Skip to content

Incorrect BLEU score calculation #3

Description

@MorinoseiMorizo

Hi,

I've just found that the calculation of BLEU score during the training is incorrect.
Current implementation of evaluateBLEU function in train.cc,
just compare hypothesis and target word IDs, which are including unk,
and not deal with these tokens.

For example, if I set the source and target vocabulary size to 4,
I can get really high BLEU score because almost all of the words are unk.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions