Inserting target-language register tokens and restricting target attention to those registers reduces off-target errors and lets a compact decoder-only MNMT model compete with far larger models.
In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Process- ing, pages 21142–21157, Miami, Florida, USA
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Registering Source Tokens to Target Language Spaces in Multilingual Neural Machine Translation
Inserting target-language register tokens and restricting target attention to those registers reduces off-target errors and lets a compact decoder-only MNMT model compete with far larger models.