<start>344</start>
<end>352</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Progressive Perturbation with KTO for Enhanced Machine Translation of Indian Languages
%A Bhaskar, Yash
%A Shetye, Ketaki
%A Mujadia, Vandan
%A Sharma, Dipti Misra
%A Krishnamurthy, Parameswari
%Y Bouillon, Pierrette
%Y Gerlach, Johanna
%Y Girletti, Sabrina
%Y Volkart, Lise
%Y Rubino, Raphael
%Y Sennrich, Rico
%Y Farinha, Ana C.
%Y Gaido, Marco
%Y Daems, Joke
%Y Kenny, Dorothy
%Y Moniz, Helena
%Y Szoc, Sara
%S Proceedings of Machine Translation Summit XX: Volume 1
%D 2025
%8 June
%I European Association for Machine Translation
%C Geneva, Switzerland
%@ 978-2-9701897-0-1
%F bhaskar-etal-2025-progressive
%X This study addresses the critical challenge of data scarcity in machine translation for Indian languages, particularly given their morphological complexity and limited parallel data. We investigate an effective strategy to maximize the utility of existing data by generating negative samples from positive training instances using a progressive perturbation approach. This is used for aligning the model with preferential data using Kahneman-Tversky Optimization (KTO). Comparing it against traditional Supervised Fine-Tuning (SFT), we demonstrate how generating negative samples and leveraging KTO enhances data efficiency. By creating rejected samples through progressively perturbed translations from the available dataset, we fine-tune the Llama 3.1 Instruct 8B model using QLoRA across 16 language directions, including English, Hindi, Bangla, Tamil, Telugu, and Santali. Our results show that KTO-based preference alignment with progressive perturbation consistently outperforms SFT, achieving significant gains in translation quality with an average BLEU increase of 1.84 to 2.47 and CHRF increase of 2.85 to 4.01 compared to SFT for selected languages, while using the same positive training samples and under similar computational constraints. This highlights the potential of our negative sample generation strategy within KTO, especially in low resource scenarios.
%U https://aclanthology.org/2025.mtsummit-1.26/
%P 344-352
Markdown (Informal)
[Progressive Perturbation with KTO for Enhanced Machine Translation of Indian Languages](https://aclanthology.org/2025.mtsummit-1.26/) (Bhaskar et al., MTSummit 2025)
ACL