Fairseq

Robust Neural Machine Translation of User-Generated Content
🎓 PhD Thesis 🎓
Lydia Nishimwe
Robust Neural Machine Translation of User-Generated Content
Making Sentence Embeddings Robust to User-Generated Content
LREC-COLING 2024
Lydia Nishimwe, Benoît Sagot, Rachel Bawden
Making Sentence Embeddings Robust to User-Generated Content
Your Fairseq-trained model might have more embedding parameters than it should.
How a bug in reading SentencePiece vocabulary files causes some Fairseq-trained models to have up to 3k extra parameters in the embedding layer.
Lydia Nishimwe, posted on Mar 16, 2024
Last updated on Jan 8, 2026
Your Fairseq-trained model might have more embedding parameters than it should.
Fairseq Bug Fix
A bug in reading SentencePiece vocabulary files causes models to have 3k extra params in the embedding layer.
Fairseq Bug Fix
RoLASER
Making LASER sentence embeddings robust to user-generated content via Knowledge Distillation and Data Augmentation.
RoLASER