Improving Korean-English Cross-Lingual Retrieval: A Data-Centric Study of Language Composition and Model Merging

Youngjoon Jang, Junyoung Son, Taemin Lee, Seongtae Hong, Heuiseok LimKorea University {dew1701, s0ny, taeminlee, ghdchlwls123, limhseok}@korea.ac.kr · 2025

Read Improving Korean-English Cross-Lingual Retrieval: A Data-Centric Study of Language Composition and Model Merging by Youngjoon Jang, Junyoung Son, Taemin Lee, Seongtae Hong, Heuiseok LimKorea University {dew1701, s0ny, taeminlee, ghdchlwls123, limhseok}@korea.ac.kr (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-08480.txt.

Description

Youngjoon Jang and co-authors study how the language mix of training data shapes Korean-English cross-lingual retrieval. Search systems that cross languages are vital in a multilingual web, yet builders rarely test what the training blend does. The team built parallel Korean-English datasets and trained retrievers on varied combinations.

A clean trade-off appeared: blends that lift cross-lingual search tend to drag same-language search down, and the two are tightly linked. Model merging, averaging trained models, eased the trade-off and delivered strong results on both sides. The data-centric lesson is practical: curate the language recipe, then merge to keep the best of each training run.

Catalog record

CREATORYoungjoon Jang, Junyoung Son, Taemin Lee, Seongtae Hong, Heuiseok LimKorea University {dew1701, s0ny, taeminlee, ghdchlwls123, limhseok}@korea.ac.kr
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE52 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-08480

Browse the full library · About the Tower of Babel