Thanks !
I think the change in TEIFormatter.java is doing nothing? Just moving stuff that do the same (simply put a default space at inline position of a figure to avoid agglutination of text).

From the discussion #956, I am updating the PR with post-processing limited to processShort(), instead of changing the general fulltext model decoding.

Coverage increased (+0.02%) to 39.928% when pulling 6c8b888 on feature/funding-statement into 8e53a7d on master.

The changes seem to work fine, on the example from #956 we obtain:

           <div type="funding">
                <div
                    xmlns="http://www.tei-c.org/ns/1.0">
                    <p>This study was supported by the South Asian Clinical Toxicology Research Collaboration, which is funded by The Wellcome Trust/ National Health and Medical Research Council International Collaborative Research Grant GR071669MA. The funding bodies had no role in analyzing or interpreting the data or writing the article.</p>
                </div>
            </div>

No runtime error on PMC_sample set and biorxiv_test_2000.

I found several cases of empty funding where the papers has the funding statement in the back of the document, the TEI serialisation calls getSectionAsTEI() which does not uses the processShort().

E.g. this document: bj4360053.pdf

Unfortunately, the result does not help, and the hack in processShort() does not work because of the different label at the first place won't fix the sequence.

FUNDING	funding	F	FU	FUN	FUND	G	NG	ING	DING	BLOCKSTART	LINESTART	ALIGNEDLEFT	NEWFONT	HIGHERFONT	1	0	ALLCAP	NODIGIT	0	NOPUNCT	0	10	0	NUMBER	0	0	I-<section>
This	this	T	Th	Thi	This	s	is	his	This	BLOCKSTART	LINESTART	ALIGNEDLEFT	NEWFONT	LOWERFONT	0	0	INITCAP	NODIGIT	0	NOPUNCT	0	10	0	NUMBER	0	0	I-<table>
work	work	w	wo	wor	work	k	rk	ork	work	BLOCKIN	LINEIN	ALIGNEDLEFT	SAMEFONT	SAMEFONTSIZE	0	0	NOCAPS	NODIGIT	0	NOPUNCT	0	10	0	NUMBER	0	0	<table>
was	was	w	wa	was	was	s	as	was	was	BLOCKIN	LINEIN	ALIGNEDLEFT	SAMEFONT	SAMEFONTSIZE	0	0	NOCAPS	NODIGIT	0	NOPUNCT	0	10	0	NUMBER	0	0	<table>
supported	supported	s	su	sup	supp	d	ed	ted	rted	BLOCKIN	LINEIN	ALIGNEDLEFT	SAMEFONT	SAMEFONTSIZE	0	0	NOCAPS	NODIGIT	0	NOPUNCT	0	10	0	NUMBER	0	0	<table>
[...]

I'm not sure how I could fix this properly

We could wait for future versions which will remove the tables and figures from the fulltext model, or apply the same fix as in processShort for every table and figure labels (not just the ones starting the sequence). What do you think?

At the moment is better to fix it, as it is as some data are lost. I can apply the same fix for processshort for all the figures and tables labels.

Actually postProcessLabeledAbstract() seems to already do this more violent stuff.

Using the biorxiv corpus for end2end evaluation, we get this result:

funding_stmt         99.79        59.4         42.42        49.49        745    

So the two problematic cases are working fine with the current version, yeah :D

I cleaned the previous hack in processShort() and replaced it with postProcessFullTextLabeledText(), to be sure it is always applied after calling processShort() (so this applied to the funding statement in header too now, otherwise it was only applied to the funding statement at the end of a paper).

Thanks !!