Is it possible to supply the EntityLinker component with a Child Class of KnowledgeBase?

[coltonflowers1]

I am trying to subclass the KnowledgeBase class, and then provide that to the EntityLinker component. However, it looks like the EntityLinker component's from_disk method contains the following:

deserialize["kb"] = lambda p: self.kb.from_disk(p)

I'm used to having from_disk be a class method, but here it is called by the object, so it's assuming that self.kb has already been initialized by the time this function gets called. However, the only time I see self.kb getting initialized at inference time(i.e. not in the initialize method) is in the __init__ method of EntityLinker in this line:

 self.kb = empty_kb(entity_vector_length)(self.vocab)

empty_kb actually returns an InMemoryLookupKB object, which leads me to believe that EntityLinker is using this subclass's from_disk method and would not use the KnowledgeBase subclass that was supplied at training time using a custom kb_loader method.

Am I missing something here? Is it the EntityLinker component not meant to be supplied with a kb_loader callable that returns something other than an InMemoryLookupKB object?

[svlandeg]

Hi @coltonflowers1. Thanks for the detailed explanation - your analysis is pretty much correct and is leaving me wondering why I implemented this as such. It's originating from this PR which actually has an explicit test for the KB being the "standard" type after IO. At the time this was KnowledgeBase, since 3.5 it would be InMemoryLookupKB as you mention, and as you can see in the current version of this test here: https://github.com/explosion/spaCy/blob/master/spacy/tests/serialize/test_serialize_kb.py#L142

This doesn't seem right - I'll look into it more and get back to you.

[svlandeg]

We're working on a fix for this here: #12320

[svlandeg]

Basically from 3.5.1 onwards, you'll be able to modify the behaviour of what was previously empty_kb() by defining this function in the config:

[components.entity_linker.generate_empty_kb]
    @misc = "kb_test.CustomEmptyKB.v1"

As example usage, you can have a look at the test_serialize_subclassed_kb test here.