filipg/geval
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

fix readme.md

Browse Source
This commit is contained in:
Filip Gralinski 2018-10-19 08:25:22 +02:00
parent 7a3a678034
commit c27a989b3f
1 changed files with 7 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
README.md
View File

@ -64,7 +64,7 @@ Let's step into the repo and run GEval (I assume you added `geval`
path to `$PATH`, so that you could just use `geval` instead of
`/full/path/to/geval`):
cd submission-01229
cd wmt-2017
geval
Well, something apparently went wrong:
@ -94,18 +94,6 @@ After a moment, you'll see the results:
WER 0.55201
Accuracy 0.01660
Ah, we forgot about the tokenization, in order to properly calculate
BLEU (or GLEU) the way it was done within the official WMT-2017
challenge, you need to tokenize the expected output and the actual
output of your system using the right tokenizer:
geval -t dev-0 --metric GLEU --metric WER --metric Accuracy --tokenizer 13a
BLEU 0.26901
WER 0.58858
GLEU 0.30514
Accuracy 0.01660
The results do not look good anyway and I'm not talking about
Accuracy, which, even for a good MT (or even a human), will be low (as
it measures how many translations are exactly the same as the golden
@ -213,6 +201,12 @@ and run GEval for one of the submissions (UEdin-NMT):
where `-i` stands for the input file, `-o` — output file, `-e` — file with expected (reference) data.
Note the tokenization, in order to properly calculate
BLEU (or GLEU) the way it was done within the official WMT-2017
challenge, you need to tokenize the expected output and the actual
output of your system using the right tokenizer. (The test set packaged
for Gonito.net challenge were already tokenized.)
Let's evaluate another system:
geval --metric BLEU --precision 4 --tokenizer 13a \