Hyperparameter optimization for NER

[FrancescoCasalegno]

In #356 we started seeing that we can play with hyperparameters to reduce the runtime while having high accuracy.

Once #321 is resolved, we can start looking into hyperparameter optimization:

• [ ] Can we further reduce the runtime without impacting — in a statistically significant way — on the generalization accuracy?
• [ ] Can we further improve the generalization accuracy of our models — in a statistically significant way — by modifying some hyperparameter?

[pafonta]

Dependency to #321

We also discussed yesterday that this task should happen after #321. Indeed, we need to know what is happening between the evaluation on the dev set and the evaluation on the test set.