feat: stop model-extracting the conference object entirely (0.9.4)

0.9.3 dropped conference date fields; this drops the rest (conferenceName,
conferenceLocation, conferenceUri, conferenceAcronym) so poster2json never
emits a conference object. Conference info comes from the repository or
platform, not the poster. Prompt no longer requests it; the placeholder and
grounding postprocess logic is removed and any conference the model emits is
stripped.

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).
 
+## [0.9.4] - 2026-06-08
+
+### Changed
+
+- **The `conference` object is no longer model-extracted.** 0.9.3 dropped the conference date fields; this release drops the rest (`conferenceName`, `conferenceLocation`, `conferenceUri`, `conferenceAcronym`) so poster2json never emits a `conference` object at all. Conference information is supplied by the repository or entered on the platform, not guessed from the poster. The prompt no longer requests it, the placeholder/grounding postprocess logic is removed, and any `conference` the model emits is stripped.
+
 ## [0.9.3] - 2026-06-05
 
 ### Changed

poster2json/extract.py

@@ -1417,7 +1417,7 @@ def _generate(model, tokenizer, prompt: str, max_tokens: int) -> str:
 EXTRACTION_PROMPT = """Convert this scientific poster text to JSON format.
 
 CRITICAL RULES:
-1. Extract ALL required fields: creators, titles, publicationYear, subjects, descriptions, conference
+1. Extract ALL required fields: creators, titles, publicationYear, subjects, descriptions
 2. Create SEPARATE sections for EACH distinct topic/header found in the poster
 3. Use the poster's OWN section headers exactly as they appear. Lines prefixed with "## " indicate detected headers from the poster layout. Standard headers (Abstract, Introduction, Methods, Results, Discussion, Conclusions, References, Acknowledgements) are common examples, but always prefer the poster's actual headers over generic ones.
 4. Each section must have its OWN "sectionTitle" and "sectionContent"
@@ -1435,7 +1435,6 @@ def _generate(model, tokenizer, prompt: str, max_tokens: int) -> str:
   "publicationYear": null,
   "subjects": [{{"subject": "keyword1"}}, {{"subject": "keyword2"}}, {{"subject": "keyword3"}}],
   "descriptions": [{{"description": "A 3-4 sentence summary of the full poster...", "descriptionType": "Abstract"}}],
-  "conference": null,
   "researchField": null,
   "content": {{
     "sections": [
@@ -1452,32 +1451,28 @@ def _generate(model, tokenizer, prompt: str, max_tokens: int) -> str:
 - publicationYear: Extract if a year is printed on the poster. If not found, set to null.
 - subjects: Extract 3-5 keywords from poster content
 - descriptions: Write a 3-4 sentence summary of the full poster.
-- titles: If the poster title is ALL CAPS, convert to proper Title Case preserving acronyms (e.g. "RESEARCH ON SARS-CoV-2" not "RESEARCH ON SARS-COV-2")
-- conference: Extract from text visible on the poster (header, footer, logos). If not found, set to null.
-- imageCaptions/tableCaptions: Include captions for figures/tables on the poster. If none exist, use [].
+- titles: If the poster title is ALL CAPS, convert to proper Title Case preserving acronyms (e.g. "RESEARCH ON SARS-CoV-2" not "RESEARCH ON SARS-COV-2")- imageCaptions/tableCaptions: Include captions for figures/tables on the poster. If none exist, use [].
 - researchField: MUST be exactly one of: "Health Sciences" | "Life Sciences" | "Physical Sciences" | "Social Sciences" — or null if unclear.
-- GROUNDING: For metadata fields (conference, publicationYear), only extract values that appear as text on the poster. If not found, use null. For section content, copy ALL text verbatim — do not skip or shorten.
+- GROUNDING: For metadata fields (publicationYear), only extract values that appear as text on the poster. If not found, use null. For section content, copy ALL text verbatim — do not skip or shorten.
 
 POSTER TEXT TO CONVERT:
 {raw_text}
 
 OUTPUT VALID JSON ONLY:"""
 
 FALLBACK_PROMPT = """Convert poster text to JSON. REQUIRED FIELDS:
-1. creators, titles, publicationYear, subjects, descriptions, conference, content
+1. creators, titles, publicationYear, subjects, descriptions, content
 2. SEPARATE section for EACH header found in the poster text. Use the poster's own headers. Lines starting with "## " are detected headers.
 3. Copy ALL text EXACTLY verbatim — every line of poster text must appear in a section
 4. If title is ALL CAPS, convert to Title Case preserving acronyms (SARS-CoV-2, not SARS-COV-2)
-5. conference: extract if visible on the poster. If not found, set to null.
-6. imageCaptions/tableCaptions: for figures/tables on the poster. If none, use [].
+5. imageCaptions/tableCaptions: for figures/tables on the poster. If none, use [].
 
