Span Finder Suggester

[thomashacker]

This PR adds a new experimental component for learning span boundaries and a custom suggester function for spancat.
It further adds a spaCy project showcasing how to use the SpanFinder component on 3 different datasets (Healthsea, ToxicSpans, Genia) with 2 configurations (tok2vec & transformer). The project also provides the possibility to train spancat with ngram and compare it to SpanFinder with a custom evaluation script that calculates the performance and overall coverage of the suggester functions.

Features

• spaCy project for comparing SpanFinder vs Ngram
• SpanFinder model
• SpanFinder component
• SpanFinder suggester
• Unit tests for component, model and suggester

[adrianeboyd]

Very nice! This is going to be annoying, but I think that we need to avoid the abbreviation sbd because it's used too much for "sentence boundary detection".

[thomashacker]

@adrianeboyd how about instead of "sbd" -> "span_bd"? 😄

[svlandeg]

This is great Edi!

From a practical point, I'm wondering whether we need all the different config files in all the different subdirectories? It's great to support various different datasets, but what if we could make the decision very early on:

• You process one of the three available datasets, with one of the three specific commands (cf below)
• The processed .spacy output is stored in the same subdir of the project, regardless of what dataset had been processed
• Now, all consequent steps don't have to bother about where the data came from originally

Is that feasible? Or are there things you're doing different in the scripts, depending on what dataset it was (besides preprocessing)

[thomashacker]

Alright, I've renamed every reference to SpanFinder 😄 I've rebuilt the set_annotation() logic so that the SpanFinder component directly produces spans and saves them to doc.spans["span_finder_candidates"] instead of having an additional token attribute. The suggester uses doc.spans["span_finder_candidates"] to produce the Ragged for the spancat.

I've also adjusted the spaCy project, added more commands, and reduced the number of configs, so that now there is only one config for all datasets.

[adrianeboyd]

As the next minor adjustment, I think that span_finder_candidates should be a configurable setting.

[adrianeboyd]

Can you rename:

• candidates_key -> predicted_key
• reference_key -> training_key

In the project, I'd also like to have paths.spans_key as vars.spans_key or something else instead, since paths seems kind of confusing.

[adrianeboyd]

I'm not sure I found everything, but can you do another pass through the docs, project, and tests?

[adrianeboyd]

The remaining issues:

• keep candidates_key as the key name for the spans suggester
• one more pass through the docs and project for the other key renaming
• a full test of the project with various option combinations after the renaming

[thomashacker]

Updated the key in the suggester and looked over the docs and the project. I also checked the spaCy project and ran the workflows with different configurations.