Improving Lesk Overlaps

[alvations]

This is more of a performance than a theoretical issue. In theory, it's implemented as they are presented with their respective papers, simple overlaps.

Going after the state-of-art will mean that the implementation is not as represented in the paper. Supervised learning part is going to be a long shot since feature extraction is another headache.

Possibly, improving the overlaps should be a better move for the current code.

1. Look at the normalization https://github.com/alvations/pywsd/blob/master/pywsd/lesk.py#L41
2. Think about the effects of lemmatized overlaps vs unlemmatized overlaps. (currently, it's lemmatized overlap by default)
3. Handling tie-breakers when #overlaps is the same.
4. Fallback on MFS (that involves extracting MFS from annotated corpus)