It has at least one bug and I am not sure why you think it is efficient.

First the bug. You have /$uniq[$i]/. This can do the wrong thing for at least two reasons. The first is that it will match "and" in the line of "where is the sand" even though the word "and" never appears. The second reason is that you haven't use \Q\E or quotemeta to ensure any metacharacters are escaped. It doesn't take case into consideration either, but I am not sure any of the solutions do (including my own).

Another potential bug is the fact that you don't consider that in the line "hello world goodbye cruel world" that each pairing of world should double and not be counted once. I am not sure if this is or isn't a requirement but it is a bug in one intepretation.

As to why it isn't efficient. This can be done in a single pass and yet your solution has you going through every line in the data (N^2 - N) / 2 times (where N = # of unique words). In other words, you check every single pairing of unique word against every single line even if the word doesn't appear on that line.