Revamp of the date pipeline

[keyber]

Description

This PR proposes a new eds.dates pipeline. It addresses the many limitations suffered by the current component.

The proposed pipeline drops the dateparser dependency. Albeit powerful, dateparser was not a perfect fit, and incurred a significant toll on the pipeline's overall complexity to work around the library's limitations.

The new date matcher handles:

• detection of absolute and relative dates, as well as durations
• direct parsing of dates, using RegEx group dicts
• normalisation using pendulum
• detection of periods (experimental)

Example

import spacy

nlp = spacy.blank('fr')
nlp.add_pipe('eds.dates', config=dict(detect_periods=True))

text = "le patient a été hospitalisé cette année pendant trois semaines"

doc = nlp(text)

for date in doc.spans['dates']:
    print(date, date._.date.dict(exclude_none=True))
# Out: cette année {'year': 0, 'direction': <Direction.CURRENT: 'CURRENT'>}
# Out: pendant trois semaines {'mode': <Mode.DURATION: 'DURATION'>, 'week': 3}

for date in doc.spans['dates']:
    print(date, date._.date.parse())
# Out: cette année 0 microseconds
# Out: pendant trois semaines 3 weeks

for p in doc.spans['periods']:
    print(p)
    print(p._.period)
# Out: cette année pendant trois semaines
# Out: {<Mode.FROM: 'FROM'>: cette année, <Mode.DURATION: 'DURATION'>: pendant trois semaines}

Label	Snippet	Parsed	Normalised
relative	cette année	{'year': 0, 'direction': <Direction.CURRENT: 'CURRENT'>}	0 microseconds
duration	pendant trois semaines	{'mode': <Mode.DURATION: 'DURATION'>, 'week': 3}	3 weeks
periode	cette année pendant trois semaines	{<Mode.FROM: 'FROM'>: cette année, <Mode.DURATION: 'DURATION'>: pendant trois semaines}

The experimental period detection algorithm uses pre detected dates.

Checklist

• If this PR is a bug fix, the bug is documented in the test suite.
• Changes were documented in the changelog (pending section).
• If necessary, changes were made to the documentation (eg new pipeline).

[keyber]

The test that fails is il y a 1 an pdt 1 mois
the test expects one extracted date but we extract two: "il y a un an" and "pdt 1 mois".
previously, this test passed because the relative_pattern had a bug.

I'm removing this single test

[bdura]

Drum roll please 🥁

[bdura]

@Thomzoy & @percevalw, I'm eager to have your opinion on this one.

[codecov-commenter]

Codecov Report

Merging #22 (da80b1c) into master (e3c66d9) will increase coverage by 0.34%.
The diff coverage is 99.13%.

@@            Coverage Diff             @@
##           master      #22      +/-   ##
==========================================
+ Coverage   95.88%   96.22%   +0.34%     
==========================================
  Files         131      139       +8     
  Lines        3107     3151      +44     
==========================================
+ Hits         2979     3032      +53     
+ Misses        128      119       -9     

Impacted Files	Coverage Δ
edsnlp/pipelines/misc/dates/patterns/current.py	100.00% <ø> (ø)
edsnlp/pipelines/misc/dates/dates.py	96.00% <96.15%> (+3.97%)	⬆️
...ines/misc/consultation_dates/consultation_dates.py	98.11% <100.00%> (ø)
edsnlp/pipelines/misc/dates/factory.py	100.00% <100.00%> (ø)
edsnlp/pipelines/misc/dates/models.py	100.00% <100.00%> (ø)
edsnlp/pipelines/misc/dates/patterns/__init__.py	100.00% <100.00%> (ø)
edsnlp/pipelines/misc/dates/patterns/absolute.py	100.00% <100.00%> (ø)
...dsnlp/pipelines/misc/dates/patterns/atomic/days.py	100.00% <100.00%> (ø)
...pipelines/misc/dates/patterns/atomic/delimiters.py	100.00% <100.00%> (ø)
...pipelines/misc/dates/patterns/atomic/directions.py	100.00% <100.00%> (ø)
... and 9 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update e3c66d9...da80b1c. Read the comment docs.

[Aremaki]

I did not review all files but what I've seen is terrific

[percevalw]

Great work !
Here's a benchmark of the new pipeline.

Benchmark on 10000 clinical docs:

Pipelines	Time
before	108s
this PR	309s

It is a bit slower but the new features are definitely worth it, and we can still optimise in the future.

[bdura]

Thanks for the benchmark! I am a bit disappointed by the results though :)

I just pushed a fix for the code block checks, I'll merge the changes as soon as the CI passes.

Thank you for your inputs !