[FEAT] Support NLP text regression

[j-adamczyk]

Is your feature request related to a problem? Please describe.

Currently only token classification and text classification are supported for NLP. However, there are important cases for text regression, for example:

• CTR prediction for advertisements
• sentiment magnitude prediction, e.g. GCP sentiment analysis predicts continous values instead of classes
• ordinal regression for texts, e.g. predicting number of stars from 1 to 5 based on review text

Describe the solution you'd like

Support for text regression, similar to tabular regression, but for NLP models, e.g. checking regression error distribution or train-test degradation for regression metrics.

[noamzbr]

Thanks for the suggestion @j-adamczyk! Any other features you'd suggest for these task types?

[j-adamczyk]

@noamzbr thank you for fast response.

This requires a mix of regression tests and NLP tests.

From tabular quickstart, interesting checks are:

• train-test performance
• regression error distribution
• prediction drift
• simple model comparison

Specifically, tabular regression checks that don't make sense for NLP are weak segments performance (since they are not well defined for NLP), boosting overfit (since NLP does not use boosting typically) and model inference time (which is naturally long for NLP).

From [NLP text classification quickstart]:

• text property outliers
• unknown tokens
• under annotated property segments
• under annotated metadata segments
• text duplicates
• special characters

Also, image regression could also be added in a very similar way (but that is outside the scope of this issue).

[j-adamczyk]

@noamzbr any news on this? As far as I understand, this is mixing 2 existing things together, and no really new code is needed