We use #grobid and the plos1000 #goldstandard as a baseline to compare the performance of LLM-based solutions.
Takeaways:
- Grobid still better choice for literature similar to the type it was trained on (mostly English-language STEM scholarship), since it is much faster & less resource-intensive
- For footnoted literature, experiments with LLamore/#Gemini show 3x better performance
Reminder: PKP invites communities to register for its Software Development Update webinar on December 16th, 2024, at 8 AM PST.
Topics
> #OpenJournalSystems (#OJS) / #OpenMonographPress (OMP) / #OpenPreprintSystems (OPS) version 3.5.0 preview and release timeline
> Typesetting workflow
> Tasks and Discussions
> Receiving emails in OJS
> Breaking the upload / download pattern with #WebDAV
> Pre-filling metadata automatically with #Grobid
Registration: https://pkp.sfu.ca/2024/11/28/pkp-software-development-update-registration-for-december-16-2024/
⚙️ You are invited to PKP's next Software Development Update webinar!
December 16 2024, 8 AM PST
Topics
* #OpenJournalSystems (#OJS) / #OpenMonographPress (OMP) / #OpenPreprintSystems (OPS) v3.5.0 preview and release timeline
* Typesetting workflow
* Tasks / Discussions
* Receiving emails in OJS
* Breaking upload / download pattern with #WebDAV
* Pre-filling #metadata automatically with #Grobid
Details and registration:
https://pkp.sfu.ca/2024/11/28/pkp-software-development-update-registration-for-december-16-2024/
Hope to meet you there!
@osma @storytracer Hi-just found this old thread - we're just working on a #referenceextraction & #evaluation workflow involving #LLMs to measure their performance using a hand-annotated dataset of older scholarly articles with #footnotes . Untrained #GROBID performs very badly but that does not mean that it will when properly trained with a good dataset.
Do you want to run the #GROBID PDF-to-#TEI conversion library/server with #Apptainer, for example for #ReferenceExtraction? There was a problem converting the #Docker image, but here's how to solve the problem: https://github.com/kermitt2/grobid/issues/1150#issuecomment-2350942263
Curious surprise!
Grobid has started using LaTeXML for processing LaTeX inputs (I think just recently), as part of its TEI-based pipeline.
Details at:
https://grobid.readthedocs.io/en/latest/Principles/
Has anyone used large language models for extracting (#bibliographic style, e.g. #DublinCore) #metadata from fulltext (PDF) documents? I tried this with a fine-tuned #OpenAI #GPT3 Curie model and the results were outrageously good at least for doctoral theses. Much better than traditional NLP methods like #GROBID.
[#BeautifulSoup #Pandas] Parsing TEI XML documents [from #grobid] with Python | Data, code and science https://komax.github.io/blog/text/python/xml/parsing_tei_xml_python/
@andrewjbtw #GROBID for academic #PDF, to get content out of PDF! https://grobid.readthedocs.io/
