fix: tag LLM summary descriptionType as Other (not Abstract); drop from prompt

CHANGELOG.md

@@ -5,6 +5,16 @@ 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).
 
+## [0.9.15] - 2026-06-10
+
+### Fixed
+
+- **The LLM summary's `descriptionType` is now `"Other"`, not `"Abstract"`.** poster2json's `descriptions[]` entry is a machine-generated summary, so its type is `"Other"`. `"Abstract"` is reserved for the author's own formal poster abstract, which the platform attaches downstream (the submitter's abstract), never poster2json's summary. Post-processing had hardcoded `"Abstract"` since v0.9.3 (2026-06-05), a regression from the `"Other"` that was set on 2026-05-04 and reverted on 2026-05-19; for roughly three weeks every extracted summary was mislabeled as an abstract. `_postprocess_json` now sets `"Other"`.
+
+### Changed
+
+- **`descriptionType` is no longer requested in the extraction prompt.** The type is set deterministically in post-processing (see above), so asking the model for it was redundant and occasionally produced a value that was immediately overwritten. Both the primary and fallback prompts drop `descriptionType` from their JSON examples; output stays schema-valid.
+
 ## [0.9.14] - 2026-06-09
 
 ### Changed

poster2json/extract.py

@@ -1433,7 +1433,7 @@ def _generate(model, tokenizer, prompt: str, max_tokens: int) -> str:
   ],
   "titles": [{{"title": "Main Poster Title"}}],
   "subjects": [{{"subject": "keyword1"}}, {{"subject": "keyword2"}}, {{"subject": "keyword3"}}],
-  "descriptions": [{{"description": "A 3-4 sentence summary of the full poster...", "descriptionType": "Abstract"}}],
+  "descriptions": [{{"description": "A 3-4 sentence summary of the full poster..."}}],
   "researchField": null,
   "content": {{
     "sections": [
@@ -1469,7 +1469,7 @@ def _generate(model, tokenizer, prompt: str, max_tokens: int) -> str:
   "creators": [{{"name": "LastName, FirstName", "givenName": "FirstName", "familyName": "LastName", "affiliation": ["Institution"]}}],
   "titles": [{{"title": "Poster Title"}}],
   "subjects": [{{"subject": "keyword1"}}, {{"subject": "keyword2"}}],
-  "descriptions": [{{"description": "3-4 sentence summary of the full poster", "descriptionType": "Abstract"}}],
+  "descriptions": [{{"description": "3-4 sentence summary of the full poster"}}],
   "researchField": null,
   "content": {{
     "sections": [{{"sectionTitle": "Header", "sectionContent": "Full verbatim text of this section..."}}]
@@ -2177,13 +2177,17 @@ def _postprocess_json(
     result.pop("version", None)
     result["publicationYear"] = None
 
-    # descriptionType is auto-filled to its default; only the description text
-    # (the summary) is model-generated.
+    # The model's description is a machine-generated summary, so its
+    # descriptionType is "Other". "Abstract" is reserved for the author's own
+    # formal abstract, which the platform attaches downstream (the submitter's
+    # poster abstract), never poster2json's summary. The type is set
+    # deterministically here; the prompt no longer asks the model for it, so
+    # only the description text (the summary) is model-generated.
     descs = result.get("descriptions")
     if isinstance(descs, list):
         for d in descs:
             if isinstance(d, dict) and d.get("description"):
-                d["descriptionType"] = "Abstract"
+                d["descriptionType"] = "Other"
 
     # Ensure caption fields exist and normalize with auto-generated IDs
     for key, ctype in [("imageCaptions", "fig"), ("tableCaptions", "table")]:

pyproject.toml

@@ -1,7 +1,7 @@
 [tool.poetry]
 
 name = "poster2json"
-version = "0.9.14"
+version = "0.9.15"
 description = "Convert scientific posters (PDF/images) to structured JSON metadata using Large Language Models"
 
 packages = [{ include = "poster2json" }]