Because ($path, $value) is a list, you get the list of submatches (list context). But if you do something like:

if ($row =~ /(.*),(.*)/)
{
  ... 
}
[download]

my $path = $row; # path is the full string by default
if ($row =~ /(.*),(.*)/)
{
  my $left_part = $1;
  my $value = $2;
  # Check if $value is a number and change $path if needed
  ...
}
[download]

The algo I would like to implement :
As I see it the steps are:
1. if the string has commas:
1.a. get the last comma and check if the last substring is a number - if so put it in hash like this: $hash{$path} = $value;
1.b. if the substring after the last comma isn't a number - $hash{$path} = 1;
2. if string has no commas: $hash{$string} = 1;

Although a good start, point 1.b. is unclear: in this case, do you want the whole string stored in $path, or just the part up until the last comma? For now I'm assuming the latter. Anyway, while there may always be "nicer" ways to write things in Perl (Update: and you haven't specified what you meant with "it doesn't look very good"), sometimes a good starting point is a direct translation:

use warnings;
use strict;
use Scalar::Util qw/looks_like_number/;
use Data::Dumper; # Debug

my %hash;
while (my $string = <DATA>) {
    chomp($string);
    # check if string has at least one comma, and at the same
    # time extract the value after the last comma
    if ( my ($path,$value) = $string=~/^(.*),([^,]*)$/ ) {
        if ( looks_like_number($value) ) {
            $hash{$path} = $value;
        }
        else {
            $hash{$path} = 1;
        }
    }
    else {
        $hash{$string} = 1;
    }
}

print Dumper(\%hash); # Debug

__DATA__
foo
bar,x
quz,5
a,b,c,42
[download]

Of course there's lots of potential for shortening that, e.g. by combining it with my example code from here. Update: A really simple shortening:

while (<DATA>) {
    chomp;
    if ( /^(.*),([^,]*)$/ )
         { $hash{$1} = looks_like_number($2) ? $2 : 1 }
    else { $hash{$_} = 1 }
}
[download]