fix: reassign affiliations from author superscript markers (0.9.10)

When a poster banner has a numbered affiliation list and authors carry
superscript markers (e.g. "Limaye 1,3"), the model often over-assigns every
institution to authors (typically the lead author absorbing all of them). A
new strictly-gated postprocess step parses the numbered affiliation list and
each author's markers from the raw text and reassigns each author exactly
their own numbered affiliations, which then resolve through ROR. No-op unless
a clean sequential 1..N list is found and every author marker resolves. On a
real 7-author poster the lead author went from 5 affiliations to his correct 2.

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).
 
+## [0.9.10] - 2026-06-09
+
+### Fixed
+
+- **Authors no longer over-assigned every affiliation.** On posters whose banner has a numbered affiliation list and whose authors carry superscript markers (e.g. "Limaye 1,3"), the fine-tuned model frequently assigned every institution to authors — most often the lead author absorbing all of them — even when the raw text was unambiguous. A new strictly-gated postprocess step (`_correct_affiliations_from_superscripts`) parses the numbered affiliation list and each author's markers from the raw text and reassigns each author exactly their own numbered affiliations (which then resolve through ROR as usual). It is a no-op unless a clean sequential `1..N` list is found, every author has a marker, and every marker resolves to a parsed affiliation — so posters without numbered affiliations are left untouched, and a `_validation` info note is emitted when it fires. On a real 7-author poster this corrected the lead author from 5 affiliations (two of them other authors' institutions) down to his correct 2.
+
 ## [0.9.9] - 2026-06-08
 
 ### Changed

poster2json/extract.py

@@ -1976,6 +1976,107 @@ def _is_section_label_only(text: str) -> bool:
     return bool(toks) and all(w in _SECTION_LABELS for w in toks)
 
 
+_AFFIL_LIST_SIGNATURE = re.compile(r";\s*\d{1,2}\s+[A-Z]")
+_AFFIL_SEGMENT_RE = re.compile(r"^(\d{1,2})\s+([A-Z].{4,})$")
+_AFFIL_FIRST_TAIL_RE = re.compile(r"(?<!\d)1\s+([A-Z].{4,})$")
+
+
+def _parse_numbered_affiliations(banner: str):
+    """Parse a poster banner's numbered affiliation list into ``{n: text}``.
+
+    Returns the map only when it is a clean, fully sequential ``1..N`` (N>=2)
+    list (affiliations are ``;``-delimited and digit-prefixed, with the first
+    entry trailing the author segment). Returns ``None`` otherwise, so callers
+    treat anything ambiguous as "do not touch".
+    """
+    parts = re.split(r"\s*;\s*", banner.strip())
+    if len(parts) < 2:
+        return None
+    amap = {}
+    for seg in parts[1:]:
+        m = _AFFIL_SEGMENT_RE.match(seg.strip())
+        if not m:
+            return None
+        amap[int(m.group(1))] = m.group(2).strip()
+    m1 = _AFFIL_FIRST_TAIL_RE.search(parts[0])
+    if m1:
+        amap[1] = m1.group(1).strip()
+    if len(amap) < 2:
+        return None
+    hi = max(amap)
+    if set(amap) != set(range(1, hi + 1)):
+        return None
+    return amap
+
+
+def _author_markers(banner: str, family: str):
+    """Return the superscript marker numbers that follow ``family`` in the
+    banner (e.g. "Limaye 1,3" -> [1, 3]), or ``None`` if no marker is found."""
+    m = re.search(re.escape(family) + r"\s{0,3}(\d{1,2}(?:\s*,\s*\d{1,2})*)", banner)
+    if not m:
+        return None
+    return [int(x) for x in re.split(r"\s*,\s*", m.group(1))]
+
+
+def _correct_affiliations_from_superscripts(result: dict, raw_text: str) -> None:
+    """Rebuild each author's affiliation list from superscript markers in the
+    raw text when the poster banner carries a numbered affiliation list.
+
+    The fine-tuned model commonly over-assigns affiliations (e.g. the lead
+    author absorbs every institution) even when the raw text is unambiguous.
+    The markers in the raw text are authoritative, so when present they are
+    used to reassign affiliations deterministically. Strictly gated: a no-op
+    unless a clean numbered list is found, every creator has a marker, and
+    every marker resolves to a parsed affiliation — so posters without this
+    structure are left untouched.
+    """
+    if not raw_text:
+        return
+    creators = result.get("creators")
+    if not isinstance(creators, list) or len(creators) < 2:
+        return
+    families = []
+    for c in creators:
+        if not isinstance(c, dict):
+            return
+        fam = c.get("familyName") or ""
+        if not fam:
+            nm = c.get("name", "")
+            fam = nm.split(",")[0].strip() if "," in nm else ""
+        if not fam:
+            return
+        families.append(fam)
+
+    banner = None
+    for line in raw_text.splitlines():
+        if families[0] in line and _AFFIL_LIST_SIGNATURE.search(line):
+            banner = line
+            break
+    if banner is None:
+        return
+
+    amap = _parse_numbered_affiliations(banner)
+    if not amap:
+        return
+
+    plan = []
+    for fam in families:
+        marks = _author_markers(banner, fam)
+        if not marks or any(n not in amap for n in marks):
+            return
+        plan.append(marks)
+
+    for c, marks in zip(creators, plan):
+        c["affiliation"] = [amap[n] for n in marks]
+    notes = result.setdefault("_validation", [])
+    if isinstance(notes, list):
+        notes.append({
+            "field": "creators",
+            "level": "info",
+            "message": "Affiliations reassigned from author superscript markers in the poster banner.",
+        })
+
+
 def _postprocess_json(
     data: dict, raw_text: str = "", extract_identifiers: bool = False
 ) -> dict:
@@ -2252,6 +2353,13 @@ def _postprocess_json(
             # honest than guessing.
             result["language"] = None
 
