Multilingual Pixel Representations for Translation and Effective Cross-lingual Transfer

Pretrained multilingual translation models using either pixel or subword (bpe) representations trained on the many-to-one parallel TED-59 dataset, accompanying the EMNLP'23 paper "Multilingual Pixel Representations for Translation and Effective Cross-lingual Transfer." 

Models can be interacted with on the command line or through a script similarly to other fairseq models, but require our code extension for rendered text with pixel representations. Each model zip file contains: the fairseq model checkpoint, vocab files, language list file, and relevant sentencepiece model(s).

We additionally package the TED-59 data here in raw extracted format for ease of comparison (original dataset release and paper by Qi et. al 2018). 

For more information, see our:

• Paper describing the method and training data  [arXiv]
• Code repository with scripts  [github]