2
$\begingroup$

We know that we can extract entities from a sentence using named entity recognition, but what if the sentence contains only an entity and no other context?

For example, we can use CRF for the following sentence:

Conditional random fields (CRFs) are a class of statistical modeling method often applied in pattern recognition and machine learning and used for structured prediction.

to extract Conditional random fields, statistical modeling, pattern recognition and etc from the sentence if we train our model with labeled data to do that.

But what if the sentences look like the following three?

Sentence 1: Conditional random fields
Sentence 2: statistical modeling
Sentence 3: Donald J. Trump

Can we use CRF again to recognize them respectively as "algorithm", "algorithm", "politician"? If not, how can we tackle that?

Improve this question
$\endgroup$
1
  • $\begingroup$ What you're talking about is actually topic classification, yes? In that case, you simply train your model to predict short snippets of text to predict a topic from a list. $\endgroup$
    – Estimate the estimators
    Commented Apr 15, 2022 at 1:20

2 Answers 2

Reset to default
1
$\begingroup$

This version of the problem has a much simpler structure. This is no longer a structured prediction problem; it's just classification. (Part of the value of a CRF is that it labels whole spans; you see where each name begins and ends. Now, the whole thing is the name.)

A decent starting point would be any typical classifier. Feed in a feature representation of the text to your model, and you can classify it.

Improve this answer
$\endgroup$
1
  • $\begingroup$ This. Just look into fine-grained topic classification or the like. You have to decide how many topics you want. I'd look at the DeepType paper. $\endgroup$
    – Estimate the estimators
    Commented Apr 15, 2022 at 1:22
0
$\begingroup$

This study shows that

We find that context does influence predictions, but the main factor driving high performance is learning the name tokens themselves.

Then if we train a BiLSTM-CRF or BERT for such sake with not only pure named entities but also entities with context the performance would be good enough.

One more option would be to extract more context from the data source and use that as a separate condition, as stated in that paper:

context-only systems are sometimes correct even when humans fail to recognize the entity type from the context
Our results suggest that designing models that explicitly operate over representations of local inputs and context, respectively, may in some cases improve performance

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.