Ufdjrw

I am new in using spacy. I want to extract text values from sentences

training_sentence="I want to add a text field having name as new data"
 OR
training_sentence=" add a field and label it as advance data"

So from the above sentence, I want to extract "new data" and "advance data"

For now, I am able to extract entities like "add", "field" and "label" using Custom NER.

But I am unable to extract text values as these value can be anything and I am not sure how to extract it using custom NER in spacy.

I have seen code snippet here of entity relations in the spacy documentation
But don't know to implement it as per my use case.

I can't share the code. Please assist how to tackle this problem

I'm not sure that framing this as a pure named entity recognition problem really makes sense here. Named entities are usually proper nouns and "real world objects" – for example, a person name like "John Doe", an organization name like "Google", or things like diseases or genes, to name examples from a more specific domain. This is also what spaCy's named entity recognizer is optimised for.

In your example, it seems like most of the clues are actually in the syntax, which is something you can usually predict pretty well out-of-the-box. For instance, you're looking for verbs like "add" and "label", and their objects ("text field") or attached prepositional phrases. If you visualize the syntax, e.g. using the displacy module, you'll see that there's a lot of relevant information in the sentence structure that you can extract programmatically:

from spacy import displacy
doc = nlp("I want to add a text field having name as new data")
displacy.serve(doc)

You can also use the rule-based matcher to find trigger tokens like "label" (with the part-of-speech tag VERB) and then check the dependency tree to find the tokens attached to them. For example, if the verb "label" is attached to a preposition "as", you can be pretty sure that the object attached to it is the name of the label. Or you could start at the root of a sentence and iterate over its subtree and check whether it contains tokens or constructions you're interested in.

You might have to experiment a little and you'll probably end up with a bunch of different rules to cover different types of constructions that are common in your data.