Update ent_id of span and tokens

[NixBiks]

I have a custom component where I need to update Span.ent_id_ and Token.ent_id_. How is that possible?

I've tried something like

for t in span:
    t.ent_id_ = "NEW_LABEL"

which gives me the desired results. BUT when I then do doc.ents = [span] then it reset the ent_id_ to what it was before. Anyone know how to fix this?

[NixBiks]

I've ended up creating a new span

span = Span(doc, span.start, span.end, span_id="NEW_LABEL", label=span.label_)
doc.ents = [span]

which seems to work as expected

[polm]

Thanks for mentioning that, and your workaround. That's definitely weird behavior, we'll think about improving it.

[adrianeboyd]

I looked into this and although it doesn't look like there's directly a bug that we can address, this is definitely confusing because some span properties are separate copies of the token-level annotation and some are a view of the underlying token annotation.

For getting or setting Doc.ents there is currently this mapping between Token and Span attributes:

• Span.label <-> Token.ent_type
• Span.kb_id <-> Token.ent_kb_id
• Span.id <-> Token.ent_id

The Span.id <-> Token.ent_id mapping is new as of the addition of SpanRuler in v3.3.1.

Span already had the existing Span.ent_id property that is just a view of the token annotation rather than an attribute of the span itself, which is also why you can't set it on the span.

For the code above, instead of setting token values within the span (that aren't reflected in the span) or having to create a new Span, you can set Span.id:

span = doc.ents[0]
span.id_ = "NEW_LABEL"
doc.ents = [span]

Because the Span.id <-> Token.ent_id mapping was added after v3.0.0, there's a backoff behavior that setting doc.ents only overrides the current Token.ent_id if the Span.id is set (not 0 or ""). This backoff would be removed in v4.

We'd like to make this more consistent in v4, feedback is welcome in #11328.

[NixBiks]

Glad I'm not the only one who got a bit confused here. But thank you for the thorugh explaination - I see how this arised over time.

[NixBiks]

But pyright does not like this

span.id_ = "NEW_LABEL"

I'm getting Literal["NEW_LABEL"] is incompatible with property - I guess id_ shouldn't be typed as a property as it was done in #11119 ?

[adrianeboyd]

Hmm, the types are a bit inconsistent here and that type does look wrong for a property with a setter. We'll take a look!

[NixBiks]

Let me know if there is an issue I can point to in my codebase 😸