Add a new indirect confirmation measure based on word2vec similarity

[macks22]

Description

As introduced in [1], topic coherence can be computed using word2vec similarity of terms. This is roughly compatible with the coherence evaluation framework posited by [2] and currently implemented in gensim. The probability estimation phase will involve training a word2vec model on the given corpus, and the confirmation measure will use the average cosine similarity between context vectors from the word2vec model. [1] did not analyze which segmentation strategy is optimal, but a reasonable guess is to use the same method as "c_v" coherence, since that method also uses context vectors in a semantic space.

The implementation should include the ability to pass in a pre-trained word2vec model or path to the same.

Steps/Code/Corpus to Reproduce

I'm envisioning an API something like the following:

cm = CoherenceModel(corpus=corpus, coherence="c_w2v", embeddings="./path-to-model.bin.gz", embeddings_kwargs=dict(binary=True))
# OR
model = gensim.models.KeyedVectors.load_word2vec_format('./GoogleNews-vectors-negative300.bin.gz', binary=True)
cm = CoherenceModel(corpus=corpus, coherence="c_w2v", embeddings=model)
cm.get_coherence()

Expected Results

Actual Results

Versions

This should be supported under all versions.

References

[1] D. O’Callaghan, D. Greene, J. Carthy, and P. Cunningham, “An analysis of the coherence of descriptors in topic modeling,” Expert Systems with Applications, vol. 42, no. 13, pp. 5645–5657, Aug. 2015.
[2] M. Röder, A. Both, and A. Hinneburg, “Exploring the Space of Topic Coherence Measures,” 2015, pp. 399–408.

[macks22]

I'm working on an implementation of this, but it will build on #1349.