Using SpanCategorizer without suggester

[thiippal]

Hi!

If I've understood correctly, the SpanCategorizer learns to suggest Spans and to predict their label.

However, if I train a SpanCategorizer, can I disable the Suggester and use the component only to predict labels for spans (e.g. sentences in a Doc)?

[thomashacker]

Hello!
you can't really disable the Suggester but instead create your own custom one (e.g. a suggester that suggests all sentences in a Doc).
We currently only have the ngram suggester which you could use as a starting point for building your own function.

[thiippal]

Does the first option involve setting the parameter gold_preproc = true under the corpora configuration?

If I set this parameter to true, I get a different error:

ValueError: [E143] Labels for component 'spancat' not initialized. This can be fixed by calling add_label, or by providing a representative batch of examples to the component's `initialize` method.

This seems strange to me, because I just checked that my training data contains the spans and their labels.

[adrianeboyd]

No, gold_preproc = true is just related to gold tokenization, not other annotation layers. For this you'd need a custom corpus reader that copies token.is_sent_start from the reference docs to the predicted docs. It would look a bit like this, just with pred = nlp.make_doc(gold.text) and token.is_sent_start instead of doc.ents:

https://github.com/explosion/projects/blob/20fd27af5122d87151e3df6277a1d6c2c3267412/tutorials/rel_component/scripts/custom_functions.py

[thiippal]

Thanks for the tip @adrianeboyd, that was easy!

My next problem: spaCy raises an IndexError when beginning to train:

File "/Users/thiippal/nlp/lib/python3.8/site-packages/thinc/layers/reduce_last.py", line 17, in forward
    Y = cast(OutT, Xr.dataXd[ends]) # type: ignore
IndexError: index 4507 is out of bounds for axis 0 with size 4487

Any ideas on how I should start debugging this?

[adrianeboyd]

It's hard to know for sure, but my first guess would be that something is wrong with the suggested spans Ragged. Try doing something really simple like suggesting the first word or first sentence in each doc and then expand to more of the annotation as smaller steps work. The suggester functions and Ragged can be a little tricky, especially around empty docs.

[thiippal]

Thanks for the help @adrianeboyd, I managed to resolve the IndexError – my lengths_array for the Ragged object did not have the correct shape. I managed to fix this by tinkering with @thomashacker's spancat.py to examine the dtypes! 🔍