[MRG] Topic coherence update 3

[devashishd12]

Changes:

• Added backtracking dictionary for storing the context vector for (w_prime or w_star, w) pair. Each such tuple is mapped one to one to a context vector. Earlier each context vector was calculated every time. Changed window size to 110.
• Added c_uci coherence measure.
• Added c_npmi coherence measure.
• Added window_size parameter to CoherenceModel init.

[devashishd12]

@tmylk should I add the benchmark testing notebooks for 20NG and Movies dataset too?

[devashishd12]

Correcting the tests.

[devashishd12]

@tmylk I've added c_uci, c_npmi coherence measures and window_size parameter to CoherenceModel init. This pr also addresses issue #765.

[devashishd12]

@tmylk I've added the benchmark testing notebook on movies dataset.

[tmylk]

what is the point of checking if isinstance(model.id2word, FakeDict): above? why is it note enough to check for None?

[devashishd12]

None doesn't work here actually. When no word->id mapping is provided while creating an LdaModel this line is called which returns a FakeDict from here. So this code:

tm1 = LdaModel(corpus=corpus, num_topics=2)
if tm1.id2word is None:
    print 'aye'
else:
    print 'naye'

actually prints naye but if I change it to isinstance(tm1.id2word, FakeDict) it outputs correctly.
Am I correct here?

[devashishd12]

@tmylk I've addressed your initial comments. I hope I've addressed them correctly.

[devashishd12]

@tmylk should I change all Args to Parameters?

[tmylk]

Explain how the indirect confirmation measures work and why they are useful similar to the competing car brands explanation in the paper.

[tmylk]

same as above

[devashishd12]

@tmylk I've addressed your comments.

[tmylk]

Merged in 6f53b31

[devashishd12]

Linking to #750 and #710.