A Comparison of Methods to Bias Translation Toward Portuguese Variants

2026-04-12 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

This study compares four methods for biasing Machine Translation (MT) systems toward European Portuguese (EP) when translating into Portuguese, addressing the resource imbalance that typically favors Brazilian Portuguese (BP). The methods evaluated target different stages of the MT lifecycle: reranking n-best MT outputs using a variant classifier, biasing hypothesis generation during inference, fine-tuning existing models specifically for EP, and employing a Large Language Model (LLM)-based approach. Researchers found that all methods successfully biased translation outputs to some degree. The LLM-based approach achieved the numerically highest results, though the influence of memorization on its performance requires further investigation.

Key takeaway

For research scientists developing MT systems for multilingual contexts, you should investigate methods to bias translation outputs towards specific language variants, especially when dealing with resource-imbalanced languages like Portuguese. Experiment with reranking, inference-time biasing, and fine-tuning, but also explore LLM-based approaches while carefully assessing potential memorization effects to ensure robust variant control.

Key insights

Biasing MT towards minority language variants is achievable through various lifecycle interventions.

Principles

• Resource imbalance favors dominant language variants.
• Multiple MT lifecycle stages allow for variant biasing.

Method

Methods include reranking n-best outputs, biasing inference generation, fine-tuning, and using LLM-based translation to favor a target language variant.

In practice

• Rerank MT outputs with a variant classifier.
• Fine-tune models for specific language variants.
• Consider LLM-based translation for variant control.

Topics

• Machine Translation
• Portuguese Variants
• European Portuguese Bias
• LLM-based Translation
• N-best Reranking

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Gender bias in machine translation: an evaluation in the English-Portuguese language pair — Machine translation systems exhibit gender bias, favoring masculine and stereotyped occupational translations from English to Portuguese.
• AMALIA: A Fully Open Large Language Model for European Portuguese — Targeted training and native benchmarking are crucial for underrepresented language variants like European Portuguese.
• Rethinking Cross-lingual Gaps from a Statistical Viewpoint — LLM cross-lingual gaps stem from response variance, not knowledge barriers, and are reducible via variance control.
• mmPISA-bench: Do LLMs Reason Equally Well Across 43 Languages? — LLMs reason effectively across 43 languages, with machine translation not degrading accuracy.
• Extending Item Response Theory for Efficient and Meaningful Multilingual Evaluation — Multilingual-IRT extends Item Response Theory to efficiently evaluate LLMs across languages, identifying translation errors and cultural bias.
• Assessing Y-Axis Influence: Bias in Multimodal Language Models on Chart-to-Table Translation — Y-axis biases in chart datasets significantly impact Multimodal Language Model performance in chart-to-table translation.

Open in AIssential →

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