# Balancing Training for Multilingual Neural Machine Translation

**Authors:** Xinyi Wang, Yulia Tsvetkov, Graham Neubig  
**arXiv:** [2004.06748](https://arxiv.org/abs/2004.06748)  
**Date:** 2020-04-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2004.06748v4/

## Key points

- Authors Xinyi Wang, Yulia Tsvetkov, and Graham Neubig address imbalanced training sets in multilingual machine translation (MT) models.
- Imbalance in training data leads to performance discrepancies due to some languages having more data than others.
- Standard up-sampling methods can impact overall model performance.
- The authors propose a novel method involving automatically learning how to weight training data using a data scorer optimized for all test languages.
- The proposed method aims to maximize translation accuracy in one-to-many and many-to-one MT settings.
- Experiments show that the proposed approach consistently outperforms heuristic baselines in terms of average performance.
- The method offers flexible control over prioritizing languages for optimization based on specific requirements or priorities.
- This study presents an innovative strategy for addressing imbalanced training data in multilingual MT models, contributing to advancements in cross-lingual communication and translation technologies.

## Summary

In their paper titled "Balancing Training for Multilingual Neural Machine Translation," authors Xinyi Wang, Yulia Tsvetkov, and Graham Neubig address the challenge of imbalanced training sets in multilingual machine translation (MT) models. They highlight the issue where some languages have significantly more training data than others, leading to performance discrepancies. The standard approach of up-sampling less-resourced languages to improve representation can impact overall model performance. To tackle this problem, the authors propose a novel method that involves automatically learning how to weight training data using a data scorer optimized to enhance performance across all test languages. By optimizing the data weighting process, their method aims to maximize translation accuracy in both one-to-many and many-to-one MT settings. Through experiments conducted on two sets of languages, the authors demonstrate that their proposed approach consistently outperforms heuristic baselines in terms of average performance. One key advantage of their method is its ability to offer flexible control over which languages are prioritized for optimization. This flexibility allows researchers and practitioners to tailor the model's performance based on specific language requirements or priorities. Overall, the study sheds light on an innovative strategy for addressing imbalanced training data in multilingual MT models, ultimately contributing to advancements in cross-lingual communication and translation technologies.