Add support for self-training meta-estimator

[desilinguist]

Nice way to take advantage of unlabeled data (based on Yarowski's old self-supervised WSD paper). See more here.