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

2026-06-26 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, medium

Summary

A data-centric study, presented at the 1st Workshop on Multilinguality in the Era of Large Language Models (MeLLM 2026) in San Diego, United States, investigates the impact of training data composition on Cross-Lingual Information Retrieval (CLIR) and Mono-Lingual Information Retrieval (Mono-IR) performance. Researchers constructed parallel Korean-English datasets and trained multilingual retrieval models using various language combinations. Experiments revealed that language composition significantly influences IR performance, showing an inter-lingual correlation where specific language pairs improve CLIR but can decline Mono-IR. The study demonstrates that simple weight-averaged model merging effectively mitigates this trade-off, achieving strong CLIR results while preserving Mono-IR capabilities. This work highlights the effects of linguistic configuration on both CLIR and Mono-IR, proposing model merging as an optimization strategy.

Key takeaway

For NLP Engineers optimizing multilingual retrieval systems, be aware that training data language composition can create a performance trade-off between Cross-Lingual and Mono-Lingual Information Retrieval. You should consider implementing simple weight-averaged model merging to effectively achieve strong CLIR results while simultaneously preserving your Mono-IR capabilities. This approach offers a viable strategy to optimize performance across both critical tasks.

Key insights

Training data composition impacts cross-lingual and monolingual retrieval, a trade-off effectively mitigated by simple weight-averaged model merging.

Principles

• Training data language composition significantly influences IR performance.
• Optimizing CLIR with specific language pairs can degrade Mono-IR.
• Weight-averaged model merging can balance CLIR and Mono-IR performance.

Method

Construct linguistically parallel Korean-English datasets, train multilingual retrieval models with varied language combinations, then apply simple weight-averaged model merging to optimize performance.

In practice

• Utilize parallel Korean-English datasets for CLIR research.
• Evaluate training data language composition for IR tasks.
• Implement weight-averaged model merging to resolve CLIR/Mono-IR trade-offs.

Topics

• Cross-Lingual Information Retrieval
• Multilingual Retrieval Models
• Training Data Composition
• Model Merging
• Korean-English Datasets
• Information Retrieval

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Merging Methods for Multilingual Knowledge Editing for Large Language Models: An Empirical Odyssey — Shared covariance in vector merging is crucial for effective multilingual knowledge editing in LLMs.
• Beyond Monolingual Deep Research: Evaluating Agents and Retrievers with Cross-Lingual BrowseComp-Plus — Cross-lingual deep research significantly degrades agent performance, revealing both retrieval and agent-side evidence integration challenges.
• When Does Mixing Help? Analyzing Query Embedding Interpolation in Multilingual Dense Retrieval — Query embedding interpolation significantly improves multilingual dense retrieval, but English dominance creates an asymmetric benefit.
• MMed-Bench-IR: A Heterogeneous Benchmark for Multilingual Medical Information Retrieval — MMed-Bench-IR reveals severe cross-lingual performance gaps in medical information retrieval, highlighting the need for benchmarks that assess multilingual and biomedical expertise jointly.
• SHIFT: Semantic Harmonization via Index-side Feature Transformation for Multilingual Information Retrieval — SHIFT is a training-free indexing method that corrects language bias in multilingual retrieval by adjusting document embeddings.
• Evaluating Monolingual and Multilingual Large Language Models for Greek Question Answering: The DemosQA Benchmark — Monolingual LLMs for under-resourced languages require specific evaluation against multilingual counterparts.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.