Phrase/Token labeling

Question

Looking for suggestions on how to define the following NLP problem and different ways in which it can be modeled to leverage machine learning. I believe there are multiple ways to model this problem. Deep-learning-based suggestions also work as there is a good amount of data is available for training.

Will evaluate different approaches for the given dataset. Please share relevant papers, blogs, or GitHub repos. Thanks!

Input: Given a sentence S having words W1 to W10.

S = W1 W2 W3 W4 W5 W6 W7 W8 W9 W10

The sentence has some syntactic and semantic patterns, but it is not exactly freely written natural language but it's in English. These are words, can be punctuation

Output: should be something like this.

Label1 - W4

Label2 - W3

Label3 - [W2 W1] continuous // semantically related. Means words [W2 W1] in-order are assigned a Label3. Also okay with solutions that don't output in-order.

Label4 - [W6 W8]

Label5- W10

Noise- W7, W9. Means words W7 and W9 independently are assigned a Label3.

Label7- W5

Need to solve the problem. Looking for research/thoughts on how this problem can be defined in different ways to exploit different patterns in the structure of sentences. Looking for similar tasks which are already defined in NLP such as token labeling, parsing which can be used.

Would be really helpful to get the suggestions to the latest research on solving/defining this problem.

Answer 1

This looks like a sequence labelling problem, the most common such problem in NLP being Named Entity Recognition (NER).

You'll find a lot of libraries and tutorials about NER. It can be done with Conditional Random Fields but there are also neural methods nowadays.

Assuming your problem is not about standard entities (like persons names, organizations, locations), you'll need to train a custom NER-like model. To do that you will need a large amount of data annotated for your specific task.