Do two direction data for parallel data is necessary?

microsoft / MASS

MASS: Masked Sequence to Sequence Pre-training for Language Generation

https://arxiv.org/pdf/1905.02450.pdf

Other

1.11k stars 206 forks source link

Do two direction data for parallel data is necessary? #165

Open SefaZeng opened 3 years ago

SefaZeng commented 3 years ago

Following the procedure of generate data, I only have train.A-B.A and train.A-B.B for parallel data. But when I run the pretrain script, it raise error about there is no para data for train.B-A. Is there some args I did not make correctly?