General MT: definition of constrained track; update on Japanese

[Tom Kocmi]

Hi All,

 

Based on your feedback from the poll, we have finalized the list of pretrained models that are allowed in constrained track.

 

The limitations for the constrained track are as follows:

* You may only use the training data allowed for this year (specified later on this page)

* You may use any publicly available metric that was evaluated on past WMT Metrics shared tasks (for example: COMET, Bleurt, etc.)

* You may ONLY use the following listed pretrained models in all publicly available model sizes: mBART, BERT, RoBERTa, sBERT, LaBSE

* Any basic linguistics tools (taggers, parsers, morphology analyzers, etc.)

 

All other models are allowed in unconstrained track.

 

Secondly, we will evaluate Japanese-English in both directions.

 

Have a lovely day,

Tom

(in Germany, he/him)