add handle_relations_with_same_arguments parameter to RETextClassificationWithIndicesTaskModule

[RainbowRivey]

If there are multiple relations with same pair of arguments, handle_relations_with_same_arguments defines if we remove both of them (keep_none) or keep the first one (keep_first). Full duplicates (if relation label is also the same) are not affected by this, one relation will be kept and a warning shown.

Note that if collect_statistics=True, final statistics do not include such "full duplicates" either as 'available' nor as 'skipped'. Also, warnings about elements collected in the statistics, e.g. skipped relations with same arguments, will not be shown if collect_statistics=True.

This PR changes the default behavior to keep_none, that's why is labeled as breaking (previously it was implicitly keep_first). Doing so prevents the model from learning conflicting relations and gaining biases for arbitrary 'first' relation occurred.

TODO:

• Test new behaviour on a downstream task (e.g. runs with 3 seeds at drugprot)

Note: keep_none reduces train instances count of drugprot by 199 (1,16%) from 17058 to 16859.

metric	baseline	with keep_none
quality (micro f1)	0.7737 (±0.0042)	0.7653 (±0.0022)
Inference time	13.5 (±0.3)	11.2 (±0.5)

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 95.53%. Comparing base (7ad87d8) to head (aba5a1e).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #155      +/-   ##
==========================================
+ Coverage   95.51%   95.53%   +0.02%     
==========================================
  Files          61       61              
  Lines        5212     5237      +25     
==========================================
+ Hits         4978     5003      +25     
  Misses        234      234              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[ArneBinder]

thanks, some comments below

EDIT: please also add some documentation about the new parameter to the docstring

[ArneBinder]

can you add a simple test for the not yet test-covered line? (from codecov report)

[ArneBinder]

feedback below