grobidOrg / grobid / 29460061878

43%

push

github

Improvement on affiliation and authors extraction - algorithm/reconciliation (#1467)

Extract the author-to-affiliation linking logic out of HeaderParser into a
dedicated, unit-tested AuthorAffiliationAssigner, and make the assignment
marker-driven with well-defined fallbacks.
Main changes:
* New AuthorAffiliationAssigner implementing a priority-based strategy:
  1. Distribution for trivial single-author / single-affiliation cases
  2. Direct marker matching (Person.getMarkers vs Affiliation.getMarker)
  3. String-search marker matching against the raw author string, as a
     fallback when the name model misses markers
  4. Proximity matching via layout-token coordinate distance
  5. Sequential assignment as a last resort when no coordinates are available
  Markerless authors are left unassigned when explicit markers are present
  on their peers, instead of being blanket-assigned. Content-duplicate
  affiliations are collapsed (substring containment above an 0.8 length ratio).
* HeaderParser slimmed down to delegate linking to the new class; author
  and affiliation handling reworked accordingly.
* AuthorParser and AffiliationAddressParser: improved marker extraction and
  author de-duplication ahead of affiliation assignment.
* Person: helpers for markers and de-duplication.
* Add a linking-aware "affiliation_linked" metric to the end-to-end
  evaluation, pairing each author with its gold counterpart and comparing
  attached affiliations. Single-affiliation gold papers are completed by
  linking every author to the sole affiliation.
* Extensive test coverage: AuthorAffiliationAssignerTest,
  PersonDeduplicationTest, HeaderParserTest, AuthorParserTest,
  AffiliationAddressParserTest, plus new evaluation fixtures.
Validated on a 6000-PDF bioRxiv set (CRF, no consolidation): no crashes or
regressions attributable to the linking logic; markerless authors correctly
left unassigned, matching gold. Known follow-ups left for separate PRs:
cross-listing auth... (continued)

9228 of 25664 branches covered (35.96%)

Branch coverage included in aggregate %.

19677 of 46526 relevant lines covered (42.29%)

1.71 hits per line