This is how I would do that, with Data::Dumper output for comparison.

#!/usr/bin/perl -w

use strict;
use warnings;
use Data::Dumper;

open my $fh, '<', $ARGV[0] or die $!;
my @lines = <$fh>;
close $fh;

my %data;

for my $line (@lines) {
    my @fields;
    @fields = split /\s+/, $line;

    $data{$fields[1]}{$fields[0]} = $fields[2];
}


for my $key (keys %data)
{
    print "$key = $data{$key}\n";

    for my $subkey (keys %{$data{$key}}) {
       print "\t$subkey = $data{$key}{$subkey}\n";
    }
}

print Dumper(\%data);
[download]

Here is a sample run and the data.

my@mybox:~/sandbox
$ ./2.pl data.txt
jones = HASH(0x40017524)
        jenny = circle
        ted = circle
knight = HASH(0x4001738c)
        ted = triangle
$VAR1 = {
          'jones' => {
                       'jenny' => 'circle',
                       'ted' => 'circle'
                     },
          'knight' => {
                        'ted' => 'triangle'
                      }
        };
my@mybox:~/sandbox
$ cat data.txt
ted jones square
ted jones circle
ted knight triangle
jenny jones circle
[download]

The key differences from Perl Best Practices :

• Don't use bareword file handles.
• Use the three-argument form of open.
• Use indirect file handles.
• Close file handles explicitly, and as soon as possible.
• Avoid C-style for statements.
• Don't use unnecessary parentheses for builtins and "honorary" builtins.

And also just the general idea of don't make global what can be local in scope ($key, $subkey, and @fields).

I did, violate the "Prefer line-based I/O to slurping," rule because we both know the sample data is tiny.