NER tags are on the tokens, not the words

[AngledLuffa]

When using an English tokenizer with MWT (such as EWT, which treats possessive clitics as MWT), the NER model puts the tags on the entire token, not just the word, and sometimes gets it wrong in situations it shouldn't

MWT tokenizer:

John's headache is bad -> no NER tags on John or 's

This makes John's headache worse -> NER tags on both John and 's

Non-MWT tokenizer:

John's headache is bad -> NER tag on just John, not 's, as expected