Update neg sampling docstring

[suamin]

Hi, thanks for the example in negative sampling docstring. Just a minor change to it.

[dfdazac]

Hi, thanks for contributing! The current example is fine though:

[[0, 3],
 [5, 3],
 [4, 6],
 [1, 7]]

It means that in the first row, the second entity that was originally 1, got replaced randomly by 3, which can also happen. Please let me know if this is not clear.

[suamin]

Hi, thank you for the clarification.

It might be that I misinterpreted this bit: they are then replaced with a random index from a row other than the row to which they belong that the selection of new indices must be from the corrupted indices i.e. [1, 2, 5, 6]. Following this with example, the updated matrix used [1, 5, 6] but not 2 so I thought it was mistake. Sorry for misunderstanding.

In case, it is helpful to illustrate the source of confusion, step-by-step:

# original
        [[0, 1],
         [2, 3],
         [4, 5],
         [6, 7]]

selected positions

        [[0, -],
         [-, 3],
         [4, -],
         [-, 7]]
selected = [1, 2, 5, 6]

now, we will replace them from the selected following the constraint that they shall not belong to the original row, so possible variations could be:

1.
        [[0, 6],
         [5, 3],
         [4, 2],
         [1, 7]]
2.
        [[0, 6],
         [5, 3],
         [4, 1],
         [2, 7]]
3.
        [[0, 2],
         [1, 3],
         [4, 6],
         [5, 7]]
etc.

the documentation confused me in thinking that we should sample from selected.