SequenceLib is slower than python-Levenshtien about errant HOT 7 CLOSED

Aha nice.

I was actually originally using the Damerau-Levenshtein code bundled with ERRANT (rdlextra.py) to do the character alignment, but found SequenceMatcher to be a lot faster. If python-Levenshtein is faster still, then that might definitely be something worth upgrading.

As for the rationale behind the custom substitution cost, it's easiest to refer you to the alignment paper, particularly section 3.2 and Table 2.

from errant.

Thanks for the reference. We noticed the speed improvements you had done that time (thanks).

BTW I have been cleaning up the code to make it easily usable as a library, making it pip installable with python 3 type annotations etc. If you plan to make this a library published to pypi, will give a pull request. Or I would consider publishing as a separate library.

One another side note, want to understand whether homophones can be added as another static rule for subsitutions, as they are among a easy class of errors people make ...

P.S. I think I we can consider moving the Damerau Levenshtein code to cython for speed improvements. But thats a separate task in itself. Realized how slow python is compared to cython in python-Levenshtein.

from errant.

@chrisjbryant Check out this refactor, https://github.com/sai-prasanna/errant. Its a bit huge, but just a basic cleanup, adding types and making it pip installable without altering the logic etc. Yet to test the commands though.

from errant.

Looks like you've been busy! Would be good to make it pip installable, but I've never done that before so have no experience with it.
It would be good to know whether it produces the same results and whether it's faster when you get a chance to test it. I guess it'll also need a new readme on how to use!

from errant.

Pip Installation:

It is pip installable from the repository (not published to pypi, will do that soon). If we conclude on what else to change we can upload to pypi under name errant(guess it would be available).

Reproduciblity:

I ran the parallel_to_m2 and compare_m2 commands on CoNLL 14 dataset sentences. Got the same results as existing repository for spacy 1.9.

But I have set spacy dependency in setup.py to 2.0 (they use the same flags) .There are some results have some diff in spacy 2.0. It would be very helpful if you could check whether the diff with spacy 2.0 is not degrading the quality.

Using 1.x is not a option for me atleast right now, and it seems spacy 2.1 will be faster when it releases (nightly is out). But if you insist on keeping 1.x we can try something like pip install errant[spacy1x] to support spacy 1.9.

README
The README is updated to include information about pip installation and how to run the cli. I will add example on programmatic access of Errant class.

from errant.

Aha ok.
From memory, I think spacy 2.0 should technically be slightly more accurate than 1.9 (at least in terms of POS tagging and parsing) but I mainly didn't upgrade because it was slower.

Since spacy 2.1 is supposed to be faster, I was intending to revisit the code when that was released. They also finally fixed the coarse POS mapping in 2.1 apparently, so we'll be able to remove the manually defined tagmap and just use the native spacy token.pos_ property instead I think.

I would insist on keeping the 1.9 version for now as that's also the version we'll be using in the BEA2019 shared task. We can still update ERRANT in the future, but it'd be good to keep the 1.9 version available in case people want to evaluate using the official shared task evaluation procedure.

from errant.

Closing this, as it has been addressed in my pull request.

from errant.