Stop superscript affiliation corrector from grabbing poster body text (0.9.18)

The numbered affiliation list's last entry was bounded only by the end of
the scanned banner region, so when the affiliation line and the abstract
shared a line (no detected header between them) the last affiliation
swallowed the body text (up to ~2000 chars of abstract/section headers on
~108 of 6,725 posters). Block parsing now bails, keeping the corrector a
strict no-op, when any parsed affiliation looks like body text: >180 chars,
or a run of 5+ consecutive lowercase words (prose) or ALL-CAPS words
(section headers). Clean numbered banners are unaffected.

Author: jimnoneill

CHANGELOG.md

@@ -5,6 +5,12 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.9.18] - 2026-06-10
+
+### Fixed
+
+- **The superscript affiliation corrector no longer grabs poster body text as an affiliation.** The numbered affiliation list's last entry was bounded only by the end of the scanned banner region, so when the affiliation line and the abstract shared a line (no detected header between them) the last affiliation swallowed the body text (observed up to 2,119 characters of abstract and section headers across ~108 of 6,725 posters). Block parsing now bails, keeping the corrector a strict no-op, when any parsed affiliation looks like it ran into body text: longer than 180 characters, or containing a run of five or more consecutive lowercase words (prose) or ALL-CAPS words (section headers). Clean numbered banners and the existing correction behavior are unaffected.
+
 ## [0.9.17] - 2026-06-10
 
 ### Changed

poster2json/extract.py

@@ -2002,6 +2002,28 @@ def _is_section_label_only(text: str) -> bool:
 # institution name (optionally space/dot separated, including accented capitals)
 _AFFIL_MARK = re.compile(r"(?:(?<=[\s;,(])|^)(\d{1,2})[.\s]{0,2}(?=[A-ZÀ-ɏ(])")
 
+# A parsed affiliation segment that ran past the institution list into the
+# poster body/abstract. The banner region can include body text when the
+# affiliation line and the abstract are not separated by a detected header
+# (e.g. they share a line), and the last numbered entry is delimited only by
+# the end of the region — so it swallows whatever follows. Real affiliations
+# are Title-Case institution/place tokens, not prose: they are not very long
+# and carry no long run of consecutive lowercase or ALL-CAPS words. Such a
+# segment is a strong ambiguity signal, so the corrector bails (no-op).
+_AFFIL_MAX_LEN = 180
+_PROSE_RUN = re.compile(r"\b[a-zà-ÿ]{2,}(?:\s+[a-zà-ÿ]{2,}){4,}\b")
+_CAPS_RUN = re.compile(r"\b[A-Z]{2,}(?:\s+[A-Z]{2,}){4,}\b")
+
+
+def _affiliation_ran_into_body(txt: str) -> bool:
+    """True if a parsed affiliation segment looks like it ran into poster body
+    text (too long, or a sentence-/header-like run of words)."""
+    return (
+        len(txt) > _AFFIL_MAX_LEN
+        or bool(_PROSE_RUN.search(txt))
+        or bool(_CAPS_RUN.search(txt))
+    )
+
 
 def _parse_marker_run(run: str):
     """Parse an author marker run ("1,3" / "1-3,6" / "4 2") into a list of ints,
@@ -2070,7 +2092,7 @@ def _parse_affiliation_block(region: str):
     for j, (num, ms, me) in enumerate(seq):
         end = seq[j + 1][1] if j + 1 < len(seq) else len(region)
         txt = region[me:end].strip().strip(",;").strip()
-        if not txt:
+        if not txt or _affiliation_ran_into_body(txt):
             return None
         amap[num] = txt
     return amap, seq[0][1]

pyproject.toml

@@ -1,7 +1,7 @@
 [tool.poetry]
 
 name = "poster2json"
-version = "0.9.17"
+version = "0.9.18"
 description = "Convert scientific posters (PDF/images) to structured JSON metadata using Large Language Models"
 
 packages = [{ include = "poster2json" }]

tests/test_normalize.py

@@ -416,6 +416,56 @@ def test_superscript_corrector_expands_ranges():
     assert len(affs[1]) == 1 and "Beta" in affs[1][0]        # Jones 2 -> Beta
 
 
+def test_superscript_corrector_noop_when_block_runs_into_abstract():
+    from poster2json.extract import _correct_affiliations_from_superscripts
+
+    # The last numbered affiliation is followed immediately by the abstract on
+    # the same banner run (no header break), so the naive parse would swallow
+    # the body text. The corrector must detect that and stay a no-op.
+    raw = (
+        "Early reproductive failure in kakapo\n"
+        "Olivia Janes 1, Jana Wold 1,2, Tammy Steeves 2\n"
+        "1 Centre National de la Recherche Scientifique (CNRS), Rennes, France; "
+        "2 University of Canterbury, New Zealand "
+        "Early reproductive failure in kakapo Kakapo are a Nationally Critical "
+        "taonga species endemic to Aotearoa New Zealand and experience high rates "
+        "of early embryo death across the breeding population studied here.\n"
+    )
+    original = ["School of Biological Sciences"]
+    result = {"creators": [
+        {"name": "Janes, Olivia", "familyName": "Janes", "affiliation": list(original)},
+        {"name": "Wold, Jana", "familyName": "Wold", "affiliation": list(original)},
+        {"name": "Steeves, Tammy", "familyName": "Steeves", "affiliation": list(original)},
+    ]}
+    _correct_affiliations_from_superscripts(result, raw)
+    # Untouched: no reassignment, no abstract text grabbed as an affiliation.
+    for c in result["creators"]:
+        assert c["affiliation"] == original
+    assert not result.get("_validation")
+
+
+def test_affiliation_ran_into_body_detector():
+    from poster2json.extract import _affiliation_ran_into_body
+
+    assert not _affiliation_ran_into_body("University of California, San Diego, CA, US")
+    assert not _affiliation_ran_into_body(
+        "LPL, Laboratoire de Physique des Lasers, Universite Paris 13, "
+        "Sorbonne Paris Cite, Villetaneuse, France"
+    )
+    # prose / abstract text
+    assert _affiliation_ran_into_body(
+        "Kakapo are a Nationally Critical taonga species endemic to Aotearoa "
+        "New Zealand and experience high rates of early embryo death"
+    )
+    # all-caps section headers grabbed from the body
+    assert _affiliation_ran_into_body(
+        "Univ. of California San Diego USA MOTIVATION THE GAS PUFF Z PINCH "
+        "THE MACHINE LEARN IMAGE RESULTS"
+    )
+    # absurdly long single segment
+    assert _affiliation_ran_into_body("Institute of " + "x" * 200)
+
+
 def test_creator_nametype_personal_vs_organizational():
     from poster2json.extract import _postprocess_json