Multilingual Pixel Representations for Translation and Effective Cross-lingual Transfer
This work addresses the challenge of building extensible multilingual models for all languages and scripts, with incremental improvements in data efficiency and transfer capabilities.
The paper tackles the problem of training multilingual machine translation models by using pixel representations instead of subword embeddings, resulting in improved performance, seamless cross-lingual transfer to unseen scripts, and greater data efficiency compared to alternatives like vocabulary expansion.
We introduce and demonstrate how to effectively train multilingual machine translation models with pixel representations. We experiment with two different data settings with a variety of language and script coverage, demonstrating improved performance compared to subword embeddings. We explore various properties of pixel representations such as parameter sharing within and across scripts to better understand where they lead to positive transfer. We observe that these properties not only enable seamless cross-lingual transfer to unseen scripts, but make pixel representations more data-efficient than alternatives such as vocabulary expansion. We hope this work contributes to more extensible multilingual models for all languages and scripts.