+    # Reassign affiliations from author superscript markers when the poster
+    # banner has a numbered affiliation list (the model commonly over-assigns,
+    # e.g. the lead author absorbing every institution). Strictly gated, so it
+    # is a no-op on posters without a numbered list. Runs before resolution so
+    # the reassigned name strings get ROR-resolved normally.
+    _correct_affiliations_from_superscripts(result, raw_text)
+
     # Affiliation normalization: coerce to the schema's array form (the model
     # sometimes emits a bare string or single object), drop any model-supplied
     # identifiers (ROR IDs are resolved from the name, never trusted from what

pyproject.toml

@@ -1,7 +1,7 @@
 [tool.poetry]
 
 name = "poster2json"
-version = "0.9.9"
+version = "0.9.10"
 description = "Convert scientific posters (PDF/images) to structured JSON metadata using Large Language Models"
 
 packages = [{ include = "poster2json" }]

tests/test_normalize.py

@@ -285,3 +285,55 @@ def test_postprocess_strips_surrogates_so_json_dumps_succeeds():
     data = {"titles": [{"title": "A\ud83dB"}], "subjects": [{"subject": "x\ud83d"}]}
     out = _postprocess_json(data, raw_text="")
     json.dumps(out, ensure_ascii=False).encode("utf-8")  # must not raise
+
+
+def test_superscript_corrector_reassigns_from_markers():
+    from poster2json.extract import _correct_affiliations_from_superscripts
+
+    raw = (
+        "Abridged Retinal Fluorescence Lifetimes in Glaucoma\n"
+        "Siddharth Limaye 1,3 Shahin Hallaj 1,2 , Maria Jessica Cruz 4 , "
+        "Saron Tedla 5 , Jalil Jalili 1,2 , Mark Christopher 1,2 , Linda M. Zangwill 1,2 "
+        "1 Hamilton Glaucoma Center, University of California, San Diego, CA, US; "
+        "2 Division of Ophthalmology Informatics, University of California, San Diego, CA, USA; "
+        "3 Carle Illinois College of Medicine, University of Illinois at Urbana-Champaign, IL, USA; "
+        "4 University of California, Davis, School of Medicine, CA, USA; "
+        "5 Oregon Health and Science University, School of Medicine, OR, USA\n"
+        "## BACKGROUND\n"
+    )
+    # The model over-assigned every institution to every author.
+    everything = ["Hamilton", "Division", "Carle", "Davis", "Oregon"]
+    result = {"creators": [
+        {"name": "Limaye, Siddharth", "familyName": "Limaye", "affiliation": list(everything)},
+        {"name": "Hallaj, Shahin", "familyName": "Hallaj", "affiliation": list(everything)},
+        {"name": "Cruz, Maria Jessica", "familyName": "Cruz", "affiliation": list(everything)},
+        {"name": "Tedla, Saron", "familyName": "Tedla", "affiliation": list(everything)},
+        {"name": "Jalili, Jalil", "familyName": "Jalili", "affiliation": list(everything)},
+        {"name": "Christopher, Mark", "familyName": "Christopher", "affiliation": list(everything)},
+        {"name": "Zangwill, Linda M.", "familyName": "Zangwill", "affiliation": list(everything)},
+    ]}
+    _correct_affiliations_from_superscripts(result, raw)
+    affs = [c["affiliation"] for c in result["creators"]]
+    # Limaye 1,3 -> Hamilton (UCSD) + Carle (UIUC), not all 5
+    assert len(affs[0]) == 2
+    assert affs[0][0].startswith("Hamilton") and "Urbana-Champaign" in affs[0][1]
+    # Cruz 4 -> only UC Davis; Tedla 5 -> only OHSU
+    assert len(affs[2]) == 1 and "Davis" in affs[2][0]
+    assert len(affs[3]) == 1 and "Oregon" in affs[3][0]
+    # 1,2 authors -> both UCSD departments
+    assert len(affs[1]) == 2
+    assert any(n.get("level") == "info" for n in result.get("_validation", []))
+
+
+def test_superscript_corrector_noop_without_numbered_list():
+    from poster2json.extract import _correct_affiliations_from_superscripts
+
+    raw = "Jane Doe, John Smith\nUniversity of Somewhere\n## INTRODUCTION\n"
+    result = {"creators": [
+        {"name": "Doe, Jane", "familyName": "Doe", "affiliation": ["University of Somewhere"]},
+        {"name": "Smith, John", "familyName": "Smith", "affiliation": ["University of Somewhere"]},
+    ]}
+    before = [list(c["affiliation"]) for c in result["creators"]]
+    _correct_affiliations_from_superscripts(result, raw)
+    assert [c["affiliation"] for c in result["creators"]] == before
+    assert "_validation" not in result