Skip to content

Multilingual Models

Multilingual neural models, particularly massively multilingual systems (M2M family), train a single model on hundreds of language pairs simultaneously. These systems leverage shared representations to translate between any pair of supported languages without pivot languages, and enable zero-shot transfer to low-resource pairs unseen during training. The scale of multilinguality enables practical deployment across the world's languages, but introduces new challenges: hallucinations in low-resource pairs, representation collapse, and bias amplification across diverse linguistic phenomena.

Massively multilingual systems have become the standard for practical machine translation deployment in industry, but their behavior across diverse language pairs—especially low-resource and non-English-centric pairs—remains understudied compared to bilingual baselines.

Key papers

  • [[2023-guerreiro-hallucinations-multilingual]] — comprehensive empirical analysis of hallucinations in M2M multilingual translation models across 100+ language pairs