NEL feature creation should not be metadata-dependent

[kdutia]

At the moment, when features are created for NEL, they rely on label and description columns already being present in the data. This is not ideal if you want to use the same pairs entity mentions and records that have been annotated as linking, but titles or descriptions are updated.

This should likely be in the form of a pipeline element that can produce the dataframe accepted by NELFeatureGenerator from the data entity_mention, source_record, target_record.

[kdutia]

added a notebook to be able to update descriptions from an old training data file. Solves this issue, although not neatly.