Keybert extractor labels all extracted keywords with the same metadata key.

[pedrocassalpacheco]

One of the interesting aspects of keyword extraction with Bert-based models is that you can extract many different keyword "types" on a single execution. The previous implementation allowed for passing a collection of terms to the model and receiving, in return, a collection of terms for each label. The current implementation hardcodes the metadata label and only passes a single string to the model for extraction.

For example, if I create metadata for a historical document, I want to know about people, places, dates, and events. And I would prefer to receive metadata looking like {"people": ["Alexander the Great," "Philip of Macedonia"...], "places": ["ancient Greece," "middle east," "Macedonia"...

I am happy to make the change if we agree that this is a good change in line with the previous implementation.

@kerinin @erichare

[epinzur]

I wasn't aware that KeyBERT can function in this way. I thought it only extracts keywords and doesn't have a way to pull a set of labels like our SpaCy and GliNER transformers function.