Fix MedLink BM25 hard-negative mining - #1195
Open
fbonc wants to merge 2 commits into
Open
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem
get_bm25_hard_negativesscored the corpus against the positive document (bm25_model.get_scores(d)) instead of the query (q), which is bad. Hard negatives are supposed to be documents the query ranks highly but that aren't the answer.A second bug too. the negative selection excluded only the current positive (
ned_d_id != d_id). Once scoring is fixed to the query, all of a query's positives rank at the top, so excluding just one lets another true positive be chosen as a negative (a false negative) whenever a query has more than one match which record linkage allows.Fix
Score against the query, and exclude all of the query's positives (
neg_d_id not in d_ids), not just the current one.Tests
Added two regression tests (controlled fake BM25): one asserting the negative comes from scoring the query (not the positive doc), one asserting no positive is ever selected as a negative when a query has multiple positives.