Missing or Possibly Incorrect Parsing of PDF

[Keegan-Vaz]

Grobid version

grobid/grobid:0.8.2-full - Docker - Deep Learning Model

Operating System and architecture (arm64, amd64, x86, etc.)

No response

What is your Java version

No response

Log and information

No response

Further information

Hello team,

We’ve noticed that GROBID is not parsing email addresses from some PDFs, even though the emails are clearly visible in the document. I’ve attached one such example where the email is right on the first page, but it isn’t captured in the extracted XML.

Additionally, for one of the PDFs, the author’s name was parsed as Mandarin characters, even though it appears in English in the original file. The corresponding PDF and the generated XML are both attached for reference.

Please let me know if you need any additional details or sample files to investigate this further.

Thanks!

file.pdf
tei.xml

The one below was parsed with the Authors in Mandarin Text
parsed_with_chinese.xml
English_PDF.pdf

[lfoppiano]

Hi @Keegan-Vaz the first PDF indeed has problems in extracting the authors, one of the two is mixed up in the affiliation.
For the second article, there is nothing that can be done either with Grobid or with any text extraction tools, because the text layer of the authors (a transparent layer not visible, but that you can verify by highlighting and copy-paste) is in Chinese, reason why you see them in the output TEI XML file.

[Keegan-Vaz]

We’ve identified another issue during XML generation. The XML file failed to extract the actual paper title and instead randomly picked one of the references as the paper’s title.

Observed Behavior:

The title field in the generated XML is populated with a reference entry instead of the correct paper title.

The actual paper title is missing.

Do let me know if anything is possible with regards to some issues like this.

Example 1
file.pdf
tei.xml

Example 2
file.pdf
tei.xml

[lfoppiano]

@Keegan-Vaz I ran this to https://huggingface.co/spaces/lfoppiano/grobid-full (0.8.2-full) and:

• Example 1 (autism) the title is mixed with the head note - nothing we can do at the moment, we need to train the header model to distinguish the two entities.

<titleStmt>
                <title level="a" type="main">UNDERSTANDING SOCIAL COGNITIVE DEFICITS IN CHILDREN WITH 1 AUTISM SPECTRUM DISORDER Understanding Social Cognitive Deficits in Children with Autism Spectrum Disorders: Implications for Clinical Practice and Educational and Societal Integration</title>
            </titleStmt>

• Example 2 (videogames) the extraction looks good to me:

<titleStmt>
                <title level="a" type="main">Perceived value of video games, but not hours played, predicts mental well-being in adult Nintendo players</title>
                <funder>
                    <orgName type="full">Huo Family Foundation</orgName>
                </funder>
                <funder ref="#_5ruyHNK">
                    <orgName type="full">Forskningsrå det för hälsa</orgName>
                </funder>
                <funder ref="#_WMpVpDr">
                    <orgName type="full">UK Economic and Social Research Council</orgName>
                </funder>
            </titleStmt>

[lfoppiano]

One immediate action that can correct the bibliographical information is to enable the consolidation ("header consolidation").

TLDR: it can help by matching data on crossref, but it will result in slower results and it's not always matching (like in example 1)

Long answer: https://grobid.readthedocs.io/en/latest/Consolidation/

[Keegan-Vaz]

Thank you so much for your quick response. I'll try running the header consolidation and check the results for the same and get back to you if i have any other questions.
Good Day

[lfoppiano]

Training data added in #1319 for the segmentation model. Header model still to be updated.