CRF for NER with many classes

Question

I'm going to develop named-entity recognition system with many (100+) classes. Assuming that they have roughly equal frequency, what algorithm should perform best? According to my understanding (sadly, far away from ideal) of how CRF works, it should be ok here. But in some sources ( google books ) I found another opinion.

So, is CRF suitable algorithm for NER with huge number of classes?

Answer 1

What kind of classes are we talking about here? Don't be coy :)

You're probably best off using a hierarchical approach:

1. Have a small number of root classes, like "product", "person", "place". First pass figures out which one is which.

2. Then for each of the root classes, have sub-classes like "car product", "electronics product", etc.