From 009be8c9584100d2e85fd8b3ad0a554e0a8a89e2 Mon Sep 17 00:00:00 2001
From: Research Assistant <research@example.com>
Date: Fri, 5 Jun 2026 15:02:00 +0800
Subject: [PATCH] fix: stop paper_title pollution, text-to-heading, table HTML
 leak, author recovery

---
 paperforge/worker/ocr_blocks.py   |  2 +-
 paperforge/worker/ocr_metadata.py | 20 ++++++++---------
 paperforge/worker/ocr_render.py   |  4 +++-
 paperforge/worker/ocr_roles.py    | 36 ++++++++++++++++++++++++++-----
 tests/test_ocr_roles.py           |  2 +-
 5 files changed, 45 insertions(+), 19 deletions(-)

diff --git a/paperforge/worker/ocr_blocks.py b/paperforge/worker/ocr_blocks.py
index abd6552f..4821d999 100644
--- a/paperforge/worker/ocr_blocks.py
+++ b/paperforge/worker/ocr_blocks.py
@@ -25,7 +25,7 @@ def build_structured_blocks(raw_blocks: list[dict]) -> list[dict]:
         index_default = True
         if role.role in {"noise", "page_header", "page_footer", "frontmatter_noise"}:
             render_default = False
-        if role.role in {"noise", "frontmatter_noise"}:
+        if role.role in {"noise", "frontmatter_noise", "table_html"}:
             index_default = False
         row = {
             "paper_id": block["paper_id"],
diff --git a/paperforge/worker/ocr_metadata.py b/paperforge/worker/ocr_metadata.py
index cdb634f6..6d8a6143 100644
--- a/paperforge/worker/ocr_metadata.py
+++ b/paperforge/worker/ocr_metadata.py
@@ -78,29 +78,27 @@ def resolve_metadata(
 
     # --- authors ---
     zotero_authors = source_metadata.get("authors", [])
+    ocr_authors_text = frontmatter_candidates.get("authors_text", "")
+    ocr_author_list = [a.strip() for a in ocr_authors_text.split(",") if a.strip()] if ocr_authors_text else []
+
     if isinstance(zotero_authors, list) and len(zotero_authors) > 0:
         resolved["authors"] = {
             "value": zotero_authors,
             "source": "zotero",
             "confidence": 0.99,
         }
+    elif ocr_author_list:
+        resolved["authors"] = {
+            "value": ocr_author_list,
+            "source": "ocr_frontmatter",
+            "confidence": 0.6,
+        }
     else:
         resolved["authors"] = {
             "value": [],
             "source": "unknown",
             "confidence": 0.3,
         }
-    ocr_authors_text = frontmatter_candidates.get("authors_text", "")
-    if ocr_authors_text:
-        ocr_author_list = [a.strip() for a in ocr_authors_text.split(",") if a.strip()]
-        if ocr_author_list and ocr_author_list != zotero_authors:
-            resolved["authors"]["alternatives"] = [
-                {
-                    "value": ocr_author_list,
-                    "source": "ocr_frontmatter",
-                    "confidence": 0.6,
-                }
-            ]
 
     # --- year ---
     zotero_year = source_metadata.get("year", 0)
diff --git a/paperforge/worker/ocr_render.py b/paperforge/worker/ocr_render.py
index 87e6bd1f..ea981ab7 100644
--- a/paperforge/worker/ocr_render.py
+++ b/paperforge/worker/ocr_render.py
@@ -92,10 +92,12 @@ def render_fulltext_markdown(
         if not block.get("render_default", True):
             continue
         role = block.get("role", "")
-        if role in ("abstract_heading", "abstract_body", "figure_caption", "table_caption", "frontmatter_noise"):
+        if role in ("abstract_heading", "abstract_body", "figure_caption", "table_caption", "frontmatter_noise", "table_html"):
             continue
 
         text = _normalize_latex(block.get("text", ""))
+        if text.strip().lower().startswith("<table"):
+            continue
         block_page = block.get("page")
 
         if block_page is not None and block_page != current_page:
diff --git a/paperforge/worker/ocr_roles.py b/paperforge/worker/ocr_roles.py
index f9be1999..c123f337 100644
--- a/paperforge/worker/ocr_roles.py
+++ b/paperforge/worker/ocr_roles.py
@@ -12,7 +12,7 @@ class RoleAssignment:
 
 
 _HEADING_NUMBER_PATTERN = re.compile(
-    r"^\d+(?:\.\d+)*\s+\w",
+    r"^\d+(?:\.\d+)*\s+[A-Z]",
 )
 
 _FIGURE_PREFIX_PATTERN = re.compile(
@@ -141,10 +141,17 @@ def assign_block_role(
                 confidence=0.85,
                 evidence=[f"paragraph_title label with numbering: {text[:60]}"],
             )
+        page_num = block.get("page", 1) or 1
+        if page_num <= 1:
+            return RoleAssignment(
+                role="paper_title",
+                confidence=0.7,
+                evidence=[f"unnumbered paragraph_title on page 1, likely paper title: {text[:60]}"],
+            )
         return RoleAssignment(
-            role="paper_title",
-            confidence=0.7,
-            evidence=[f"unnumbered paragraph_title, likely paper title: {text[:60]}"],
+            role="section_heading",
+            confidence=0.6,
+            evidence=[f"unnumbered paragraph_title on page {page_num}, treated as heading: {text[:60]}"],
         )
 
     if raw_label == "figure_title":
@@ -189,6 +196,13 @@ def assign_block_role(
             evidence=["abstract label from Paddle OCR"],
         )
 
+    if raw_label == "reference_content":
+        return RoleAssignment(
+            role="reference_item",
+            confidence=0.85,
+            evidence=[f"reference content label: {text[:60]}"],
+        )
+
     # text with reference-like pattern
     if _looks_like_reference(text):
         return RoleAssignment(
@@ -202,6 +216,14 @@ def assign_block_role(
         stripped = text.strip().lstrip("*•·-–—_")
         lower_txt = stripped.lower()
 
+        # Check for inline table HTML
+        if text.strip().lower().startswith("<table"):
+            return RoleAssignment(
+                role="table_html",
+                confidence=0.95,
+                evidence=[f"inline table HTML"],
+            )
+
         # Check for abstract heading (may appear as text block, not paragraph_title)
         if lower_txt.startswith("abstract") and len(text) < 30:
             return RoleAssignment(
@@ -290,7 +312,11 @@ def assign_block_role(
                 confidence=0.7,
                 evidence=[f"frontmatter noise text: {text[:60]}"],
             )
-        if _has_heading_numbering(text):
+        if (
+            _has_heading_numbering(text)
+            and len(text) < 80
+            and ". " not in text
+        ):
             return RoleAssignment(
                 role="section_heading" if re.match(r"^\d+\s", text) else "subsection_heading",
                 confidence=0.65,
diff --git a/tests/test_ocr_roles.py b/tests/test_ocr_roles.py
index 40999d32..639bc75c 100644
--- a/tests/test_ocr_roles.py
+++ b/tests/test_ocr_roles.py
@@ -123,7 +123,7 @@ def test_stabilize_role_classifies_paper_title() -> None:
     from paperforge.worker.ocr_roles import assign_block_role
 
     assignment = assign_block_role(
-        block={"block_label": "paragraph_title", "block_content": "Test Paper About Science"},
+        block={"block_label": "paragraph_title", "block_content": "Test Paper About Science", "page": 1},
         page_blocks=[],
     )