fix: update docling prediction provider to include word cells #118
Conversation
Signed-off-by: samiullahchattha <[email protected]>
Merge ProtectionsYour pull request matches the following merge protections and will not be merged until they are valid. 🔴 Require two reviewer for test updatesThis rule is failing.When test data is updated, we require two reviewers
🟢 Enforce conventional commitWonderful, this rule succeeded.Make sure that we follow https://www.conventionalcommits.org/en/v1.0.0/
|
Signed-off-by: samiullahchattha <[email protected]>
| } | ||
| pred_record.status = res.status | ||
|
|
||
| return pred_record | ||
|
|
||
| def set_word_cells(self, page: Page): | ||
| if page.parsed_page is None: |
There was a problem hiding this comment.
This check is already done on the caller side, through ... for p in res.pages if p.parsed_page. No need to repeat it here.
There was a problem hiding this comment.
Actually, the type checker was throwing error as it still sees parsed_page as Optional[SegmentedPdfPage] and throws the union-attr error. So I had to add this check explicitly
error: Item "None" of "SegmentedPdfPage | None" has no attribute "word_cells" [union-attr]
page.parsed_page.word_cells = page.cells
^~~~~~~~~~~~~~~~~~~~~~~~~~~
| f"Page {page.page_no} has no parsed_page, cannot set word_cells." | ||
| ) | ||
| return page.parsed_page | ||
| page.parsed_page.word_cells = page.cells |
There was a problem hiding this comment.
This is going to inject fake cells into the word_cells of a parsed_page. The only case in which that is currently a valid case is when the document was fully OCRed, hence this is a very partial workaround.
Also please consider that...
- The standard OCR options in Docling are generating line-level cells (TextCellUnit.LINE), not word-level cells. So this would not compare.
- The
parsed_pagewill be populated from Docling, this code would simply overwrite the word_cells without checking if some are present (they will be present whenever the source PDF hat programmatic cells)
Co-authored-by: Christoph Auer <[email protected]> Signed-off-by: samiuc <[email protected]>
Co-authored-by: Christoph Auer <[email protected]> Signed-off-by: samiuc <[email protected]>
Signed-off-by: samiullahchattha <[email protected]>
|
This is addressed here: docling-project/docling#1745 for consistent line-level cells. Word-level cells cannot be easily retrieved across all OCR backends. |
|
@samiuc We have released now a docling version where the OCR cells are included in the SegmentedPage (parsed_page) output on the line-level cells. |
DoclingPredictionProvider returned 0 F1, Recall, and Precision scores during evaluation and upon investigation, I found out that the word cells were not being populated in the SegmentedPage object that OCREvaluator uses for calculating metrics, so added a fix to copy cells from the original page to the parsed_page (SegmentedPage) word cells object.