docx parsing headings with outline format numbered-style applied is not working properly.

[jbtelice]

Bug

msword_backend.py doesn't parse docx files with headings ("Heading 1, Heading 2", etc) properly. When Outline format is applied to a heading:

and has "Numbering style" activated as it follows:

the parser label that paragraph as a list (which is true) but no as a section_header ( which is true as well). That's ambiguous and it potentially prone to unnoticed errors( Imagine when you have to process a lot of docx files in which you are NOT allowed to edit...)

The consequences are:

• If there is no proper parsing, that carries errors to the HierarchicalChunker too.
• The user cannot proper access to the document structure, proceeding as explained here

Here is an screenshot of what I mean:

Clean headings (No outline format)

Headings with outline format

Steps to reproduce

Here are the docx samples:

bug_example.docx

bug_example.docx

bug_example_with_list_headings.docx

bug_example_with_list_headings.docx

from docling.document_converter import DocumentConverter

DOC_SOURCE_ORIGINAL = "./bug_example.docx"
DOC_SOURCE_MODIFIED = "./bug_example_with_list_headings.docx"

doc_original = DocumentConverter().convert(source=DOC_SOURCE_ORIGINAL).document
doc_edited = DocumentConverter().convert(source=DOC_SOURCE_MODIFIED).document

print(f"---- {DOC_SOURCE_ORIGINAL} ----")

doc_original.print_element_tree()

print(f"---- {DOC_SOURCE_MODIFIED} ----")

doc_edited.print_element_tree()

Here is the output:

---- ./bug_example.docx ----
 0: unspecified with name=_root_
  1: section with name=header-0
   2: section_header    
    3: paragraph  # empty
    4: paragraph
    5: paragraph  # empty
    6: section_header  
     7: paragraph # empty
     8: paragraph
     9: paragraph # empty
    10: section_header
     11: paragraph   # empty
     12: paragraph
     13: paragraph  # empty
     14: paragraph
     15: paragraph  # empty
     16: section_header
      17: paragraph  # empty
      18: paragraph
      19: paragraph  # empty
   20: section_header
    21: paragraph  # empty
    22: paragraph
    23: paragraph  # empty

---- ./bug_example_with_list_headings.docx ----
 0: unspecified with name=_root_
  1: list with name=list    # should be header
   2: list_item
  3: paragraph
  4: paragraph
  5: paragraph
  6: list with name=list   # should be header
   7: list_item
  8: paragraph
  9: paragraph
  10: paragraph
  11: list with name=list   # should be header
   12: list_item
  13: paragraph
  14: paragraph
  15: paragraph
  16: paragraph
  17: paragraph
  18: list with name=list   # should be header
   19: list_item
  20: paragraph
  21: paragraph
  22: paragraph
  23: list with name=list   # should be header
   24: list_item
  25: paragraph
  26: paragraph
  27: paragraph
(.venv) 

Be aware that in the first example, there is no name (label.name) in section_headers

Docling version

Docling version: 2.15.1
Docling Core version: 2.15.1
Docling IBM Models version: 3.2.1
Docling Parse version: 3.1.1

Python version

Python 3.11.9

Additional Context

Digging into the code, I notice some other things worth to explore too:

[split_text_and_number](

https://github.com/DS4SD/docling/blob/1976584be15b6dea7451001adbfd8b9dc3422235/docling/backend/msword_backend.py#L170) . That regex is not trimming the match.groups()

That implies wrong label parsing

Regarding style

line 200

¿Shouldn't be name instead of style_id?

Hope it helps,

Let me know if you need more information.

Have a nice day!

[MiguelAngelTorres]

I found the same problem parsing other file that combines both outline format and non-outline format in the same file.

The non-outline format produces that the outline format output is neither generated correctly.

Docling version

Docling version: 2.15.1
Docling Core version: 2.15.1
Docling IBM Models version: 3.2.1
Docling Parse version: 3.1.1

[MiguelAngelTorres]

Update

It seems the problem is not the outline. Testing with different outlines and formats doesn't resolve the issue.

The point is the label parsing:

• There is an additional whitespace on labels.
• Microsoft Word changes the style_id depending on the language settings. I found a post in Microsoft's forum about that.

Coding in local the changes suggested by @jbtelice worked for me, but it would be nice to include them in the core code.

[jbtelice]

Hi!

There is an additional whitespace on labels.

Yep, indeed, that's what I mean (In the additional context section). In fact, the mechanism for assigning label_str and label_level is language-dependant if you rely on "style_id", and could be fixed if you take name instead of (label name seems to be language-invariant, but I didn't have the time to check it out carefully)

Microsoft Word changes the style_id depending on the language settings. I found a post in Microsoft's forum about that.

Thanks @MiguelAngelTorres for pointing out the reference. 👀

[maxmnemonic]

Same issue: #612

[maxmnemonic]

Thanks @jbtelice, @MiguelAngelTorres,
We investigated and with this PR, issue should be resolved: #842