Sort for similarity with training data

[dlangenk]

In the annotation candidates step it would be nice if you could see candidates similar to the training data at the top. This is especially useful if prediction results are bad.

This could be achieved in different ways:

1. Just randomly select one/few training samples and use the already implemented similarity search
2. Show a sidescrolling list of all training samples at the top and allow to pin them
3. Use clustering to find possible classes and then do 1) for each cluster
4. Use techniques similar to active learning, i.e. search for training samples that are very different from each other and then do 1)
5. Something even more sophisticated

[mzur]

I like the idea if it is implemented as optional feature.