New Ranker with Word Mover's Distance

[Lolik111]

Ranker based on Word Mover's Distance that uses word embedding for calculating a distance between two documents (see Kusner, Matt, et al. "From word embeddings to document distances." 2015 - http://mkusner.github.io/publications/WMD.pdf).
config.toml example:

[ranker]
method = "wmd-base"
mode = "rwmd"
distance-function = "cosine"

Available modes:

• emd - more precise according to paper above but very slow
• rwmd - relaxed wmd - faster and close to emd in precision
• wcd - word centroid discance - fast, but not precise
• prefetch-prune - combination of all modes that allow to reduce possible candidates before running emd

Available distance functions:

• cosine - uses cosine distance between two numeric vectors
• l2diff - uses Euclidean distance (L2-norm) between two numeric vectors