Extracting from pdfs

[efuae]

Hello, I am using Grobid for my project and I am working with PDF Drug Labels. I have noticed a few things that happen when the pdf is extracted into xml:

1. It often times does not extract the text that comes right after an image
2. It sometimes captures a new head into the preceding header. For example after extracting section 12.3, it extracts section 12.4 as a continuation of the preceding header.

Could this be looked at please?

[lfoppiano]

HI @efuae,
which version of Grobid are you using?

I'm working on the first problem, but it's not yet finished. You can test it using this branch. It's still experimental, but it's trying to solve the problem you're experiencing.

For the second problem, could you please provide some examples? I'd say we would need more training data.

Regards
Luca

[efuae]

I am using Grobid 0.8.1 via Docker container.

For the second problem, this is an example below. I want to extract the section 12.5 and everything under that section however, because 12.5 is nested in 12.4 it does not extract that part. Sometimes not just the section number but the information captured under that section would be fused with the preceding section -> 7409.pdf

Pharmacodynamics

BENTYL can inhibit the secretion of saliva and sweat, decrease gastrointestinal secretions and motility, cause drowsiness, dilate the pupils, increase heart rate, and depress motor function 12.5 Pharmacokinetics

Absorption and Distribution

In man, dicyclomine is rapidly absorbed after oral administration, reaching peak values within 60-90 minutes. Mean volume of distribution for a 20 mg oral dose is approximately 3.65 L/kg suggesting extensive distribution in tissues.

There are also instances where it misses paragraphs in the extraction. In this xml that was extracted from a pdf, there are 3 subheadings before "excretion" but it completely missed them :( -> 6134.pdf

lesser extent CYP2C9 and CYP2D6, are responsible for conversion of methadone to EDDP and other inactive metabolites, which are excreted mainly in the urine.

Excretion

The elimination of methadone is mediated by extensive biotransformation, followed by renal and fecal excretion. Published reports indicate that after multiple dose administration the terminal half-life (T 1/2 ) was highly variable and ranged between 8 to 59 hours in different studies. Since methadone is lipophilic, it has been known to persist in the liver and other tissues. The slow release from the liver and other tissues may prolong the duration of methadone action despite low plasma concentrations.

Pharmacokinetics in Special Populations Pregnancy

The disposition of oral methadone has been studie

[lfoppiano]

Hi @efuae,
I did a more dept analysis of the problems.

7409.pdf

the first file's issue with sections 12.4 (correclty identified as <section>:

12	12	1	12	12	12	2	12	12	12	BLOCKIN	LINESTART	ALIGNEDLEFT	NEWFONT	SAMEFONTSIZE	1	0	NOCAPS	ALLDIGIT	0	NOPUNCT	10	3	0	UNKNOWN	0	0	I-<section>
.	.	.	.	.	.	.	.	.	.	BLOCKIN	LINEIN	ALIGNEDLEFT	SAMEFONT	SAMEFONTSIZE	1	0	ALLCAP	NODIGIT	1	DOT	10	3	0	UNKNOWN	0	0	<section>
4	4	4	4	4	4	4	4	4	4	BLOCKIN	LINEIN	ALIGNEDLEFT	SAMEFONT	SAMEFONTSIZE	1	0	NOCAPS	ALLDIGIT	1	NOPUNCT	10	3	0	UNKNOWN	0	0	<section>
Pharmacodynamics	pharmacodynamics	P	Ph	Pha	Phar	s	cs	ics	mics	BLOCKIN	LINEEND	ALIGNEDLEFT	SAMEFONT	SAMEFONTSIZE	1	0	INITCAP	NODIGIT	0	NOPUNCT	10	3	0	UNKNOWN	0	0	<section>

and 12.5 incorrectly classified as <paragraph>:

