Wrong layout extraction (texts blended between two columns)

[jeromemassot]

Bug

The text extracted is interleaved from the two columns of the original paper.
You can retrieve the paper at https://pmc.ncbi.nlm.nih.gov/articles/PMC113777

Output:
=== 0 ===
('Journal of Biomedicine and Biotechnology · 1:2 (2001)89-90 · PII. '
'S111072430100016X · http://jbb.hindawi.com/ CORRESPONDENCEARTICLE')
=== 1 ===
'The role of neuraminidase inhibitors in the treatment'
=== 2 ===
'and prevention of influenza'
=== 3 ===
('Naem Shahrour * American International Health Council, 414 South Craig St. '
'#269, Pittsburgh, PA 15213, USA The causative agents of acute respiratory '
'infections in children and adults are mostly thought to be viruses. Many '
'types of viruses could cause similar symptoms of ARI. Among them, influenza '
'viruses A and B and respiratory syncytial virus (RSV)are thought to be the '
'most important because of the severity of illness after infection and their '
'high communicability in the human population [1]. Type C influenza virus '
'modalities have been investigated and some have already been introduced to '
'practice. The most recent breakthrough in that direction is the introduction '
'of neuraminidase inhibitors for the treatment and prevention of influenza '
'infection. The neuraminidase is a surface glycoprotein that is composed of '
'eleven conserved residues [4] which catalyze the cleavage of sialic acid '
'residues terminally linked to glycopro-')

Third TextItem interleaves text from the two columns.

Steps to reproduce

def convert_doc(path):

    # define the conversion options
    pipeline_options = PdfPipelineOptions()
    pipeline_options.do_ocr = True
    pipeline_options.do_table_structure = True
    pipeline_options.table_structure_options.do_cell_matching = True
    pipeline_options.ocr_options.lang = ["en"]
    pipeline_options.accelerator_options = AcceleratorOptions(
        num_threads=8, device=AcceleratorDevice.AUTO
    )

    # create the doc converter
    doc_converter = DocumentConverter(
        format_options={
            InputFormat.PDF: PdfFormatOption(pipeline_options=pipeline_options)
        }
    )

    # convert the document
    result = doc_converter.convert(path)
    return result

doc = convert_doc(os.path.join("PMC113777.pdf")).document

i = 0
for item, level in doc.iterate_items():
    if isinstance(item, TextItem):
        print(f"=== {i} ===")
        pprint.pprint(item.text)
        i += 1

Docling version

2.14.0

Python version

3.12.3

[MahmoudAtef4499]

Me too.

[cau-git]

@jeromemassot confirmed, this is apparently a bug in docling-parse. I opened an issue to track it there, and I will close this issue.

For verification, you can run this to visualize the problem:

docling --debug-visualize-cells --debug-visualize-layout S111072430100016X.pdf    

and this workaround will avoid the problem by using the pypdfium backend instead:

docling --debug-visualize-cells --debug-visualize-layout --pdf-backend pypdfium2 S111072430100016X.pdf