the cds formula (context distribution smoothing) is wrong

[xiaoouwang]

The original formula is

However, the code implementation is reversed, see below (the sum is above)

self.d_alpha = np.sum(np.power([self.terms_counts[c] for c in self.terms_counts], self.cds_alpha))
bar.update()
self.terms_counts_cds_powered = {word: self.d_alpha / np.power(self.terms_counts[word], self.cds_alpha) for word in self.terms_counts}

I've made a pull request.