1
$\begingroup$

Removing stop words can significantly speed up named entity recognition (NER) modeling by reducing the number of tokens in a document.

Are stop words critical to get correct NER performance?

$\endgroup$
1
$\begingroup$

It depends how you recognise the entities.

If you do a simple gazetteer lookup, then it could be faster, as you have fewer tokens to deal with.

However, if you use contextual rules, then stop words might be vital to identify certain contexts, so by removing stop words you lose information about the entity's environment. For example, if [work] at {organisation} is a rule you use to identify companies etc, then this wouldn't work if you take out the at.

You will also have problems if the stop words are part of an entity, eg the town Stoke-on-Trent. If hyphens are used to split tokens, then you won't be able to recognise it if the on is discarded.

In general I think stop words should mostly be kept; apart from information retrieval where they are not all that useful in an inverted index you will always lose something. If stop words were really pointless, then they would have been discarded from languages a long time ago. In the old days they were a useful way of reducing the demand on computational resources without losing too much, but nowadays I would say this is not really a problem anymore.

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.