`sherlock/feature_converters` directory is in a confusing place or confusingly named

[GabrielKP]

The concept of InputFeatures is dependent on the representations and functions used in the transformers package.

With the support of AllenNLP (#8) there will be an additional set of something which is equivalent to InputFeatures but is based on AllenNLP.

Neither the current positioning nor the name of the directory sherlock/feature_converters and its contents reflect that they are transformers-only.

I propose to change this, but do not exactly know how. Some ideas (and my subjective feeling about them):

• move feature_converters into annotators folder (this seems cumbersome later on, e.g. if somebody wants to use them)
• create a generic featuer_converter class, register the current feature converter and the allennlp there and call the current ones accordingly somethin with "transformers" (potential overengineer, unless we want to implement another library, this seems related #14)
• rename sherlock/feature_converters into something like sherlock/feature_converters_transformer, create sherlock/feature_converters_allennlp and call it a day (this is not clean)

@leonhardhennig Opinions?

[GabrielKP]

I also could implement the allennlp support without touching sherlock/feature_converters and handle all the dataconversion in the annotator itself. This would ensure (a) backward compability and (b) ease of implementation. Still, the confusion about feature_converter would remain too.

What is the usecase of a separate feature_converter?

[GabrielKP]

This issue is solved, feature_converter now relies on the framwork parameter and thus is not only dependent on transformers anymore