Path to sense disambiguation for a new NER model

My NER model will have spans that apply to multiple NER labels. For example, 'grill' could be labelled as a 'cooking product' or as a 'car part'. This is similar to the "Michael Jordan" or "Washington" problem.

Is there a document that discusses moving towards sense disambiguation from a standard NER model? Is Entity Linking the right move for this situation?

Hi! It sounds like perhaps your NER challenge isn't really a typical "entity recognition" challenge. It might be that the NER approach is still working for you, but just for reference I also want to point you to the new spancat implementation in spaCy, and the Prodigy docs here:

Whether Entity Linking is appropriate, kind of depends on that as well. Entity Linking typically means resolving an ambiguous mention to a unique identifier, and that identifier typically refers to a real-world object/person. There is not a single unique "car part" in the world, so WSD works slightly differently. Though I do think that you could use similar algorithms for EL as for WSD.

Here's a different thought: I don't know how well that fits your data/use-case, but if you have a limited number of senses/domains, you could also work with a textcat component? If you've found the mention "grill" in a sentence that is classified as being in the domain "car" rather than "cooking", then your grill will be further defined by that class/domain.