Rare class and batch size

[kermitt2]

For ELMo, which is using reduced batch size because of memory constrains, it might be necessary to review how the batch are created to ensure that rare classes are well represented in each batch, with automatic over-sampling techniques for instance.

See issue #7 (although ontonotes work finally fine without sampling, we could have more extremely unbalanced training data)