Can you provide the inference？

[wulouzhu]

No description provided.

[snukky]

What do you mean exactly? There is a script correcting common test sets (https://github.com/grammatical/neural-naacl2018/blob/master/models/evaluate.sh), from which you can extract a command to run on any input, e.g.:

cat text.err \
    | bash preprocess.sh \
    | $MARIAN/marian-decoder -c model/config.yml -d $GPUS -w 4000 --quiet \
    | bash postprocess.sh \
> text.out

[wulouzhu]

@snukky
Should I tokenize error sentences first?

[snukky]

Yes, with the NLTK tokenizer.

[wulouzhu]

Thank you!