The edits for inserting multiple words are sometimes wrong about m2scorer

[tamhd]

We get incorrect evaluation using M2Scorer as well (please double check).

Hypothesis:

It 's Thursday , is n't it ?

M2 file:

S Thursday , is it not ?
A 0 0|||Mec|||It 's|||REQUIRED|||-NONE-|||0
A 3 5|||Mec|||n't it|||REQUIRED|||-NONE-|||0

Based on my understanding, the reason is the initialization step of the levenshtein matrix. It creates the first row using the index of the hypothesis, i.e: the necessary edits to transform an empty sentence into the hypothesis. I think the start and end of such edits should always be 0.

My proposed solution is to change line 827of file levenshtein.py into:

edit = ("ins", 0, 0, '', second[j-1], 0) # always insert at the beginning

It appears to me that the modification will fix the problem, yet I am not sure whether other errors may arise.

Tam

[kavehtp]

I tried it. The generated edits are correct with the change. Can you make a pull request?

[tamhd]

I tried it. The modification does not change the result of the 13 teams participating in CoNLL-2014 shared task (Table 7).

[kavehtp]

OK. Thanks.