Training a spaCy SpanCat Model to Annotate in Texts more quickly in Prodigy | SpanCat 03

[@GrahamAndersonis]

Thanks! What video do you recommend after this spancat 03? Feels like there is more to know here

[@python-programming]

Thank you so much for this!! 🎉

[@python-programming]

There shoud be an 04. It looks like it never uploaded to youtube. I will be sure to fix that soon!

[@BSP77]

This is a wonderful series! I look forward to the next video, thank you!

[@jsr7599]

Does part 4 not exist because you ran into issues with it not predicting anything? Confused on why there’s no docs online about finishing this spancat process, but a lot of posts online about it not predicting correctly

[@python-programming]

Thanks for the comment! No, it works fine. I lost the video footage and I need to re-record it. I'm trying to get it done ASAP. I use it on a lot of projects and spancat does work well. You need more training data for it, typically.

[@davidrussell9662]

Please do. I was looking forward to it@@python-programming

[@JonasWindey]

Bump! Been waiting for the next video for a while now :)

[@GrahamAndersonis]

For spancat, is it better to treat a section of sentences as single doc for tagging, or is it better do tag sentence by sentence. In my case , there are sentences, external doc references, tables, figures, code, and other stuff that describe a section.

[@python-programming]

It depends on how much context is needed to accurately predict a span. If it relies on larger context, go larger (up to 250 tokens or so).

[@GrahamAndersonis]

@@python-programming if the token size is larger than 250, do you simply make section 1a, and section 1b? In my case I have some control over where I divide the section.

[@python-programming]

spaCy will automatically handle the chunking of the text for you when you run the model. This is just for training the model. If you have some control, then yes, just find a natural breaking point and separate there (such as a paragraph)@@GrahamAndersonis

[@GrahamAndersonis]

@@python-programming for future ref, do you consult and/or have a discord?

[@python-programming]

@@GrahamAndersonis I do! You can reach me via the form on my site: wjbmattingly.com/

[@pcxxy]

super helpful video, looking forward to video 04 keep up to great work!

[@dariaglushkina2036]

Hello! Thanks a lot for your tutorials! Could you please make a new video on how to correctly create and modify config files? I've tried to train a spancat model upon en_core_web_lg and en_core_web_trf models (I want to have both ner and spancat), but it did not work because of some errors in config files. I think this topic will be very useful also for others. Thank you again.

[@python-programming]

Absolutely! I will try to do that soon! Thanks for the idea!

[@paulmiller591]

Very helpful Cheers!

[@python-programming]

I'm so happy to hear it helped!

[@shawnmarcy4413]

🎉🎉🎉

[@judithnathanail3742]

Enjoyed the video. Would love to see a video using the Prodigy pdf plugin - Prodigy_pdf - to annotate some pdfs in Prodigy and then train a model in Spacy (or something else); followed by applying the created model to some unknown pdfs. Lots of humanities materials are pdfs. There is a nice video on annotating papers (kzbin.info/www/bejne/qKjcq5hqbtOYbqs) but to be useful, we need to use the annotated output to train a model.