Make docs clearer on alpha parameter in LDA model

[xh2]

Summary

Modified docstring on alpha in LDA model around 'symmetric' and 'asymmetric' options

Motivation

• When I first read the doc, I saw by default alpha='symmetric', but in the docs, only 'asymmetric' and 'auto' are listed as acceptable strings, so my first instinct was it was a typo. To be clearer, especially to new users, I think we should spell out all possible options, including the default 'symmetric' in the doc
• The current 'asymmetric' description doesn't seem to fit the actual formula in the code

[piskvorky]

You're right, thanks for the fix!

The current 'asymmetric' description doesn't seem to fit the actual formula in the code

Which is correct – the code or the documentation? I don't remember the motivation or original paper for the asymmetric prior any more, unfortunately.

[xh2]

Which is correct – the code or the documentation? I don't remember the motivation or original paper for the asymmetric prior any more, unfortunately.

The paper by Hoffman only used symmetric priors, while citing Wallach, Hanna & Mimno, David & Mccallum, Andrew. (2009). Rethinking LDA: Why priors matter. NIPS. 23. 1973-1981. for asymmetric priors. But I couldn't find either formula in that paper either, so not really sure what it is supposed to be.

The existing doc on asymmetric alpha, however, is actually suggesting a symmetric distribution instead? No? That was also what confused me and made me think it was a typo.

[piskvorky]

I don't think so – with asymmetric, different topics get a different alpha. With symmatric, all topics get the same alpha.

[xh2]

yes agree! but isn't 1.0 / topicno (current docstring description for asymmetric) uniform? (the actual code is indeed asymmetric). unless here topicno means index not number of topics?

[piskvorky]

Yes. For sure that's meant as "index of the topic", not "total number of topics". Another piece of documentation that could use a fix!