Add top_features functions

[jamesalster]

Suggested addition of functions for top_features(), for Corpus, DocumentTermMatrix, Document and lexicons. I've provided docs and tests. It's my first time contributing to this package, so let me know if I've missed anything.

One question is if/how we sort terms with tied counts: I've gone for no sorting, using julia's default sort!(OrderedDict(...)) since it seemed simplest and I couldn't think of an obvious solution.

[jamesalster]

I enforced alphabetic sorting for ties after the online tests were having problems reproducing the order (it was possibly run-or-machine specific)

[rssdev10]

Computational complexity of this method is very high. Consider something like this one:

top_features(D::DocumentTermMatrix, n::Int) = first.(top_features(D, Val(n)))

function top_features(D::DocumentTermMatrix, ::Val{N}) where {N}
    counts = @view(sum(D.dtm; dims=1)[1, :])
    n = min(N, length(counts))
    idx = partialsortperm(counts, 1:n; rev=true)
    collect(zip(D.terms[idx], counts[idx]))
end

Here, we avoid exceeding the memory allocation and the need to sort everything, opting instead to take only n-elements.

[rssdev10]

I'm not sure this is needed.

[rssdev10]

I'm thinking about the name top_features. Don't you think top_terms or most_frequent_terms would be clearer?

[rssdev10]

The comment regarding sort!(sort!(OrderedDict(zip... is related to all 3 places with full sorting.

[jamesalster]

Thank you for the review @rssdev10 ! You're right, I forgot how inefficient the full sort!() could become.

Changes:

• partialsortperm() implemented, with numeric sorting using alphabetic order to break ties, consistent with quanteda
• Method without the n argument dropped, since on reflection it's unncessary
• OrderedDict kept, since I think the order of the terms (keys) being strictly consistent is definitely expected user behaviour. If you feel strongly about the dependency, we could discard the count information and return a string vector?
• Naming was taken from quanteda::topfeatures() but I've changed it to keep it consistent with TextAnalysis.

[rssdev10]

• OrderedDict kept, since I think the order of the terms (keys) being strictly consistent is definitely expected user behaviour. If you feel strongly about the dependency, we could discard the count information and return a string vector?

I think it would be better not to add OrderedDict as a new package dependency. Users who need it can simply use top_terms(...) |> OrderedDict in their own code.

You could use either a Vector{Pair} or a Vector{Tuple} struct as the result instead. See allowable arguments types - https://github.com/JuliaCollections/OrderedCollections.jl/blob/master/src/ordered\_dict.jl\#L61-L68

Also, please add a new line to the end of all files you have modified. GitHub indicates the absence of a free line in places like this one - https://github.com/JuliaText/TextAnalysis.jl/pull/289/changes#diff-c6b4870a7fa201c05bd4ce8bf468b4f092f94fb6d18bc8a913e9e6009b342956R415