mBART50

mBART50, short for multilingual BART 50, is a pretrained multilingual sequence-to-sequence model developed by Facebook AI Research for neural machine translation across 50 languages. It extends the mBART framework by using a single shared encoder–decoder and a common multilingual vocabulary to enable translation between any language pair in the set.

Architecture and training: The model follows a Transformer-based encoder-decoder architecture inspired by BART. It is trained with a denoising objective over multilingual text and, where available, parallel corpora, to learn cross-lingual representations. As a 50-language model, it supports many-to-many translation, including zero-shot translation between language pairs without direct training data.

Applications and use: mBART50 provides a strong foundation for multilingual translation and cross-lingual transfer in NLP tasks. It is widely used as a backbone for multilingual MT pipelines and for experiments requiring translation between non-English languages. The model is released publicly, for example on Hugging Face as facebook/mbart-large-50-many-to-many-mmt.

Limitations: Translation quality varies by language and data availability; lower-resource languages often lag behind higher-resource ones. The model can reflect biases in training data and benefits from fine-tuning on target language pairs. Large compute resources are typically required to train or fine-tune.

See also: mBART; multilingual BART; neural machine translation.