Thank you for your excellent work!
In the current implementation of qa_evaluate, the has_intersection function used for comparing predicted answers with gold standard answers splits strings on whitespace and checks for exact matches between words.
This can lead to false negatives when comparing strings with minor differences in punctuation.
The current method may incorrectly evaluate answers as different when they are semantically the same but differ in punctuation. For example, "Yes," and "Yes" are currently treated as different answers.
I suggest implementing an improved has_intersection function with removing punctuation.
Thank you for your excellent work!
In the current implementation of
qa_evaluate, thehas_intersectionfunction used for comparing predicted answers with gold standard answers splits strings on whitespace and checks for exact matches between words.This can lead to false negatives when comparing strings with minor differences in punctuation.
The current method may incorrectly evaluate answers as different when they are semantically the same but differ in punctuation. For example, "Yes," and "Yes" are currently treated as different answers.
I suggest implementing an improved
has_intersectionfunction with removing punctuation.