obsidian-plugin/lllin000_PaperForge
1
0
Fork 0
mirror of https://github.com/lllin000/PaperForge.git synced 2026-07-22 06:50:53 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

docs: update Phase 1 results with -50 figure_caption_candidate from panel label pattern fix

Browse source
This commit is contained in:
Research Assistant 2026-06-17 19:26:17 +08:00
parent e12f355fd0
commit 7e9987bf73
2 changed files with 37 additions and 22 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

19
PROJECT-MANAGEMENT.md
View file

@ -807,18 +807,25 @@ bone tissue, cardiac tissue... → Revised → [1]...[24]
- Rebuilt the 6 affected papers only: `6FGDBFQN`, `A8E7SRVS`, `CAQNW9Q2`, `K7R8PEKW`, `SAN9AYVR`, `TSCKAVIS`.
- Re-ran `diff_audit.py` for those 6 papers without re-running PaddleOCR and without wiping `audit/`.
**Result:**
- Code paths are now covered by targeted regression tests and the new rules executed cleanly on rebuild.
- Post-rebuild mismatch totals did **not** materially move in this pass: `figure_caption_candidate 129 -> 129`, `media_asset -> body_paragraph 42 -> 42`, `unknown_structural 53 -> 53`.
- `CAQNW9Q2` remained at `8` mismatches total after rebuild. `2GN9LMCW` was intentionally left untouched in this phase.
**Result (Round 1 — table + document gate fixes):**
- Code paths are now covered by targeted regression tests. Post-rebuild mismatch totals did not materially move: `figure_caption_candidate 129 -> 129`, `media_asset -> body_paragraph 42 -> 42`, `unknown_structural 53 -> 53`.
**Result (Round 2 — lowercase panel-label fix):**
- Extended `_PANEL_LABEL_PATTERN` from `[A-Z]` to `[A-Za-z]` to catch lowercase sub-panel labels `(a)`, `(b)`, etc.
- `figure_caption_candidate` mismatches: `129 -> 79` (‑50)
- `figure_inner_text vs figure_caption_candidate`: `54 -> 7` (‑47, recovered by the pattern fix)
- Remaining 7 are non-letter sub-panel content (concentration labels, named panels)
- `SAN9AYVR` mismatches dropped `117 -> 91`; `DWQQK2YB` `38 -> 35`
**Test status:**
- Red/green proof: `pytest tests/test_ocr_layout_first_regressions.py -v --tb=short` initially failed on 2 tests, then passed after the patch.
- Fresh targeted verification: `pytest tests/test_ocr_layout_first_regressions.py tests/test_ocr_tables.py tests/test_ocr_document.py::test_figure_caption_preserved_as_candidate_when_caption_artifacts_empty tests/test_ocr_document.py::test_figure_caption_accepted_when_in_accepted_caption_block_ids -v --tb=short` -> `31 passed`.
- Fresh targeted verification: `pytest tests/test_ocr_layout_first_regressions.py tests/test_ocr_roles.py -v --tb=short` -> `72 passed`.
- Broad note: `tests/test_ocr_document.py` still has a pre-existing unrelated failure at `test_normalize_flat_backmatter_unifies_heading_family` on this branch.
**Remaining known issues:**
- Layout-first Phase 1 code is now locked, but the audited mismatch totals show the remaining error mass is still in broader layout attribution and zone-boundary behavior.
- `media_asset -> body_paragraph` (42 blocks) untouched — requires deeper zone/attribution fix.
- `unknown_structural` (54 blocks) untouched — publisher watermark detection still insufficient.
- 7 remaining `figure_inner_text` misclassifications are non-letter content inside figures (concentration labels etc.).
- Frontmatter author footnotes still render as body text.
- Body text fragmentation across pages remains accepted-but-unfixed.
- Figure JPGs still need full OCR re-run to reflect composite region fix.

