pevnak

Thanks. This is nice.

I like the `decompose` a lot. It is nicer and cleaner over what I have written. The indexing is also very nice. To which part you refer as not being...

This is crystal clear answer. I will work on relevant PRs.

This did the job for me ```julia function Base.intersect(s1::AbstractIndices, s2::AbstractIndices) length(s1) > length(s2) ? filter!(in(s2), copy(s1)) : filter!(in(s1), copy(s2)) end ```

I cannot judge that. I am usually concerned with speed.

Since there is a stable_sort and sort, would it make sense to have something similar? I apologize for abusing the comments, but the difference in speed I am getting in...

But it can be narrowed for types where this will be performant.

I was not. I thought that they are already default. Thanks a lot, I will try it. Tomas

I guess that dedup in NGramMatrices will offer the highest benefit. Multiplication of OneHot is essentially a copying, and Dense matrices should not contain that many duplicates (although they might).

U know we have talked about it, I just did not know, if it fits there an where? Should I move it?