How to create a nlp object with a dict instead of a string?

[lingvisa]

The typical way of creating a nlp and doc object is:

nlp = spacy.load("en_core_web_sm")
doc = nlp("Apple is looking at buying U.K. startup for $1 billion")

Is it possible that my input is not a str, but a dict like:

input = {'text': 'text', 'summary': 'summary', 'id': '001', 'type': 'News'}

Is it possible that a doc can created like:

doc = nlp(input)

The reason is that, my custom component needs to access other fields of the input, beyond the usual 'text' field. Please advise ideas if this is possible.

[polm]

You can't exactly pass a dict as input, but in 3.2 we added a feature for exactly this kind of use case such that you can use a Doc as input to a pipeline (pipeline = nlp object). So what you can do is have a minimal pipeline that does maybe just tokenization, then add your data to the Doc as underscore attributes, and then pass that to another pipeline with your custom components.

[lingvisa]

Good to know the new features. What about the component_cfg = {} argument? Can I access this argument from a nlp object?

[lingvisa]

Yes, the contents of component_cfg are passed to the respective components as keyword arguments.

Please tell how to access the cocmponent_cfg's contents in a component? Thanks.

[polm]

Here's an example of using component_cfg.

import spacy
from spacy.tokens import Doc
from spacy.language import Language

nlp = spacy.blank("en")
Doc.set_extension("bloop", default=None)

@Language.component("bloop")
def bloop(doc, msg=None):
    if msg is not None:
        doc._.bloop = msg
        print(msg)
    return doc

nlp.add_pipe("bloop")

# prints nothing
doc = nlp("I like peanut butter")
# is None
print(doc._.bloop)
# prints "hello there"
doc = nlp("I like sushi", component_cfg={"bloop":{"msg":"hello there"}})
# is "hello there"
print(doc._.bloop)

The top level keys in component_cfg correspond to component names, and the contents are passed as keyword arguments to the component.

[lingvisa]

If bloop is a component class, then it should be like this :

 @Language.factory("bloop")
def create_bloop(nlp:Language, name:str):
    return Bloop(nlp)

class Bloop(object):
   def __init__(nlp:Language):
      ...

  def __call__(self, doc, user_args):
     print(user_args)

Then I can do:
doc = nlp("I like sushi", component_cfg={"bloop":{"user_args":{"test1":"hello", "test2":{"test3":"hello2"}}}})

This looks like exactly what I need. I tested and it worked! Thanks. I don't have to upgrade to spacy 3.2 for the time being.

[polm]

It might be helpful to realize the that the component_cfg is just the same as what you would put in a config file - check the docs on how the config is loaded. The values in the config are basically passed in the same way (though at the point the component is created, not on individual calls).

[lingvisa]

That's true. Thanks.