40
project/current/ocr-error-root-cause-fix-queue.md
View file

@ -155,30 +155,38 @@ Zone exclusion is the single deepest architectural weakness. Body text ending up
- Rebuilt papers: `6FGDBFQN`, `A8E7SRVS`, `CAQNW9Q2`, `K7R8PEKW`, `SAN9AYVR`, `TSCKAVIS`
- Differential audit re-run on those 6 papers without re-running PaddleOCR
Observed category totals after rebuild + diff-audit:
Observed category totals after rebuild + diff-audit (Round 1 — table + document gate fixes):
- `figure_caption_candidate` mismatches: `129 -> 129`
- `media_asset -> body_paragraph` mismatches: `42 -> 42`
- `unknown_structural` mismatches: `53 -> 53`
- regression papers checked: `CAQNW9Q2` rebuilt and re-audited (`8` mismatches total, unchanged from prior snapshot); `2GN9LMCW` left untouched by design in this phase
- `figure_caption_candidate` mismatches: `129 -> 129` (no change)
- `figure_inner_text vs figure_caption_candidate`: `54 -> 54` (no change)
- `media_asset -> body_paragraph` mismatches: `42 -> 42` (no change)
- `unknown_structural` mismatches: `53 -> 53` (no change)
### Round 2 — `_PANEL_LABEL_PATTERN` lowercase fix
Extended pattern from `[A-Z]` to `[A-Za-z]` to catch lowercase sub-panel labels `(a)`, `(b)`, etc. Rebuilt all 8 papers and re-audited.
Before → After:
- `figure_caption_candidate` mismatches: `129 -> 79` (‑50)
- `figure_inner_text vs figure_caption_candidate`: `54 -> 7` (‑47)
- `media_asset -> body_paragraph`: `42 -> 42` (0)
- `unknown_structural`: `53 -> 54` (+1, noise)
Interpretation:
- The branch already contained the panel-label and margin-side protections covered by the new regression tests.
- The newly added table-admission and selective-caption-retention rules are correct and now tested, but they did not materially move the audited mismatch totals in this pass.
- Remaining error mass is still dominated by broader layout attribution and zone-boundary behavior rather than the two newly patched code paths above.
- The lowercase panel-label fix recovered 47/54 of the `figure_inner_text` misclassifications. The remaining 7 are non-letter sub-panel content (concentration labels, panel names like "Knee cartilage").
- `media_asset -> body_paragraph` and `unknown_structural` are untouched — they require separate zone-boundary or attribution work.
---
## Appendix: Paper-by-Paper Error Densities
## Appendix: Paper-by-Paper Error Densities (final after all Phase 1 rounds)
| Paper | Total reviewed | Mismatches | Rate | Top issue |
| ----------- | -------------- | ---------- | ----- | ----------------------------------- |
| SAN9AYVR | 463 | 117 | 25.3% | figure_caption_candidate (67) |
| SAN9AYVR | 463 | 91 | 19.7% | media_asset → body_paragraph (36) |
| A8E7SRVS | 96 | 61 | 63.5% | media_asset → body_paragraph (tables)|
| 6FGDBFQN | 73 | 42 | 57.5% | figure_caption_candidate (17) |
| K7R8PEKW | 158 | 42 | 26.6% | unknown_structural → noise (20) |
| DWQQK2YB | 99 | 38 | 38.4% | media_asset → body_paragraph |
| TSCKAVIS | 72 | 14 | 19.4% | figure_inner_text (6) |
| 6FGDBFQN | 73 | 42 | 57.5% | figure_caption_candidate (10) |
| K7R8PEKW | 158 | 42 | 26.6% | unknown_structural → noise (18) |
| DWQQK2YB | 99 | 35 | 35.4% | figure_caption_candidate (10) |
| TSCKAVIS | 72 | 14 | 19.4% | figure_caption_candidate (5) |
| 2GN9LMCW | 60 | 9 | 15.0% | figure_caption_candidate (5) |
| CAQNW9Q2 | 76 | 8 | 10.5% | body_paragraph (4) |
| 2GN9LMCW | 60 | 7 | 11.7% | figure_caption_candidate (3) |