 {{
   "creators": [{{"name": "LastName, FirstName", "givenName": "FirstName", "familyName": "LastName", "affiliation": ["Institution"]}}],
   "titles": [{{"title": "Poster Title"}}],
   "publicationYear": null,
   "subjects": [{{"subject": "keyword1"}}, {{"subject": "keyword2"}}],
   "descriptions": [{{"description": "3-4 sentence summary of the full poster", "descriptionType": "Abstract"}}],
-  "conference": null,
   "researchField": null,
   "content": {{
     "sections": [{{"sectionTitle": "Header", "sectionContent": "Full verbatim text of this section..."}}]
@@ -2026,39 +2021,9 @@ def _postprocess_json(
         elif isinstance(result[key], (dict, list)):
             result[key] = _normalize_captions(result[key], caption_type=ctype)
 
-    # Strip placeholder/hallucinated conference values
-    conf = result.get("conference")
-    if isinstance(conf, dict):
-        # Conference dates are not extracted by the model; they come from the
-        # repository or are entered on the platform.
-        for _k in ("conferenceStartDate", "conferenceEndDate", "conferenceYear"):
-            conf.pop(_k, None)
-        cn = conf.get("conferenceName", "")
-        if isinstance(cn, str) and _is_placeholder(cn):
-            conf.pop("conferenceName", None)
-        # Ground-truth check: conference name must appear in the poster text
-        cn = conf.get("conferenceName", "")
-        if isinstance(cn, str) and cn and raw_text:
-            cn_lower = cn.lower().strip()
-            if cn_lower not in raw_text.lower():
-                result["conference"] = None
-                conf = None
-        if conf is not None:
-            cl = conf.get("conferenceLocation", "")
-            if isinstance(cl, str) and _is_placeholder(cl):
-                conf.pop("conferenceLocation", None)
-            cu = conf.get("conferenceUrl", "")
-            if isinstance(cu, str) and _is_placeholder(cu):
-                conf.pop("conferenceUrl", None)
-            # Fix 4: validate conferenceUri is a real URL
-            for uri_key in ("conferenceUri", "conferenceUrl"):
-                val = conf.get(uri_key, "")
-                if isinstance(val, str) and val and not val.startswith(("http://", "https://")):
-                    conf.pop(uri_key, None)
-            # Collapse to null if no meaningful fields remain
-            meaningful = {k for k, v in conf.items() if v}
-            if not meaningful:
-                result["conference"] = None
+    # Conference is not extracted by the model at all; it is supplied by the
+    # repository or entered on the platform. Drop anything the model emits.
+    result.pop("conference", None)
 
     # Publisher is auto-set by posters.science — strip if the LLM emitted one
     result.pop("publisher", None)

pyproject.toml

@@ -1,7 +1,7 @@
 [tool.poetry]
 
 name = "poster2json"
-version = "0.9.3"
+version = "0.9.4"
 description = "Convert scientific posters (PDF/images) to structured JSON metadata using Large Language Models"
 
 packages = [{ include = "poster2json" }]

tests/test_normalize.py

@@ -193,17 +193,16 @@ def test_conference_placeholder_stripped():
         }
     }
     out = _postprocess_json(data, raw_text="")
-    # conferenceName is a placeholder and conferenceYear is dropped (dates are
-    # not model-extracted), so the object collapses to null.
-    assert out["conference"] is None
+    # conference is no longer model-extracted at all; it is dropped entirely.
+    assert "conference" not in out
 
 
 def test_conference_all_placeholder_collapses_to_null():
     from poster2json.extract import _postprocess_json
 
     data = {"conference": {"conferenceName": "Name of Conference", "conferenceLocation": "City, Country"}}
     out = _postprocess_json(data, raw_text="")
-    assert out["conference"] is None
+    assert "conference" not in out
 
 
 def test_bogus_table_caption_filtered():
@@ -230,25 +229,27 @@ def test_publisher_stripped_from_output():
     assert "publisher" not in out
 
 
-def test_real_conference_preserved():
+def test_conference_dropped_even_when_real():
     from poster2json.extract import _postprocess_json
 
     data = {"conference": {"conferenceName": "US-RSE'25", "conferenceYear": 2025}}
     out = _postprocess_json(data, raw_text="")
-    assert out["conference"]["conferenceName"] == "US-RSE'25"
+    # Even a real, well-formed conference is dropped — it comes from metadata.
+    assert "conference" not in out
 
 
-def test_conference_grounded_in_poster_text():
+def test_conference_dropped_even_when_grounded():
     from poster2json.extract import _postprocess_json
 
     data = {"conference": {"conferenceName": "ICML 2025", "conferenceYear": 2025}}
     out = _postprocess_json(data, raw_text="Presented at ICML 2025, Vancouver")
-    assert out["conference"]["conferenceName"] == "ICML 2025"
+    # No grounding logic remains; conference is dropped regardless of the text.
+    assert "conference" not in out
 
 
 def test_conference_not_in_poster_text_stripped():
     from poster2json.extract import _postprocess_json
 
     data = {"conference": {"conferenceName": "US-RSE'25", "conferenceYear": 2025}}
     out = _postprocess_json(data, raw_text="Drug-polymer interactions in nanoparticle delivery systems")
-    assert out["conference"] is None
+    assert "conference" not in out