There are some great posts on combining NER with Classification. I have made good progress, but would like to make sure I am doing this correctly.
I have an NER model called lpo_ner_model. I would like to use the entities this model can identify for training the classification model.
What does this command do?
prodigy textcat.teach lpo_cat lpo_ner_model data/comments.jsonl --label labels.txt
How is it using the NER model to teach the classification model?
Is it the best way to use the NER Model to start categorizing the data in comments.jsonl?
Hi! In spaCy, the
textcat_multilabel are separate, so the textcat predictions don't depend on the named entities, and vice versa. However, the textcat component will still take the whole text into account so the words mentioned will have an impact on the predictions – but it won't rely on the entities and entity labels for that.
So you can collect your NER and textcat annotations in separate datasets and then use
prodigy train with
--ner pointing at your NER dataset(s) and
--textcat-multilabel pointing at your textcat datasets. This will train a model containing both components, trained on the respective data.
Thankyou this makes sense.
What is the best approach for trying to use the NER model predictions to train the Classification model?
I found this reference, is this basically the approach to use?
Using predictions from preceding components V3.1](https://spacy.io/usage/training#annotating-components)
Sorry if these are elementary questions, I am still learning what the out of box capabilities are. I love the annotation interface, not super familiar with spaCy yet.
In theory, you could build your own text classifier implementation that uses the named entities as features but this is likely overkill I'm really not sure that's necessary and will actually give you an advantage of just training the default text classifier separately. The text classifier will still get to see the same texts including the entity mentions, so it can still take this information into account without the predicted entities being explicit features of the model.
My data is human entered cause and effect data. I am trying to extract cause, subject and effect from this data.
The NER model seems to be pretty good at extracting entities from these comments, however, since the comments are from humans there are lots of synonyms for the same cause, subject, and effect. It may be that I just need to generate a synonym list rather than a classification model.
I can imagine that a classification model may have trouble trying to figure out cause and effect when different nouns are used sometimes as a cause and sometimes as an effect, and the comments are very terse in many cases.
Thanks for your help, got lots of things to try!
Hi, the same problem here. I also want to use predicted entities for text classification. Have you found any solution?
thanks for the reply. will check it.