Training semantic role labeling model (with version 3)?

[egumasa]

I am finding a way to train a semantic role labeling (and predicate disambiguation) using Ontonote5 dataset. I encountered this post from the past, and now that SpaCy v3 has more user-friendly pipelines, I believe that the hurdle is much lower.

To create this training project, which project template would you recommend to start with? I am thinking to leverage the already existing template for either dependency parsing or Ner or entity linking, to start my project.

I appreciate any recommendation on how to get this project started. I am imagining that training a model with spacy can be efficient than importing AllenNLP models such as this one, and like this.

I appreciate any suggestions/recommendations! Thank you in advance!

[egumasa]

Could anyone provide their insights on how I might do this? OR is this impossible with the current version of spaCy?

[polm]

This isn't impossible with the current version of spaCy, but it's also not a built-in feature or something we're currently working on specifically. You could try using the recently released SpanCategorizer for it, but the main challenge with semantic role labelling is that it's hard to learn in general, and there may be better specialized architectures or approaches.

[egumasa]

Thank you for your response on this one! Much appreciated. I see, I think I will stay with the current architecture for now (e.g., importing AllenNLP model into space pipeline). Thanks for the pointer to SpanCategorizer. I think this is interesting new feature! I might consider this for other component I am thinking (e.g., modal verb categorization/sense disambiguation, etc.)

Thank you so much for your guidance!