Make the "variants" an array and push items onto it. Then sift out unique values once all the data has been read.

use strict;
use warnings;

use Data::Dumper;

open my $inFH, q{<}, \ <<EOD or die $!;
The    DT    the
International    NN    International
for    IN    for
well    NN    well
preparation    NN    preparation
preparation    NN    preparation
in    IN    in
conference    NN    conference
conference    NN    conference
conferences    NN    conference
good    VVG    good
EOD

do { my $discard = <$inFH> };
my %hash;

while ( <$inFH> )
{
    my @tags = split;
    next unless $tags[ 1 ] eq q{NN};
    
    $hash{ $tags[ 2 ] }->{ frequency } ++;
    push @{ $hash{ $tags[ 2 ] }->{ variants } }, $tags[ 0 ];
}

@{ $hash{ $_ }->{ variants } } = do {
   my %seen;
   grep { not $seen{ $_ } ++ } @{ $hash{ $_ }->{ variants } };
   } for keys %hash;

print Data::Dumper->Dumpxs( [ \ %hash ], [ qw{ *hash } ] );
[download]

The output.

%hash = (
          'preparation' => {
                             'frequency' => 2,
                             'variants' => [
                                             'preparation'
                                           ]
                           },
          'conference' => {
                            'frequency' => 3,
                            'variants' => [
                                            'conference',
                                            'conferences'
                                          ]
                          },
          'well' => {
                      'frequency' => 1,
                      'variants' => [
                                      'well'
                                    ]
                    },
          'International' => {
                               'variants' => [
                                               'International'
                                             ],
                               'frequency' => 1
                             }
        );
[download]

I hope this is helpful.

Update: Perhaps simpler would be to keep a ->{ seen }->{ $tags[ 0 ] } sub-sub-HoH to filter out duplicates and delete it at the end.

...
while ( <$inFH> )
{
    my @tags = split;
    next unless $tags[ 1 ] eq q{NN};
    
    $hash{ $tags[ 2 ] }->{ frequency } ++;
    push @{ $hash{ $tags[ 2 ] }->{ variants } }, $tags[ 0 ]
       unless $hash{ $tags[ 2 ] }->{ seen }->{ $tags[ 0 ] } ++;
}

delete $hash{ $_ }->{ seen } for keys %hash;
...
[download]