Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Balanced Multi-Factor In-Context Learning for Multilingual Large Language Models

Masahiro Kaneko, Alham Fikri Aji, Timothy Baldwin

TL;DR

The paper addresses the sensitivity of multilingual in-context learning (MICL) to example selection by identifying three key factors: semantic similarity, linguistic alignment, and language-specific performance. It introduces Balanced Multi-Factor ICL (BMF-ICL), which defines explicit metrics for each factor and combines them into a total score $score^{(j)} = \alpha score^{(j)}_{sem} + \beta score^{(j)}_{lag} + \gamma score^{(j)}_{per}$ with $\alpha+\beta+\gamma=1$, selecting top-k multilingual examples with weights tuned on development data. The authors validate BMF-ICL on mCSQA and TYDI across four MLLMs, reporting consistent gains over baselines and demonstrating that all three factors, as well as cross-language diversity, contribute to improved performance. The results highlight the practical value of optimizing multilingual example selection for cross-lingual knowledge transfer and suggest broader applicability to diverse generation tasks beyond question answering.

Abstract

Multilingual large language models (MLLMs) are able to leverage in-context learning (ICL) to achieve high performance by leveraging cross-lingual knowledge transfer without parameter updates. However, their effectiveness is highly sensitive to example selection, particularly in multilingual settings. Based on the findings of existing work, three key factors influence multilingual ICL: (1) semantic similarity, (2) linguistic alignment, and (3) language-specific performance. However, existing approaches address these factors independently, without explicitly disentangling their combined impact, leaving optimal example selection underexplored. To address this gap, we propose balanced multi-factor ICL (\textbf{BMF-ICL}), a method that quantifies and optimally balances these factors for improved example selection. Experiments on mCSQA and TYDI across four MLLMs demonstrate that BMF-ICL outperforms existing methods. Further analysis highlights the importance of incorporating all three factors and the importance of selecting examples from multiple languages.

Balanced Multi-Factor In-Context Learning for Multilingual Large Language Models

TL;DR

The paper addresses the sensitivity of multilingual in-context learning (MICL) to example selection by identifying three key factors: semantic similarity, linguistic alignment, and language-specific performance. It introduces Balanced Multi-Factor ICL (BMF-ICL), which defines explicit metrics for each factor and combines them into a total score with , selecting top-k multilingual examples with weights tuned on development data. The authors validate BMF-ICL on mCSQA and TYDI across four MLLMs, reporting consistent gains over baselines and demonstrating that all three factors, as well as cross-language diversity, contribute to improved performance. The results highlight the practical value of optimizing multilingual example selection for cross-lingual knowledge transfer and suggest broader applicability to diverse generation tasks beyond question answering.

Abstract

Multilingual large language models (MLLMs) are able to leverage in-context learning (ICL) to achieve high performance by leveraging cross-lingual knowledge transfer without parameter updates. However, their effectiveness is highly sensitive to example selection, particularly in multilingual settings. Based on the findings of existing work, three key factors influence multilingual ICL: (1) semantic similarity, (2) linguistic alignment, and (3) language-specific performance. However, existing approaches address these factors independently, without explicitly disentangling their combined impact, leaving optimal example selection underexplored. To address this gap, we propose balanced multi-factor ICL (\textbf{BMF-ICL}), a method that quantifies and optimally balances these factors for improved example selection. Experiments on mCSQA and TYDI across four MLLMs demonstrate that BMF-ICL outperforms existing methods. Further analysis highlights the importance of incorporating all three factors and the importance of selecting examples from multiple languages.

Paper Structure

This paper contains 24 sections, 8 equations, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Balanced Multi-Factor In-Context Learning (BMF-ICL)
  3. In-Context Learning
  4. Example Selection via Multi-Factor
  5. Semantic similarity
  6. Linguistic alignment
  7. Language-specific performance
  8. Experiments
  9. Settings
  10. Dataset
  11. Model
  12. ICL Setup
  13. Weight Selection for BMF-ICL
  14. Baseline
  15. Experimental Results
  16. ...and 9 more sections

Figures (3)

  • Figure 1: Our proposed method, BMF-ICL, selects multilingual examples for ICL by considering three factors: semantic similarity, linguistic alignment, and language-specific performance.
  • Figure 2: An overview of how BMF-ICL computes semantic similarity, linguistic alignment, and language-specific performance scores to select multilingual examples.
  • Figure 3: Results of the ablation study. Semantic similarity, linguistic alignment, and language-specific performance are denoted as SS, LA, and LP, respectively.