[ENH] Keywords - replace embedding with MBERT #932
Conversation
Codecov Report
Additional details and impacted files@@ Coverage Diff @@
## master #932 +/- ##
==========================================
+ Coverage 77.54% 77.63% +0.08%
==========================================
Files 86 86
Lines 12280 12266 -14
Branches 1609 1602 -7
==========================================
Hits 9523 9523
+ Misses 2458 2446 -12
+ Partials 299 297 -2 |
| from .decorators import * | ||
| from .widgets import * | ||
| from .concurrent import asynchronous | ||
|
|
||
|
|
||
| def enum2int(enum: Union[Enum, IntEnum]) -> int: |
There was a problem hiding this comment.
Why isn't this imported from Orange.widgets.utils?
There was a problem hiding this comment.
This one is actually already merged. I forgot to rebase this PR. :) Anyway enum2int from Orange is not released yet. I will add a test that will start to fail after some time and warn about removing it from text after some time.
There was a problem hiding this comment.
Actually it was written here before it existed in Orange. I propose a time bomb that will start to fail and warn about removing it. #951
| return b64encode(zlib.compress(instance.encode("utf-8", "replace"), level=9)) | ||
|
|
||
|
|
||
| if __name__ == "__main__": |
There was a problem hiding this comment.
When I run this, it crashes.
|
Event with the two comments I'm happy to merge. |
PrimozGodec
force-pushed
the
keywords-mbert
branch
from
3859ad1 to
839444a
Compare
|
@VesnaT since you approved it I am merging it. |
Issue
Embedding keyword has poor performance and is slow.
Description of changes
Replace Embedding keyword extractor with MBERT with excellent performance. This keyword extractor extracts keywords with higher F1 scores than all existing extractors. The downside of this extractor is that it, on average extracts less than 5 keywords per document and also 0 many times.
Includes