function	function	f	fu	fun	func	n	on	ion	tion	BLOCKIN	LINEEND	ALIGNEDLEFT	SAMEFONT	SAMEFONTSIZE	0	0	NOCAPS	NODIGIT	0	NOPUNCT	10	4	0	UNKNOWN	0	0	<paragraph>
12	12	1	12	12	12	2	12	12	12	BLOCKIN	LINESTART	ALIGNEDLEFT	NEWFONT	SAMEFONTSIZE	1	0	NOCAPS	ALLDIGIT	0	NOPUNCT	10	4	0	UNKNOWN	0	0	<paragraph>
.	.	.	.	.	.	.	.	.	.	BLOCKIN	LINEIN	ALIGNEDLEFT	SAMEFONT	SAMEFONTSIZE	1	0	ALLCAP	NODIGIT	1	DOT	10	4	0	UNKNOWN	0	0	<paragraph>
5	5	5	5	5	5	5	5	5	5	BLOCKIN	LINEIN	ALIGNEDLEFT	SAMEFONT	SAMEFONTSIZE	1	0	NOCAPS	ALLDIGIT	1	NOPUNCT	10	4	0	UNKNOWN	0	0	<paragraph>
Pharmacokinetics	pharmacokinetics	P	Ph	Pha	Phar	s	cs	ics	tics	BLOCKIN	LINEEND	ALIGNEDLEFT	SAMEFONT	SAMEFONTSIZE	1	0	INITCAP	NODIGIT	0	NOPUNCT	10	4	0	UNKNOWN	0	0	<paragraph>
Absorption	absorption	A	Ab	Abs	Abso	n	on	ion	tion	BLOCKIN	LINESTART	LINEINDENT	NEWFONT	SAMEFONTSIZE	0	0	INITCAP	NODIGIT	0	NOPUNCT	10	4	0	UNKNOWN	0	0	I-<section>
and	and	a	an	and	and	d	nd	and	and	BLOCKIN	LINEIN	LINEINDENT	SAMEFONT	SAMEFONTSIZE	0	0	NOCAPS	NODIGIT	0	NOPUNCT	10	4	0	UNKNOWN	0	0	<section>
Distribution	distribution	D	Di	Dis	Dist	n	on	ion	tion	BLOCKIN	LINEEND	LINEINDENT	SAMEFONT	SAMEFONTSIZE	0	0	INITCAP	NODIGIT	0	NOPUNCT	10	4	0	UNKNOWN	0	0	<section>
In	in	I	In	In	In	n	In	In	In	BLOCKIN	LINESTART	LINEINDENT	SAMEFONT	SAMEFONTSIZE	0	0	INITCAP	NODIGIT	0	NOPUNCT	10	4	0	UNKNOWN	0	0	I-<paragraph>
man	man	m	ma	man	man	n	an	man	man	BLOCKIN	LINEIN	LINEINDENT	

as a result the head of section is lost:

<div
                xmlns="http://www.tei-c.org/ns/1.0">
                <head n="12.4">Pharmacodynamics</head>
                <p>BENTYL can inhibit the secretion of saliva and sweat, decrease gastrointestinal secretions and motility, cause drowsiness, dilate the pupils, increase heart rate, and depress motor function 12.5 Pharmacokinetics</p>
            </div>

Solution for this issue would be to add more training data for the fulltext and segmentation model.

6134.pdf

The second article's, the heading before "excretion" are classified as abstract. You could still extract them with a trick, if you consider the abstract as the first part of the article, since your pdfs are not articles and they don't have abstract, you could stack the abstract before the body. It's not guaranteed that will work always but it could be a short term solution.

For solving the second issue, we could create training data for the segmentation using your articles, and create a new process flavor or specialisation that override the standard segmentation model and should improve the extraction of your articles.
You can check the concept and some examples here: https://grobid.readthedocs.io/en/latest/Grobid-specialized-processes/

This is a new feature that will be released with version 0.8.2 (master version), but it has been tested and can be used. However, it might need some more information in the documentation, so feel free to let me know if there is anything not clear.

UPDATE: I've checked with the flavor light/article-ref and the result is better for 6134.pdf, except that for the head of sections that will definitely need more training data. For 7409.pdf no much difference, it seems that only one model might require training data (fulltext model)

[efuae]

Hi,
Thank you so much for the detailed response and the suggestion. I’ll definitely try the light/article-ref flavor and see how it improves the results for my PDFs. I’ll also explore the documentation you shared.

If I run into any issues or have questions while setting it up, I’ll be sure to reach out.

Thanks again for your support!

[lfoppiano]

@efuae could you give it a try with the latest master version? We've ran a retrain of the fulltext model and the extraction of head / paragraphs should have improved, even your documents weren't used in the training.

Archive.zip