Refactor no answer handling in Question Answering #258

Timoeller · 2020-02-20T16:08:34Z

This PR will need adjustments from downstream applications consuming QA inference, since we now return None for no answer (instead of empty string)

Change naming to no answer boost.
Change how no answer score is computed.
Predictions contain a no answer along positive answers, the no answer score is relative to positive answers.

tholor

Looks good. Minor comments

examples/question_answering.py

farm/modeling/prediction_head.py

Timoeller added 4 commits February 20, 2020 14:36

First try

bfd1cbb

Adjust doc for inference debugging

1c978fc

Adjust var naming

98e016a

Adjust naming and concepts of no answer boost

cfb0c60

Timoeller requested a review from tholor February 20, 2020 16:08

Change no_ans_boost

2f28dad

tholor requested changes Feb 20, 2020

View reviewed changes

examples/question_answering.py Outdated Show resolved Hide resolved

farm/modeling/prediction_head.py Outdated Show resolved Hide resolved

Timoeller added 5 commits February 20, 2020 18:15

Change no answer output

e8f3391

Revert dataset names

a106d41

Adjust docstring

c7c9913

Adjust comments

8f627c0

Adjust comments

f0ae123

Timoeller mentioned this pull request Feb 21, 2020

Add no answer handling deepset-ai/haystack#26

Merged

Timoeller requested review from tholor and removed request for tholor February 24, 2020 09:24

Timoeller self-assigned this Feb 24, 2020

Timoeller merged commit 875f8c6 into master Feb 24, 2020

Timoeller deleted the no_ans_boost branch February 24, 2020 09:45

Provide feedback